A Handbook For Seismic Attributes

Downloaded 12/13/16 to 155.247.166.234. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.
org/
Downloaded 12/13/16 to 155.247.166.234. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Handbook of Poststack
Seismic Attributes
Arthur E. Barnes
Geophysical References Series No. 21

Elizabeth Lorenzetti Harvey, volume editor
Rebecca Latimer, managing editor
SM
Society of Exploration Geophysicists
8801 S. Yale, Ste. 500
Tulsa, OK 74137-3575
# 2016 Society of Exploration Geophysicists

All rights reserved. This book or parts hereof may not be reproduced in any form without written permission
from the publisher.
Published 2016
Printed in the United States of America
Library of Congress Control Number: 2016945977
ISBN 978-0-931830-47-1 (Series)

ISBN 978-1-56080-331-7 (Volume)
To Irene
Dedication
This page has been intentionally left blank

Contents
About the Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
Chapter 1: Overview of Poststack Seismic Attributes .............................. 1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Attributes and methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Characteristics of seismic attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Seismic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Chapter 2: History of Seismic Attributes ......................................... 13

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Digital recording and bright spots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Signal attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Nigel Anstey’s attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Complex seismic trace analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Seismic stratigraphy and inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Proliferation and disillusionment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Discontinuity and attribute revival . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Multiattribute analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Recent developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Chapter 3: Attribute Maps and Interval Attributes ............................... 27

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Horizon attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Interval attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Gallery of interval attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Chapter 4: Complex Seismic Trace Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
1D complex seismic trace analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Spikes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Response attributes and average attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3D complex seismic trace analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
v
vi Handbook of Poststack Seismic Attributes
Gallery of complex trace attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Chapter 5: Structural and Stratigraphic Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Dip and azimuth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Seismic shaded relief . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Volume curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Stratigraphic attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Chapter 6: Seismic Discontinuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Discontinuity based on energy ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Discontinuity based on derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
Improving discontinuity attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Chapter 7: Spectral Decomposition and Waveform Classification ................ 113

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Thin beds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Spectral decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Waveform classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Chapter 8: Relative Acoustic Impedance and Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Relative acoustic impedance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Chapter 9: Multiattribute Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Volume blending . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Crossplots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Principal component analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Automatic pattern recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Contents vii
Chapter 10: Applying Seismic Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Choosing suitable attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Reconnaissance and presentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Bright spots and amplitude mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Low-frequency shadows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Diapirs and gas chimneys . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
Geobody extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Seismic attributes in data processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
Appendix A: Analysis Windows ................................................. 201

Appendix B: Hilbert Transform ................................................. 207
Appendix C: Derivative Filter ................................................... 211
Appendix D: Discrete Formulas for Approximating Instantaneous Frequency
and Relative Amplitude Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Appendix E: Vector Traces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Glossary ........................................................................ 223
References ...................................................................... 237
Index ........................................................................... 249

About the Author
Arthur E. Barnes received a BS (1974) in physics from Denison University, an MS (1980)

in geophysics from the University of Arizona, and a PhD (1990) in geophysics from Cornell
University. His experience includes seismic data acquisition and processing, software
development, software pre-sales, and research. His employers have included Western Geo-
physical, Conoco, Ecole Polytechnique de Montreal, Landmark (Halliburton), Paradigm,
and PETRONAS Research. He has applied and developed seismic attributes throughout
his career. Currently a consultant in seismic software development and services, his pro-
fessional interests include seismic attributes, interpretive processing, and seismic pattern
recognition. He is a member of SEG, EAGE, and AAPG.
ix

Preface
The Handbook of Poststack Seismic Attributes is a general reference for poststack

seismic attributes intended for reflection seismologists in petroleum exploration. The
goal of the book is to bring greater understanding and order to the important and rapidly
evolving science of seismic attributes, so that geophysicists can apply attributes more
effectively to interpret seismic data. To this end, I emphasize what all attributes have in
common, what they mean, and what they measure, arguing that the meaning of an attribute
should guide both its implementation and its application. I freely judge certain attributes to
be useful and others to be useless, and I consider the advantages as well as the shortcomings
of attribute analysis. I provide sufficient mathematics to implement the attributes, favoring
clarity and simplicity over mathematical rigor. In the manner of a handbook, I cover
methods and ideas that are more likely to be encountered in practice, but I make no pretense
of being comprehensive. Indeed, ponderous books can be written on topics that are treated
here only cursorily, such as spectral decomposition and seismic pattern recognition.
I begin the book by introducing the fundamental ideas that underlie all seismic attribute
analysis and reviewing the history of seismic attributes from their origins to current devel-
opments. I describe the characteristics of key and familiar poststack attributes, starting with
attribute maps and interval statistics, and progressing through to complex trace attributes,
3D attributes that quantify aspects of geologic structure and stratigraphy, seismic discon-
tinuity attributes, spectral decomposition, thin-bed analysis, waveform classification,
recursive inversion for relative acoustic impedance, and spectral ratioing for Q estimation.
I discuss how attributes are usefully combined through multiattribute analysis through
volume blending, cross-plotting, principal component analysis, and unsupervised classifi-
cation. I end the book with a brief overview of how seismic attributes aid data interpret-
ation, with a look at bright spots, frequency shadows, faults, channels, diapirs, and data
reconnaissance. A glossary provides definitions of seismic attributes and methods, and
appendices provide necessary background mathematics.
It is my sincere hope that this book instills a greater appreciation for seismic attributes,
moderated by a clearer awareness of their limitations.
xi

Acknowledgments
This book is a product of several decades spent studying seismic attributes. I thank the
many colleagues who have influenced my thinking, particularly Larry Fink, Kenny Laugh-
lin, Boshara Arshin, and the late Turhan Taner. I am sincerely grateful to Nigel Anstey for a
long and lively letter detailing his early development of seismic attributes, as well as for
copies of his famous but rare reports on attributes from 1972 and 1973. I especially
thank Tracy Stark and Jeffrey Thurston for many stimulating discussions and for openly
sharing ideas and insights.
I acknowledge Crown Minerals and the New Zealand Ministry of Economic Develop-
ment for permission to show images of their data from the Taranaki Basin, offshore New
Zealand, from which most of the data examples are taken.
I warmly thank the editors who have devoted many hours in helping put this book
together, namely Rebecca Latimer, Elizabeth Lorenzetti Harvey, Sergey Fomel, Marilyn
Perlberg, and Susan Stamm. Their valuable suggestions and many corrections have
improved the book significantly. However, errors and shortcomings are inevitable, and I
encourage interested readers to point them out and to offer their comments.
Finally, I thank my wife, Irene, for her constant support and encouragement over the
many years during which this book gradually took form.
xiii

Chapter 1
Overview of Poststack Seismic Attributes
Introduction
Seismic attributes are tools for inferring geology from seismic reflection data. Seismic
attributes aid seismic interpretation by revealing hidden features, by identifying similar
patterns, and by quantifying specific properties (Figure 1). Attribute analysis is a vital
facet of reflection seismology for petroleum exploration and finds wide application,
from anomaly identification to feature extraction to lithological prediction.
Seismic attributes quantify properties of seismic data; seismic attributes describe seis-
mic data. As seismic data can be described in countless ways, the potential number of
seismic attributes is likewise countless. Hundreds of diverse attributes have been invented
and more appear each year. Their interpretation remains largely a matter of qualitative
investigations with individual attributes, but quantitative multiattribute analysis is
slowly growing.
Seismic attributes are constituents of seismic data. In a sense, seismic data are the sum
of their attributes. Attribute analysis decomposes data into attributes, but the decompo-
sition is informal because no rules govern how to compute attributes or what they represent.
Attribute computations act as filters that remove some component of the signal to reveal
another component. It is often argued that attributes are never as good as the original
seismic data because they have less information. This criticism misses the mark entirely:
attributes are useful precisely because they have less information.
Seismic attributes record information from prestack data gathers or poststack
data volumes. Prestack attributes include P- and S-wave velocities and impedances,
amplitude variation with offset intercept and gradient, anisotropy, attenuation, and
seismic wave illumination. Poststack attributes include complex trace attributes, interval
statistics, discontinuity, time-frequency attributes, waveform, and 4D differences. Prestack
attributes treat seismic data as records of seismic reflections. Poststack attributes treat
seismic data as images of the earth. Prestack attributes are derived through involved
methods of geophysical inversion. They provide valuable clues about lithology and fluid
content, but they are relatively expensive, demand sophisticated interpretation, and
require specific kinds of data or processing. Poststack attributes are derived through
filters, transforms, and statistics. They quantify stratigraphic and structural properties
and are easy to compute and apply, but they lack the direct ties to lithology that are of
primary interest.
Though prestack and poststack attributes have many elements in common, profound
differences in theory and application justify separate treatment. This book surveys post-
stack seismic attributes.
1
2 Handbook of Poststack Seismic Attributes
Figure 1. Seismic attributes reveal features hidden in seismic data. Here faults and diapirs
are much clearer in the attribute display on the right than in the original seismic data on the left.
The attribute display blends discontinuity (red) with seismic shaded relief (blue).
Attributes and methods

Seismic attributes are subsets of the information in seismic reflection data. Typically,
they retain the form of the data from which they derive, so that for every point in a seismic
volume, or for every point on a seismic horizon, there is a corresponding point in the
derived attribute volume or map. Seismic attributes vary by the seismic properties they
measure and by the way they measure the properties. There are many seismic properties
and many ways to measure them, so there are many types of attributes.
Types of attributes
Seismic attributes are geological, geophysical, or mathematical. Geological attributes
are the most useful, and mathematical attributes the least useful; geophysical attributes
have intermediate utility.
Geological attributes record structural, stratigraphic, or lithological properties of
seismic data. Structural attributes include dip, azimuth, curvature, and discontinuity.
Here, stratigraphic attributes refer largely to 3D reflection patterns as described by the
ideas of seismic stratigraphy. They include reflection spacing, parallelism, and thin-bed
thickness. Lithological attributes measure porosity, density, fluid content, and sand
percent. Velocity, impedance, and quality factor also can be considered lithological attri-
butes. Poststack seismic attributes adequately quantify structural and stratigraphic proper-
ties but not lithological properties, which require prestack, multicomponent, or vertical
seismic profile data for their measurement. Geological attributes are easy to understand,
but are often difficult to quantify.
Geophysical attributes record properties of seismic waves and wavelets. They include
amplitude, phase, frequency, and bandwidth. They must be inverted or interpreted to obtain
Chapter 1: Overview of Poststack Seismic Attributes 3
geological information. Geophysical attributes are relatively straightforward to measure,

but they can be difficult to understand and relate to geology.
Mathematical attributes record averages, variances, counts, ratios, and other statistics
of seismic data. They are the most numerous category of seismic attributes because they are
easy to create. Mathematical attributes have purely mathematical definitions and lack
inherent geological significance.
Methods of computation
Poststack seismic attributes are generated through a wide variety of methods, some
purely empirical, and others based firmly on theory. These methods include statistics,
map computations, complex seismic trace analysis, correlation, semblance, principal com-
ponent analysis, filtering, spectral decomposition, dip scanning, plane-wave destruction,
gradient squared tensor, and unsupervised classification (Table 1). Many attributes can
be computed through several methods.
Table 1. Methods for computing poststack seismic attributes with representative attributes.
Many attributes can be computed through several methods.
Method Representative attributes
Statistics Average value, rms amplitude, largest value, maximum

trough, total energy, variance, skew, kurtosis, number of
peaks, percent above threshold
Map computations Discontinuity, dip, azimuth, curvature, shaded relief
Complex seismic trace analysis Reflection strength, phase, frequency, bandwidth, amplitude
change, dip, azimuth, parallelism, curvature, quality
factor, thin-bed indicator
Response attributes Apparent polarity, response amplitude, response phase,
response frequency, sweetness
Correlation, semblance, Discontinuity, dip, azimuth, curvature, parallelism
principal component analysis
Dip scanning, plane-wave Discontinuity, dip, azimuth, curvature, parallelism
destruction, gradient
squared tensor
Spectral decomposition Frequency components, average frequency, bandwidth,
tuning thickness
Pattern recognition Waveform maps, attribute classes or seismic facies
Miscellaneous Zero-crossing frequency, arc length, energy half-time,
relative acoustic impedance, quality factor
Statistics provide the basis for a large set of mathematical attributes that include the
mean value, root-mean-square (rms) amplitude, variance, largest value, smallest value,
number of peaks, and average energy. Statistical attributes usually are presented as maps.
Map computations apply difference operators to interpreted seismic horizons to
measure structural properties, primarily dip, azimuth, and curvature.
Complex seismic trace analysis separates seismic data into two basic components,
amplitude and phase, from which other attributes derive through differentiation, averaging,
and combination. Of all the methods for attribute computation, complex trace analysis is
the most flexible and produces the widest variety of attributes, chiefly reflection strength,
instantaneous phase, instantaneous frequency, and amplitude change. Response attributes
are a subset of complex trace attributes that include apparent polarity, response phase,
response frequency, and sweetness.
Crosscorrelation, semblance, and principal component analysis measure how much
neighboring seismic traces resemble each other. They are applied to compute structural
and stratigraphic attributes, primarily discontinuity, dip, azimuth, curvature, and parallelism.
Dip scanning, the plane-wave destructor, and the gradient squared tensor employ 3D
derivatives in an averaging window to measure dip, azimuth, curvature, and continuity.
Spectral decomposition and other time-frequency methods produce geophysical
attributes that record time-varying spectral properties of seismic data. These properties
include frequency components, mean frequency, peak spectral frequency, tuning fre-
quency, and spectral bandwidth.
Methods of automatic pattern recognition classify seismic data according to character-
istic patterns in the data. They are applied to seismic waveforms and to sets of attributes to
produce waveform maps and seismic facies volumes.
A few miscellaneous attributes are computed through specific methods. These include
arc length, energy half-time, zero-crossing frequency, and thin-bed indicator.
Seismic attributes are instantaneous, computed “at a point,” or local, computed in a
window around a point. Instantaneous attributes usually are derivatives, and local attributes
are most frequently averages. The term “instantaneous attribute” originally referred
to attributes in time. It has since come to refer to any attribute computed at a point,
whether in time or in depth. For discrete seismic data, computing at a point usually requires
an operator that spans multiple data points, such as a difference operator.
Categorization
Categorizing seismic attributes helps make sense of their great number and confusing
variety. Attributes can be categorized by computational method (Table 1), by application
(Chen and Sidney, 1997), or by the property measured (Chen and Sidney, 1997; Brown,
2011, p. 248; Figure 2).
Characteristics of seismic attributes

Seismic attributes are filters that quantify properties of seismic images. The most useful
seismic attributes are unique, comparable, easy to use, and geologically meaningful.
Seismic attributes
Mathematical Geophysical Geological
Statistics Miscellaneous Wave Wavelet Reflection patterns

Mean Principal components Amplitude Polarity
Variance Spectral components Phase Response amplitude
Skew Signal complexity Discontinuity
Frequency Response phase
Kurtosis Arc length Lateral amplitude change
Bandwidth Response frequency
Total energy Energy half-time Inst. quality factor Tuning frequency
Largest value Effective bandwidth
Smallest value
Average absolute value Reflection strength
Largest absolute value Root-mean-square ampliitude Inst. frequency
Number of peaks Relative amplitude change Peak frequency Structural Stratigraphic Lithological
Number of troughs Amplitude acceleration Average frequency
Ratio peaks/troughs Root-mean-square frequency Dip or slope Reflection spacing Quality factor
Percent above threshold Zero-crossing frequency Azimuth Spacing change Porosity
Frequency change Shaded relief Sweetness Sand indicator
Curvature Parallelism Shale indicator
Fault indicator Divergence Fluid indicator
Waviness Impedance
Chaos measure P-wave velocity
Mean curvature Thin bed indicator S-wave velocity
Maximum curvature Amplitude variance
Minimum curvature Thin bed thickness
Most positive curvature Waveform Relative acoustic impedance
Most negative curvature P-wave impedance
Gaussian curvature S-wave impedance
Dip curvature Acoustic impedance
Strike curvature Elastic impedance
Amplitude curvature
Curvature gradient
Figure 2. Seismic attributes categorized by property. Most lithological attributes are prestack attributes and are not discussed in this book.
Chapter 1: Overview of Poststack Seismic Attributes
5
Attributes quantify properties

Seismic attributes quantify properties of seismic data. Attributes inherit the meanings
of the properties that they quantify. Define a seismic attribute by the property it measures,
not by its mathematics.
Seismic attribute analysis is a kind of image analysis applied to seismic images. Seis-
mic images are seismic data for which geophysical corrections are complete and
geological interpretation can proceed. They can be treated as pictures of the Earth’s
subsurface made up of a complex sequence of seismic reflections, with a modest level
of noise. Seismic reflections are characterized by continuity in phase and amplitude.
These idealizations provide the basis on which seismic properties and attributes are defined.
Attributes are filters

Seismic attributes are filters in that they remove some component of the signal to high-
light another component. Attributes necessarily have less information than the data from
which they derive. In this way, they simplify seismic data to aid interpretation.
Many attributes are derived within analysis windows. The size of the window deter-
mines the attribute resolution. Small windows enhance resolution, whereas large
windows reduce noise. Design attribute windows to reduce artifacts, such as Gibbs’
phenomenon (Appendix A).
Attributes should have useful meaning

Seismic attributes should have clear and useful meanings. If you don’t know what
an attribute means, don’t use it. If you know what it means but it isn’t useful, don’t use it.
Seismic interpretation demands attributes with geological or geophysical meaning.
It is nearly impossible to establish geological meaning for an attribute that lacks inherent
ties to geology. Empirical relations between mathematical attributes and geological prop-
erties rarely have more than local validity. Prefer attributes with geological meaning and
avoid attributes with purely mathematical meaning.
Attributes should be unique

Seismic attributes should contain some unique information. You need only one
attribute to describe a property. Where multiple attributes measure the same property,
choose the one that works best and discard the others.
Attributes should be easy to use

Seismic attributes should be easy to use. Parameters should be few and obvious. Attri-
butes with complicated parameters often are poorly conceived. Prefer attributes with
simple parameters.
Attributes should have data distributions that are convenient for display and analysis.
This sometimes requires rescaling or clipping to spread attribute values more evenly across
the display color scale.
Seismic properties
Seismic properties are geological, geophysical, or mathematical. Properties measured
down seismic traces are 1D; properties measured through seismic volumes are 3D. Most
mathematical and geophysical properties are 1D, and most geological properties are 3D.
Seismic data have countless properties, though relatively few are important for attribute
analysis. Key 1D poststack seismic properties include amplitude, frequency, bandwidth,
amplitude change, tuning thickness, and waveform. Key 3D properties include dip,
azimuth, discontinuity, curvature, and parallelism.
This section describes key seismic properties and their sign conventions. The coordi-
nate systems employed are those peculiar to reflection seismology, in which time or depth
increase downward (Figure 3).
Amplitude
Amplitude refers to the magnitude of the seismic trace values or trace envelope.
Amplitude is always positive and is independent of the polarity or phase of the data. It is
the most important seismic property and has more attributes than any other. The most
common amplitude attributes are reflection strength, rms amplitude, maximum amplitude,
average absolute amplitude, and total energy. Unlike most properties, amplitude is not gen-
erally comparable between different data sets, due to the vagaries of seismic data acquisi-
tion and processing. Thus the same reflection can have wildly different amplitudes when
imaged on different data sets.
Reflection strength is the best amplitude measure for most purposes. The term “reflec-
tion strength” was coined by Nigel Anstey (1972) and originally referred to an amplitude
a) b)
Figure 3. The (a) 2D and (b) 3D coordinate systems of reflection seismology for data in time; time
increases downward. For depth data, depth variable z replaces time variable t. The 3D coordinate
system obeys a left-hand rule.
attribute computed by smoothing a rectified

seismic trace. Reflection strength now is com-
puted through complex seismic trace analysis

and is synonymous with trace envelope and
instantaneous amplitude.
Amplitude attributes reveal bright spots
and dim spots caused by gas, tuning, hard
streaks, or porosity changes.
Figure 4. Phase is a measure of relative
position along a waveform x(t). By the
convention followed in this book, peaks have Phase and polarity
08 phase, troughs have 1808, downgoing zero
crossings have +908, and upgoing zero Phase refers to a relative position along a
crossings have 2908. waveform (Figure 4). It is independent of
amplitude. Though phase differences are the
basis for many important attributes, phase
itself produces only a few attributes of limited value, chiefly instantaneous phase, response
phase, and apparent polarity.
Phase also refers to the average spectral phase of a seismic wavelet. The phase of
the seismic wavelet in the data is unimportant for most poststack seismic attributes, but
it must be zero for tuning attributes, apparent polarity, and relative acoustic impedance.
The phase convention followed here is that a cosine wave has 08 phase and a sine
wave has 2908 phase. A process that transforms a cosine wave into a sine wave
changes the phase by 2908, and a process that transforms a sine wave into a cosine
wave changes the phase by +908. Table 2 summarizes the phase changes that are intro-
duced by standard operations of seismic attribute analysis.
Polarity refers to the sign of a reflection at its maximum or at its envelope peak.
Polarity also refers to the sign of reflections in general with respect to a convention.
The convention followed here is that positive polarity corresponds to a positive reflection
coefficient and indicates an increase in impedance.
Frequency
Frequency refers to the number of sinusoidal cycles that occur along a waveform in a
given time interval. Frequency attributes are inherently trace attributes of time data and
Table 2. Phase changes introduced by standard operations in seismic attribute analysis. Processes
that employ these operations are given along with their effect on the phase of a pure sinusoid.
The Hilbert transform is reviewed in Appendix B and differentiation is reviewed in Appendix C.
Operation Phase change Effect Process
Differentiation +908 sin(u) cos (u) Derivative filter

Integration 2908 cos(u) sin (u) Relative acoustic impedance
Hilbert transform 2908 cos(u) sin(u) Quadrature filter
Phase rotation f cos(u) cos(u + f) Phase rotation filter
include instantaneous frequency, zero-

crossing frequency, average spectral fre-
quency, rms frequency, and tuning

frequency.
Frequency attributes serve conflicting
purposes. They are applied to measure
bed thickness, a geological property, or
to measure seismic attenuation, a geophy-
sical property. For measuring bed thick-
ness, the influence of the wavelet must Figure 5. A common and useful measure of
be negligible so that it does not mask the bandwidth f is the standard deviation of the
b
stratigraphy. For measuring attenuation, spectral power P( f ) about the average spectral
the influence of the stratigraphy must be frequency fa.
negligible in order to detect changes in
the wavelet. The same frequency measure cannot adequately serve both purposes.
Average frequency in a large window is best for detecting attenuation; tuning frequency
in a short window is best for estimating bed thickness.
Bandwidth
Bandwidth refers to the breadth of the frequency power spectrum of a waveform. For a
seismic trace in time, bandwidth is a function of the changes in frequency and amplitude
along the trace. Bandwidth attributes are applied much like frequency attri-
butes to distinguish stratigraphic features or identify attenuation. In stratigraphic analysis,
relatively small bandwidth suggests uniform or predictable stratigraphy, whereas high
bandwidth suggests complex stratigraphy or noise. Though a fundamental property of
seismic data, bandwidth has proven less useful than frequency and has found limited appli-
cation in seismic attribute analysis.
There is no standard measure of bandwidth. A useful and mathematically convenient
measure is the standard deviation of the power spectrum about the mean spectral frequency
(Berkhout, 1984, p. 28; Cohen, 1995, p. 8; Figure 5). This measure might better be termed
“half bandwidth” because it provides an estimate of bandwidth that is about one half to one
quarter of the breadth of the power spectrum.
Amplitude change
Amplitude change is how much the seismic amplitude or envelope changes over an
interval in a given direction. Relative amplitude change is amplitude change normalized
by the amplitude. Relative amplitude change is comparable between different data sets.
Measured down seismic traces in time, it is scaled to have units of Hz and is an instan-
taneous measure of bandwidth.
Amplitude change attributes reveal faults, channel edges, and other details hidden in
the seismic amplitudes. Being directional, they look like illuminated topography when
displayed in monochrome.
Slope, dip, and azimuth

Dip and azimuth together quantify reflection orientation. The terms “slope” and “dip”
are commonly used interchangeably, but there is a distinction. For depth data, slope is the
dimensionless ratio of the change in depth of a reflection over a certain horizontal distance.
Dip is the angle in degrees that a seismic reflection makes with the horizontal, which is the
arctangent of the slope. For time data, slope becomes inverse apparent velocity with units
of milliseconds per meter, and dip must be estimated using a conversion velocity. Slope is a
geophysical property and dip is a geological property.
Slope in 2D and slope components in
3D are signed according to the convention
shown in Figure 6. In 3D, slope and dip are
unsigned magnitudes that correspond to
the maximum slope or dip, which is in the
direction given by the azimuth (Figure 7).
Azimuth is the angle measured clockwise
from geographic north of the downdip direc-
tion of maximum slope or dip (Figure 8).
Figure 6. Sign convention for dip and slope in
Reflection orientations can be cast as unit
2D. Slope px ¼ dz/dx. Dip angle g is the
arctangent of px, and has the same sign as slope. vectors that are everywhere normal to
The 3D slope components px and py also follow reflection surfaces. Orientation vectors
this sign convention. For time data, dip form a vector field that describes the
calculation requires depth conversion, and imaged geological structure. This idea is
slopes have units of slowness. straightforward for seismic data in depth,
but it requires adjustment for seismic data
in time due to the inconsistency in units
between the spatial horizontal axes and the
temporal vertical axis. Vector fields are
important in the design of qualitative 3D
stratigraphic attributes, and of reflection-
guided processes, such as coherency
filtering.
Curvature
Curvature refers to the rate of change of
dip and azimuth along a seismic reflection
or horizon. Dips and azimuths can vary dif-
ferently in different directions, so curvature
Figure 7. Dip and slope are unsigned in 3D. forms a complicated set of properties, which
For seismic data in depth, dip angle g is the includes mean curvature, Gaussian curva-
arctangent of slope magnitude p, where p is ture, maximum curvature, minimum curva-
given by p ¼ |dz/dr|, and dr is defined by ture, most positive curvature, most negative
1/dr 2 ¼ 1/(dx 2 + dy 2). curvature, dip curvature, strike curvature,
curvedness, and shape index (Roberts, 2001). Mean and Gaussian curvatures are funda-
mental curvature properties from which many others derive. However, only most positive
and most negative curvatures have found general application as seismic attributes, though
technically they are not true curvature measures (Rich, 2008).
Curvature is important chiefly because of its association with fractures (Lisle, 1994).
Fractures tend to develop in folded layers of brittle rock where the layer curvature is
large. Thus fractures tend to occur at anticlinal tops, synclinal bottoms, and at flexures.
Strong curvature does not indicate that fractures exist, or whether they are open if they do
exist; it simply indicates where fractures are more likely to develop. Curvature cannot
image fractures directly if their size is below seismic resolution, which is often the case.
Curvature attributes involve second
derivatives, so they are particularly
susceptible to noise and acquisition arti-
facts. The remedy is to incorporate aver-
aging into the computation through large
operators, or to smooth the seismic data
prior to computing attributes. Curvature
reveals faults, channels, anticlinal tops,
synclinal bottoms, and flexures. Large
operators are required to identify broad
anticlines and synclines (Bergbauer
et al., 2003), but small operators suffice
for routine purposes.
Figure 9 illustrates the sign con-
vention for curvature. This convention
Figure 8. Reflection azimuth f is the angle holds for maps as well as volumes and
measured clockwise from north of the downdip is consistent with the sign convention
direction of the maximum dip. The x and y axes are for dips and slopes shown in Figure 7.
the coordinate system of the seismic survey; angle By this convention, anticlines and reflec-
fo refers the survey to north. tion bumps have positive curvature,
Figure 9. Sign convention for reflection curvature. Anticlines have positive curvature, synclines
have negative curvature, and regions of constant dip have zero curvature. Faults often exhibit
positive curvature on the upthrown side and negative curvature on the downthrown side. Channels
tend to have positive curvature on their sides and negative curvature in their interiors.
whereas synclines and reflection sags have negative curvature. Faults and flexures have
both positive and negative curvature.
Discontinuity
Seismic continuity is the degree to which seismic reflections have consistent ampli-
tude and phase. Seismic discontinuity is the opposite of continuity, and refers to breaks
in the continuity of seismic reflections. Discontinuity attributes reveal these breaks,
which are caused by faults, diapirs, channels, pinch-outs, noise, and artifacts. They are
the most important 3D attributes.
Seismic discontinuity is often called coherence or similarity. Because discontinuity
is applied to image geological features, the term “discontinuity” is preferable because
it suggests a geological property, whereas “coherence” suggests a geophysical property
and “similarity” suggests a mathematical property.
Outline
This handbook of poststack seismic attributes is a general reference for reflection seis-
mologists. It covers attribute theory, meaning, computation, and application.
This chapter prepares the ground for the subsequent chapters by explaining the basic
ideas that underlie all seismic attributes.
Chapter 2 tells the history of seismic attributes, from their origins in the 1950s to
current developments.
Chapter 3 discusses attribute maps and interval statistics. It introduces basic measures
that appear throughout attribute analysis.
Chapter 4 presents complex seismic trace analysis. It introduces instantaneous and
weighted average trace attributes, and extends the 1D theory to 3D to quantify reflection
dip and azimuth.
Chapter 5 describes 3D attributes that quantify elements of geological structure and
stratigraphy, primarily dip, azimuth, curvature, reflection spacing, and parallelism.
Chapter 6 develops seismic discontinuity attributes as variances or differences and
shows how they are improved through dip corrections and filters.
Chapter 7 covers spectral decomposition, thin-bed analysis, and waveform
classification.
Chapter 8 details two methods that purport to record rock properties, relative acoustic
impedance, and Q estimation through spectral ratioing.
Chapter 9 reviews multiattribute analysis through volume blending, cross-plotting,
principal component analysis, and unsupervised classification.
Chapter 10 concludes the book with a brisk tour of the application of seismic attributes
in seismic data interpretation. It looks at bright spots, frequency shadows, faults, channels,
diapirs, and data reconnaissance. It suggests which attributes are effective for these purposes.
A glossary provides concise definitions of seismic attributes and methods, and appen-
dices provide background mathematics.
Chapter 2
History of Seismic Attributes
Introduction
The history of seismic attributes began in the 1950s with the earliest efforts to extract
geological information from seismic reflection character. These efforts became feasible
only with the digital revolution of the 1960s, which was sparked by digital recording in
the field. Digital recording facilitated data processing and greatly increased dynamic
range, making amplitude analysis possible. This led to the watershed discovery of bright
spots, which conclusively proved the value of reflection character and stratigraphic
interpretation. The success of bright spots encouraged the search for other ways to quantify
reflection character. This search produced frequency analysis, impedance inversion,
complex trace attributes, and color displays in the 1970s. The popularity of complex
trace attributes in the 1980s stimulated the development of new poststack attributes as
well as the first prestack attributes, but the lack of geological significance bred disillusion-
ment. Attributes returned to favor in the mid-1990s with the introduction of attribute maps,
seismic discontinuity, spectral decomposition, and multiattribute techniques. Recent years
have seen the invention of curvature attributes, automatic fault extraction, advanced
methods of impedance inversion, and prestack attributes tied to rock properties. These attri-
butes and methods address the demand for greater detail, clearer ties to geology, and faster
ways to interpret seismic data. These demands continue to drive the development of
seismic attributes.
This is the history of seismic attributes in brief. A fuller description follows, with
emphasis on poststack attributes and an assessment of their significance.
Digital recording and bright spots

From the first practical seismic reflection experiments in 1921 until the early 1960s,
seismic reflection data interpretation was largely a matter of identifying reflections and
converting their times to depth to map subsurface geological structure. It was difficult to
glean much else from the data. Structural interpretation ruled, and stratigraphic interpret-
ation languished (Figure 1).
The early technology of reflection seismology was inadequate for more than rudimen-
tary structural mapping, but this began to change with advances in the 1950s. Analog mag-
netic recording, introduced about 1952, replaced paper records and enabled data to be
redisplayed or processed. Data processing was initially analog and limited to simple pro-
cesses such as automatic gain and filtering. Analog-to-digital converters appeared by the
late 1950s, facilitating data processing on digital computers, which were more versatile
13
Figure 1. A seismic section recorded in 1956 in Egypt, with interpreted horizons (from Masson
and Agnich, 1958, Figure 7, p. 339). Prior to the advent of digital recording, seismic interpreters had
little alternative to basic structural interpretation.
than analog computers. These innovations were not enough to support routine stratigraphic
interpretation, but they laid the groundwork for the coming digital revolution.
Even with analog data, a few intrepid visionaries recognized that seismic reflection
character contained clues to stratigraphy. Ben Rummerfield was foremost among them.
In 1954, he published his famous paper on mapping “reflection quality” to reveal subtle
stratigraphic features (Rummerfield, 1954; Figure 2). Lindseth (1982, p. 9.2) considers
this idea a forerunner of bright spot analysis, but it is as much a forerunner of seismic attri-
bute analysis in general. Rummerfield was remarkably prescient because he foresaw that
with improvements in reflection seismology, one could deduce fluid content, porosity,
and facies changes. Otto Koefoed was no less visionary. In 1955, he suggested that litho-
logical properties might be inferred from amplitude variation with offset (AVO) effects
observed on prestack data gathers (Koefoed, 1955), and in 1960 he investigated the record-
ing of true seismic amplitudes and the extraction of lithological information (Koefoed,
1960). Another visionary, Eduardo Merlini, developed the “summarizer trace” in 1959
to provide a crude form of true amplitude recording (Merlini, 1960; Savit, 1960; Lindseth,
Chapter 2: History of Seismic Attributes 15
Figure 2. A conceptual illustration of a dendritic channel system mapped by hand according to

Rummerfield’s ideas of reflection quality (from Rummerfield, 1954, Figure 5, p. 690).
1982, p. 9.2; Figure 3). However, these exceptions only prove the rule. Overall, geophysi-
cists mapped geological structure and paid little attention to seismic reflection amplitude or
character.
Figure 3. Eduardo Merlini’s (1960) summarizer trace, an early method for detecting amplitude
anomalies in seismic data. (a) Seismic line. (b) Derived summarizer traces (from Savit, 1960,
Figures 1 and 2, pp. 314 – 315).
Reflection seismology changed dramatically in 1963 with the introduction of digital

recording of exploration seismic data in the field (Van Melle et al., 1963; Dobrin, 1976,
p. 68). Its acceptance was so rapid that by 1968 fully half of all new seismic recording
was digital, and by 1975 nearly all was (Sheriff and Geldart, 1995, pp. 21, 23; Figure 4).
Chapter 2: History of Seismic Attributes
Figure 4. The history of the digital revolution as seen through advertisements published in GEOPHYSICS . Shown are a workflow for
stratigraphic exploration from 1964 (Geophysical Service Inc., 1964), a seismic data-processing center in 1965 (Geophysical Service Inc.,
17
1965), and an early digital recording system from 1966 (SDS Data Systems, 1966).
Digital recording greatly enhanced the dynamic range of seismic data and facilitated pro-
cessing to improve data quality. Stratigraphic interpretation and routine investigation of
amplitude variations could finally proceed. This led to the discovery of bright spots and
the recognition that they were direct hydrocarbon indicators.
Early research on bright spots was published in the Soviet Union in the late 1960s,
but this research was largely unavailable elsewhere and consequently had limited influ-
ence. Instead, methods of bright spot exploration were developed independently in
secrecy among oil companies and seismic contractors exploring in the Gulf of Mexico
in the late 1960s and early 1970s (Proubasta, 2000). By 1972, bright spot exploration
was widespread, though companies still jealously guarded the technology (Dobrin 1976,
p. 339; Waters, 1981, p. 199; Sheriff and Geldart, 1995, p. 21). While bright spots received
most of the attention, researchers began to consider other amplitude variations (Mathieu
and Rice, 1969).
The stunning success of bright spot exploration quickly established it as a key work-
flow of exploration geophysics. Perhaps even more important than its direct contribution
to finding oil and gas, bright spot exploration proved the value of reflection character, pre-
paring the way for more powerful methods to come.
Thus, the first seismic attribute was reflection amplitude. In various guises, it remains
the most important attribute today.
Signal attenuation
The successful example of amplitude as a direct hydrocarbon indicator stimulated
a search for other indicators. With buoyant hopes, researchers investigated frequency.
They reasoned that seismic energy suffers anomalous attenuation when it passes
through a gas reservoir, causing an abrupt shift to lower frequencies on reflections
from beneath the reservoir. This effect, the well-known “low-frequency shadow,”
could serve as another direct hydrocarbon indicator (Sheriff, 1975). The fondest hope
was that this attenuation could be quantified to estimate rock quality factor (Dobrin,
1976, p. 289). Many workers sought these frequency changes and a corresponding
means to display them. A. H. Balch of Marathon Oil was the first to demonstrate progress
(Balch, 1971).
Balch employed spectral decomposition and red-green-blue (RGB) color blending to
make color “sonograms” that recorded the time-variant average frequency of seismic
data. His interest lay more in measuring frequency changes than in interpreting their
origin. To encourage oil finders, he speculated that his technique might detect seismic
attenuation in gas-filled reefs. Though he developed only a frequency attribute, he
suggested that other quantities could be displayed usefully in color.
Balch’s paper is a historical curiosity. Its awkward methods and uninspiring displays
were soon forgotten. Nonetheless, it stands as a milestone in reflection seismology, as it
was the earliest published paper to display seismic data in color and the first to apply spec-
tral decomposition to seismic data.
Nigel Anstey’s attributes

Nigel Anstey’s landmark exposition of seismic attributes followed on the heels of

Balch’s study. Working for a seismic processing and acquisition contractor in the late
1960s and early 1970s, Anstey developed a novel procedure for attribute analysis, which
he detailed in two famous internal reports alive with brilliant color displays of seismic
data (Anstey, 1972, 1973a; see also Anstey, 1977, 2001, 2005). Color figures were too
expensive at the time for wide reproduction, so only a few copies of the reports were
made and given to key clients. The reports aroused keen interest among the few who
had access to them. Lively and lucid, they make fresh and insightful reading even today.
Anstey was excited by the potential of bright spots. Standard displays of relative
amplitude seismic data were inherently limited for bright spot exploration because the
apparent brightness of a seismic reflection is influenced by its polarity and phase. A
partial but inconvenient remedy was to produce both normal and reversed polarity plots
of relative amplitude processed data (Figure 5). Anstey discovered that he could highlight
bright spots better by displaying a crude measure of the trace envelope (Figure 6). He called
this measure “reflection strength,” deliberately choosing a descriptive name instead of
a technical one. Reflection strength
removes differences due to polarity and
phase from the seismic amplitudes, ren-
dering bright spots more visible and per-
mitting fairer comparisons between
amplitude anomalies. This simple inno-
vation was a big step forward. Anstey
also invented measures for apparent
polarity, frequency, frequency difference,
cross dip, and stack coherence, and pro-
moted interval velocity as an attribute.
He took care to compute his frequency
attribute on data that had not been decon-
volved or filtered, but he was skeptical
that it could reveal attenuation anomalies.
Throughout, he emphasized the meanings
of the attributes, not their mathematics.
Anstey championed the application
of color to seismic data displays and pio-
Figure 5. An example of a bright event that neered the simultaneous display of
stands out better with reverse polarity than with
seismic and attribute data to facilitate
normal polarity. In the 1970s, before amplitude
attributes were widely available, true amplitude comparison (Anstey, 1973b). He accom-
seismic data were routinely displayed with both plished this by displaying attributes in
normal and reverse polarity. In this way, bright color and over-plotting them with the
events of either polarity could be recognized original seismic data in variable area
equally well. format. His method of display was
Figure 6. Nigel Anstey’s reflection strength (after Anstey, 1972, reprinted in Barnes, 2001,
Figure 2, p. 42).
adopted universally and is the forerunner of today’s volume blending techniques. His first
color displays of seismic data were produced on early color plotters by a tedious and costly
process, but they were nonetheless a tremendous improvement over displays produced by
competing methods, and they set the standard for many years (Anstey, 2005).
Anstey concluded his famous report on seismic attributes (1973a) with a bold assertion:
“We are saying, then, that we are entering a new age of seismic prospecting – one
that yields a new insight into the geology, one that makes the seismic method far
more quantitative, and one which requires a whole new arsenal of seismic
interpretation skills. The seismic method is in the course of a great leap forward.”
Seismic attributes had arrived, but it would be several years more before the geophy-
sical world took notice.
Complex seismic trace analysis

Anstey left the seismic contractor in 1975, entrusting seismic attributes to the able
hands of his colleagues Turhan Taner, Fulton Koehler, and Robert Sheriff. They replaced
Anstey’s various empirical methods with a single mathematical framework, complex
seismic trace analysis.
The origins of complex seismic trace analysis date back to the 1920s and 1930s, when
the signal envelope and instantaneous frequency were invented as elements of signal modu-
lation for radio communication. Dennis Gabor (1946) generalized these ideas in the ana-
lytic signal, or complex trace, which Bracewell (1965) helped popularize through his
influential book on Fourier transforms.
Gelchinsky et al. (1969) introduced the complex trace to seismology in the Soviet
Union to compute the instantaneous amplitudes of seismic waves. In an overlooked but
prescient chapter of his Ph.D. thesis, Berkhout (1970) applied the complex trace to the
study of minimum phase. Farnbach (1975) followed with a study of instantaneous attributes
applied to earthquake records. Nonetheless, the complex seismic trace remained a curiosity
until Taner and Koehler harnessed it to attribute computations for exploration seismology.
Complex seismic trace analysis debuted at the 1976 annual meeting of the SEG and
subsequently was published in two seminal papers, Taner and Sheriff (1977) and Taner
et al. (1979). The timing was propitious. Complex seismic trace analysis arrived alongside
seismic stratigraphy during the oil crisis years of the late 1970s, and was followed fortui-
tously by the first practical color plotters. This potent mix of geophysics, geology, color,
and money was irresistible. Complex trace attributes vaulted to prominence, and the industry
excitedly embraced them as exotic new tools for seismic interpretation. Displays of reflec-
tion strength and other attributes were soon commonplace in prospect review meetings.
Taner and Sheriff (1977) developed five attributes: instantaneous amplitude, instan-
taneous phase, instantaneous polarity, instantaneous frequency, and weighted average fre-
quency. They patterned instantaneous amplitude after Anstey’s reflection strength, and
adopted the same name. Instantaneous polarity likewise followed Anstey’s design. For
these two attributes, mathematics follows meaning.
In contrast, instantaneous phase and frequency were new attributes that were intro-
duced by the mathematics of the complex trace. Their geophysical or geological meanings
must be inferred empirically. Taner and Sheriff noted only that instantaneous phase reveals
reflection continuity and instantaneous frequency reveals attenuation anomalies.
Seismic stratigraphy and inversion

While Taner and his colleagues were developing complex seismic trace analysis, Peter
Vail and his colleagues were formulating the principles of seismic stratigraphy. Seismic
stratigraphy is the stratigraphic interpretation of seismic reflection patterns to determine
depositional environments and infer lithology. First presented at the 1975 AAPG annual
meeting, it profoundly transformed the way geophysicists look at seismic data.
Vail learned about complex trace analysis in 1976. Sensing that it complemented his
own work, he invited Taner to present seismic attributes to his team. The rest is history:
complex seismic trace analysis first appeared in 1977 in AAPG Memoir 26, alongside seis-
mic stratigraphy (Taner and Sheriff, 1977). This famous and influential book kindled intense
interest in seismic attributes and bestowed upon them a gloss of scientific respectability.
Seismic stratigraphy draws geological inferences from observations of seismic facies
parameters. Seismic facies parameters are seismic properties related to stratigraphy and
include amplitude, spacing, continuity, parallelism, and other reflection configurations.

The early seismic stratigraphers saw seismic attributes as quantifying seismic facies param-
eters, so they adopted them enthusiastically. They anticipated that new attributes would
quantify other facies parameters. Sangree and Widmier (1977, p.183) wrote, “routine com-
puter procedures to model, store, and map a variety of seismic facies data promise to be
useful.” They added, “quantification of seismic parameters . . . holds advantages for objec-
tive prediction of facies.” This splendid vision still awaits fulfillment.
The 1970s also saw the invention of acoustic impedance derivation through recursive
inversion, which was the first attribute method, other than velocity analysis, to forge a link
between seismic data and lithology (Lavergne and Willm, 1977; Lindseth, 1979). Sparse
spike inversion soon superseded recursive methods (Oldenburg et al., 1983). Amplitude-
variation-with-offset analysis arrived in the early 1980s and promised more lithological
information derived from prestack data (Ostrander, 1982). These methods launched
seismic attributes on the path toward quantifying lithological properties, though progress
was to prove slower and more frustrating than envisaged.
Proliferation and disillusionment

If a few attributes are good, a lot of them must be better. Faithful to this logic, research-
ers in the 1980s invented a host of new seismic attributes.
Root-mean-square amplitude, zero-crossing frequency, and cosine of the phase were
developed in the early 1980s as more comprehensible substitutes for complex trace attri-
butes. These were followed by less comprehensible attributes, including the perigram,
arc length, energy half-time, instantaneous bandwidth, and various empirical amplitude
and frequency measures.
The rise in the mid-1980s of 3D seismic data and computer systems for seismic
interpretation led to the invention of attribute maps. The first attribute maps were simple
extractions from seismic volumes (Denham and Nelson, 1986); horizon attributes and
horizon-guided interval attributes came a few years later (Dalley et al., 1989; Dorn and
Fisher, 1989; Sonneland et al., 1989; Rijks and Jauffred, 1991; Bahorich and Bridges,
1992; Hoetz and Watters, 1992). Attribute maps provided an easy yet powerful new way
to view seismic data, and they quickly became the most common form of presentation
for seismic attributes.
As attribute analysis matured and evolved from exotic novelty to prosaic tool,
nagging questions arose (Figure 7). How do you use them? Which are best? What do
they mean? Especially, what do they mean? Inventing attributes proved easier than explain-
ing what they meant. Too many attributes had elegant mathematical definitions but obscure
geological meanings. Seismic interpreters could only guess what information they
conveyed.
In the quest for meaning, geophysicists related seismic attributes to wavelet properties.
Robertson and Nogami (1984) and Saha (1987) showed that the instantaneous frequency
at the peak of a wavelet equals the wavelet’s average Fourier spectral frequency weighted
by its amplitude spectrum. Bodine (1984) used similar relations to develop “response
Figure 7. The earliest published example of a complex trace attribute is this advertisement from
the June 1977 issue of GEOPHYSICS (Seiscom Delta, 1977). It introduces instantaneous frequency
merely to promote a new method of deconvolution.
attributes.” Robertson and Fisher (1988) championed response attributes as time-variant

measures of the seismic wavelet, spurring their wide adoption.
These efforts to make sense of seismic attributes gave mathematics in place of the geo-
logical insights that interpreters needed. Further, the hopeful claims made for response
attributes did not hold in general (White, 1991). Questions became doubts; enthusiasm
yielded to skepticism. By the mid-1980s, attributes had lost their initial gloss of scientific
respectability. Roy Lindseth (1982, p. 9.15) observed, “except for amplitude, they have
never become very popular, nor are they used extensively in interpretation. The reason
for this seems to lie in the fact that most of them cannot be tied directly to geology.”
Regarding complex seismic trace attributes, Hatton et al. (1986, p. 25) opined, “this
concept is a little difficult to grasp intuitively. . . . While these functions do provide alterna-
tive and sometimes valuable clues in the interpretation of seismic data, cf. Taner et al.
(1979), it is probably fair to say that their usage has not been as widespread as it might
have been due to their somewhat esoteric nature.” Yilmaz (1987, p. 484) cautiously
wrote, “The instantaneous frequency may have a high degree of variation, which may be
related to stratigraphy. However, it also may be difficult to interpret all this variation.”
Robertson and Fisher (1988, p. 23) added, “The mix of meaningful and meaningless
values is probably the major factor that has frustrated interpreters looking for physical sig-
nificance in the actual numbers on attribute sections.”
Disillusionment with seismic attributes spread and was not dispelled until the introduc-
tion of 3D discontinuity attributes.
Discontinuity and attribute revival

Picou and Utzmann (1962) developed the first multidimensional seismic attributes,
including a continuity attribute and vector attributes. Far ahead of their time, it was only
in the 1980s that their ideas became practical, enabled by increasing computer power
and driven by the growing requirements of seismic data interpretation. Diverse 2D con-
tinuity and dip measures appeared first, but aroused little interest (Waters, 1981, p. 262;
Conticini, 1984; Scheuer and Oldenburg, 1988; Milkereit and Spencer, 1990; Claerbout,
1992). Multidimensional attributes remained a sideshow until Bahorich and Farmer
(1995) recast 2D continuity as 3D discontinuity and displayed it along time slices and
horizons to reveal faults, salt domes, and channels with a clarity never seen before. Dis-
continuity was an immediate sensation. It aroused such tremendous excitement in the
industry that it revitalized all of attribute analysis. A multitude of competing measures
soon emerged, bearing a multitude of competing names. These were followed by measures
for azimuth, dip, curvature, parallelism, and other 3D seismic properties (Oliveros and
Radovich, 1997; Marfurt et al., 1998; Taner, 2000; Randen et al., 2000).
Spectral decomposition arrived two years later with an insightful new way to visualize
channels and thin beds in map view (Gridley and Partyka, 1997). Coupled with tuning
analysis, it provided a means to estimate thin-bed thicknesses, a small but consequential
step toward making seismic attributes more quantitative.
Spectral decomposition was adopted almost as quickly as discontinuity attributes had
been. These methods owed much of their success to clear and useful geological meaning.
Driven by improved methods and better understanding, prestack attributes steadily
gained wider acceptance. Key developments include the fluid factor attribute (Smith and
Gidlow, 1987), the recognition of AVO classes (Rutherford and Williams, 1989; Castagna
et al., 1998), attributes expressed in terms of Lamé parameters (Goodway et al., 1997), and
the concept of elastic impedance (Connolly, 1999). These advances broadened the appeal
of prestack attributes by providing a framework for their interpretation, and by linking them
more closely to rock properties.
Multiattribute analysis
The potential of multiattribute analysis has been recognized since the earliest days
of attribute analysis. In 1977, Taner and Sheriff wrote, “More information is obtainable
by using a set of displays of different attributes synergetically [sic] than by interpreting
them individually.” A few years later, ambitious researchers were already imagining
sophisticated multiattribute workflows to identify seismic facies automatically (de Fig-
ueiredo, 1982; Sonneland, 1983; Conticini, 1984). Sonneland (1983, p. 549) described
the ultimate goal: “Finally, automated interpretation techniques might release the
interpreter from tedious parts of the interpretation and thereby contribute to faster
turnaround.”
It was only in the 1990s that multiattribute analysis found routine practical application
in seismic interpretation. Supervised methods of pattern recognition were favored at first
because they could be trained to produce results with inherent geological significance.
A typical supervised method involved a neural network, trained with seismic attributes at
well locations, to predict sand thickness, porosity, or some other reservoir property (John-
ston, 1993; Hampson et al., 2001). Unfortunately, training entails considerable effort and
experience. Unsupervised methods require no training and therefore are much easier to
develop and apply. Following the popular success of unsupervised waveform classification
in 1997 (Addy, 1997; Poupon et al., 1999), unsupervised classification was applied to
seismic attributes (De Groot, 1999; Walls et al., 2002). Introducing geological meaning
into unsupervised classification remains challenging.
It was soon recognized that multiattribute analysis for predicting reservoir properties
can produce spurious correlations between attributes and the reservoir geology if the attri-
butes are not related directly to the properties (Kalkomey, 1997). This problem continues to
challenge multiattribute analysis.
Recent developments
Development in seismic attributes is driven by the quest for greater resolution, the
desire to facilitate rapid interpretation of large volumes, and the need for quantitative, geo-
logically meaningful measures.
Automatic fault extraction is perhaps the greatest recent advance in seismic attri-
butes (Pedersen et al., 2002). It promises to pick large faults much faster than can be
picked by hand, as well as to pick many more small faults than is currently practicable.
Automatic fault extraction is destined to become a routine tool of seismic interpre-
tation like automatic horizon picking. Difficult problems must yet be overcome, such as
ensuring fault ensembles are geologically reasonable or automatically recognizing listric
faults.
A wide variety of volume curvature attributes have been invented, which complement
discontinuity attributes by revealing different kinds of structural detail (Hakami et al.,
2004; Al-Dossary and Marfurt, 2006). The chief challenge now is to better understand
their geological meaning.
Inversion methods continue to increase in sophistication, resolution, and popularity.
New methods of spectral decomposition are accompanied increasingly by claims of
improved resolution in both the time and frequency domains (Puryear et al., 2012).
However, most improvements seem largely incremental.
Progress in pattern recognition of seismic data is slow. The prevalence of false corre-
lations, coupled with an increased awareness of other inherent difficulties, has greatly tem-
pered ardor. The goal of automated seismic facies analysis with clear geological
significance remains as alluring as ever, and as stubbornly distant.
Intriguing new methods of automatic full-volume flattening promise a powerful new
way to look at stratigraphy. They flatten every reflection in a seismic volume, in effect
transforming the vertical axis from record time to “relative geological time,” thereby sim-
plifying stratigraphic study (Stark, 2004; Lomask et al., 2006; Wu and Hale, 2014). Results
are encouraging, though practical application lags.
Structurally guided computations are increasingly important in seismic attribute analy-
sis. Originally developed for discontinuity attributes, they are applied now in other
structural attributes, as well as in stratigraphic attributes, 3D filters, and pattern recognition

algorithms. Structurally guided computations will soon become routine and widespread.
Following the example of Goodway et al. (1997), a major trend in prestack attribute
analysis is the development of more attributes related to rock properties. These include
measures for “fracability” and brittleness to aid exploration in shale (Treadgold et al.,
2011; Sharma and Chopra, 2012).
This concludes the history of seismic attributes. I now explore their theory and appli-
cation, starting with the most basic ideas, those of attribute maps and statistical attributes.
Chapter 3
Attribute Maps and Interval Attributes
Introduction
Seismic attributes can be presented as a map, as a cross section, or as a volume. Maps
are the most common form of presentation because they are easy to generate and con-
venient to interpret.
Attribute maps derive from seismic horizons. A seismic horizon is an interpreted map
that represents the time or depth of a stratigraphic surface observed in seismic data. On
clean uncomplicated data, horizons follow continuous and unambiguous seismic reflec-
tions characterized by uniform phase, amplitude, and waveform. On noisy or structurally
complex data, horizon interpretation is challenging, and consequently a horizon may
follow trends that are only suggested by the data. Horizons inherit characteristics from
the seismic data, and, in turn, attribute maps inherit characteristics from the horizons.
Attribute maps record horizon attributes or horizon-guided interval attributes. Horizon
attributes derive directly from seismic horizons and are limited to structural attributes.
Interval attributes are numerous and varied and measure statistics of the seismic data or
basic geophysical properties, such as amplitude, frequency, and bandwidth. Interval attri-
butes become trace attributes by running the interval down the seismic trace.
This chapter presents horizon attributes and interval attributes. It reviews the structural
attributes that derive directly from horizons, and introduces interval attributes that quantify
various data statistics as well as amplitude, frequency, and bandwidth. It considers both
map interval attributes and trace interval attributes.
Waveform classification and spectral decomposition also produce attribute maps.
Their methods and interpretation are involved and require separate discussion, which is
given in Chapter 7.
Horizon attributes
Horizon attributes quantify structural properties of interpreted seismic horizons: slope,
dip, azimuth, discontinuity, and curvature. Because horizons are often smoothed, horizon
attributes tend to be smoother than the same attributes derived directly from the seismic
data. Horizon attributes highlight faults, fractures, channel edges, anticlines, basins, and
other structural features. They also reveal problems in horizon interpretation caused by
pinch-outs, noise, or mispicks, particularly across faults.
In the following, z(x, y), or simply z, represents a horizon in time or depth as a function
of map coordinates x and y.
27
Slope, dip, and azimuth

Horizon dip and azimuth have units of degrees and derive from the horizon slopes in
the x and y directions. Horizon slope in the x direction, px, is defined by
∂z
px = . (1)
∂x
The corresponding slope in the y direction, py, is
∂z
py = . (2)
∂y
Horizon slope magnitude p combines slopes px and py according to

p = p2x + p2y . (3)
For horizons in depth, slope is dimensionless and horizon dip g is
g = arctan p. (4)
For horizons in time, slope is an inverse velocity or slowness with units of milliseconds per
meter, and dip must be estimated given a conversion velocity.
Horizon azimuth f is the downdip direction of the horizon in degrees from north, and is
given by

px
f = arctan + fo , (5)
py
where fo refers the survey to true north (refer to Figure 8 in Chapter 1). This holds whether
the horizon is in time or in depth.
Dip and azimuth are combined into a single display either through a 2D color bar or,
more intuitively, through shaded relief (Dalley et al., 1989).
Edge detection
Edge detection is a standard method of image processing that changes pictures into line
drawings. The lines represent boundaries or “edges” between distinct regions of the picture.
Applied to seismic horizons, edge detection highlights abrupt changes in horizon time or
depth, which indicate faults, channel edges, cycle skipping, or problems due to noise.
Most methods for horizon edge detection employ gradients, such as Sobel operators
(Mlsna and Rodriguez, 2009). They are applied as two orthogonal convolutional operators
to compute the gradients in the x and y directions. Combining these two gradients as a mag-
nitude produces the edge map.
An alternative measure for edge detection is defined as the difference between a
horizon and a smoothed version of itself. The smoothing operator can be linear or non-
linear. The results are comparable to those of gradient methods.
Chapter 3: Attribute Maps and Interval Attributes 29
Curvature
Horizon curvature attributes measure properties of the local curvature to reveal faults,
flexures, anticlines, and synclines. Because curvature attributes are interpreted qualitat-
ively, the units are ignored and the same mathematics are employed whether the horizon
is in time or depth.
Curvature properties are defined in terms of the coefficients of a second-order poly-
nomial in x and y that locally fits the horizon z(x, y):
z(x, y) = Cxx x2 + Cyy y2 + Cxy xy + Cx x + Cy y + Co (6)
(Roberts, 2001). Coefficient Co is the time or depth of the horizon at x ¼ 0, y ¼ 0. It has

no influence on curvature and is ignored. The other coefficients derive from first- and
second-order partial derivatives of horizon time or depth z with respect to x and y,
and usually are expressed in terms of horizon slopes px and py. The equations for the coef-
ficients are
1 ∂2 z 1 ∂px
Cxx = = , (7)
2 ∂x2 2 ∂x
1 ∂2 z 1 ∂py
Cyy = = , (8)
2 ∂y2 2 ∂y
∂2 z ∂px ∂py
Cxy = = = , (9)
∂x∂y ∂y ∂x
∂z
Cx = = px , and (10)
∂x
∂z
Cy = = py . (11)
∂y
In most applications, the first and second partial

derivatives in the formulas for the curvature coefficients
usually are approximated by three point-difference
equations. For three points in a line, z1, z2, and z3, with
Figure 1. Grid of nine points for
uniform spacing Ds, the first derivative is approximated
estimating curvature on a horizon,
with times or depths z1 through z9. by (z3 – z1)/2Ds, and the second derivative by (z3 –
Adjacent points are separated by 2z2 + z1)/Ds 2. For a set of nine horizon points arranged
distance Dx in the x direction and in a 3 × 3 grid with uniform spacing Dx and Dy in the x
Dy in the y direction. Curvature is and y directions (Figure 1), the difference equations are
assigned to the center point. modified to incorporate simple averages over the grid.
The approximate formulas for the curvature coefficients become

z1 + z3 + z4 + z6 + z7 + z9 z2 + z5 + z8
Cxx = − , (12)
6Dx2 3Dx2
z1 + z2 + z3 + z7 + z8 + z9 z4 + z5 + z6
Cyy = − , (13)
6Dy2 3Dy2
z1 + z9 − z3 − z7
Cxy = , (14)
4DxDy
z3 + z6 + z9 − z1 − z4 − z7
Cx = , and (15)
6Dx
z7 + z8 + z9 − z1 − z2 − z3
Cy = . (16)
6Dy
The two fundamental curvature properties are mean curvature kmean and Gaussian
curvature kgauss. They are expressed terms of the curvature coefficients as
Cxx (1 + Cy2 ) + Cyy (1 + Cx2 ) − Cxy Cx Cy

kmean = , (17)
3
1 + Cx2 + Cy2
and
4Cxx Cyy − Cxy
2
kgauss = . (18)
(1 + Cx2 + Cy2 )2
Maximum and minimum curvatures, kmin and kmax, are expressed in terms of kmean and
kgauss as

kmax = kmean + k2mean − kgauss , and (19)

kmin = kmean − k2mean − kgauss . (20)
Rearranging these equations, kmean is the average of kmin and kmax,

kmax + kmin
kmean = , (21)
2
and kgauss is their product,
kgauss = kmax kmin . (22)
Most positive curvature kpos is given by

kpos = Cxx + Cyy + (Cxx − Cyy )2 + Cxy
2 , (23)
and most negative curvature kneg by

kneg = Cxx + Cyy − (Cxx − Cyy )2 + Cxy

2 . (24)
Other curvature properties that are derived from the curvature coefficients include dip
curvature, strike curvature, curvedness, and shape index. Minimum, maximum, most posi-
tive, and most negative curvatures have shown the most value as attributes.
Where horizon dip is small and can be neglected, the equations for mean and Gaussian
curvatures reduce to the zero dip approximations,
kmean = Cxx + Cyy , and (25)

kgauss = 4Cxx Cyy − Cxy
2
. (26)
Maximum curvature reduces to most positive curvature, and minimum curvature reduces to
most negative curvature, but the equations for most positive and most negative curvatures
remain unchanged.
Examples of horizon attributes

Figure 2 shows a seismic horizon and its structural attributes. Shaded relief combines
the information of dip and azimuth to create a more natural view of the horizon structure.
Dip highlights discontinuities well, whether due to geological structure or problems in the
horizon interpretation. Edge detection and curvature attributes likewise highlight disconti-
nuities; the discontinuities are naturally sharper when derived from unsmoothed horizons.
Azimuth shows regions in the data that have similar reflection orientation, and it highlights
regions with anomalous orientation. Problems in horizon interpretation become evident on
attribute maps. Here the problems are due to difficulties tracking the reflection through
pinch-outs.
Interval attributes
Interval attributes measure seismic properties in short intervals on seismic traces. Most
interval attributes are mathematical attributes that record statistics of the data. These stat-
istics typically represent averages, variances, selections, totals, and counts. Some interval
attributes record statistics of the autocorrelation or Fourier transform of the data. The most
useful interval attributes measure amplitude, frequency, or bandwidth.
Interval attributes are either map attributes or trace attributes. For trace attributes, the
interval is a window that runs down the seismic trace. The window typically is tapered to
reduce the ill effects of Gibbs’ phenomenon (Appendix A), and its length sets the resolution
of the attribute. For map attributes, the interval is defined either as a constant vertical length
about a guide horizon, or as the region bounded by two horizons (Figure 3). An interval that
does not have constant length is unsuitable for attributes that represent counts or sums,
such as number of peaks or total energy. Intervals are rarely tapered for map attributes
because the ill effects of Gibbs’ phenomenon are not evident on maps. As a rule, the
a) b)
c) d)
e) f)
g) h)
Figure 2. A smoothed seismic horizon in time and its attributes. (a) Horizon with 20-ms contours.
(b) Shaded relief. (c) Dip. (d) Azimuth. (e) Edge detection of horizon before smoothing. (f ) Edge
detection of smoothed horizon. (g) Maximum curvature. (h) Minimum curvature. Data from the
Taranaki Basin, offshore New Zealand.
a) x
b) x
Top horizon
tl
Guide horizon
Bottom horizon
t t
Figure 3. (a) Horizon interval defined as a constant time length tl about a guide horizon.
(b) Horizon interval defined as the region between two horizons.
x Figure 4. For attribute maps, intervals should

be interpolated so that samples fall on the guide
horizon. Here, the blue zone represents a
horizon-guided interval. Vertical lines are
seismic traces and horizontal lines are original
discrete sample times; black dots represent trace
Guide horizon samples that must be selected for attribute
computation. For the central trace, the required
samples coincide with existing trace samples, so
interpolation is not needed. For the two other
traces, the required samples must be obtained
through interpolation.
interval is centered on a zone or reflection of interest with its length a little longer than the
zone. If the zone of interest is bounded by zones of markedly different character, the inter-
val should exclude adjacent zones.
Guide horizons should be clean and free of defects. For an interval defined by a single
horizon, the seismic data should be interpolated so that samples fall exactly on the horizon
(Figure 4). This is particularly important for short analysis windows guided by a single
horizon. Neglecting interpolation and selecting existing samples instead produces stair-
stepped artifacts that parallel horizon isochrons (Figure 5).
Statistical measures
Statistical measures are the most common interval attributes, and they take the form of
largest or smallest values, simple counts or ratios, or basic distribution statistics. Statistical
measures are mathematical attributes and lack direct connection with geology or
a) b) c)
Figure 5. Illustration of the role of interpolation in computing horizon interval attributes.

(a) Interpreted horizon; blue is 350 ms deeper than red. The vertical stripes are an acquisition
footprint. (b) Root-mean-square amplitude computed in a window of five samples with
interpolation, and (c) without interpolation; red are strong amplitudes and blue are weak. Vertical
striping in the attribute is inherited from the horizon. When computed without interpolation, the
attribute also exhibits stripes that parallel horizon isochrons. Such stripes are more closely spaced
the steeper the dip. They become more pronounced with shorter analysis intervals.
geophysics. Many more statistical measures have been implemented than can be reviewed
here; most have obscure value.
The largest or smallest sample values in a trace interval are common statistical
measures. They are prone to outliers and tend to be noisy; apply them only where required.
The interval must be chosen carefully and usually is kept small. With good quality data,
these attributes produce sharper maps than attributes based on averages.
Maximum peak and maximum trough attributes are similar to the largest and smallest
values, but usually they are interpolated to determine maxima between samples. Applied
directly to standard seismic data, attributes that record maxima are more appropriate for
short horizon intervals centered on reflection peaks. Similarly, attributes that record
minima are more appropriate for short horizon intervals centered on reflection troughs.
Distribution statistics record how the seismic trace values vary in amplitude. Distri-
bution statistics include the mean, variance, skew, and kurtosis. Some attribute software
also offers the geometric mean and harmonic mean. None of these measures are well
suited for standard seismic data, though they are appropriate for seismic velocities, impe-
dances, and similar data. Avoid distribution statistics.
Attributes that count the number of peaks or troughs in an interval act as crude and
inferior measures of relative frequency. They are sensitive to small changes in the position
or length of the interval, so they should be applied only where counts are of direct interest,
such as in tracking channels or pinch-outs.
Amplitude and energy

Amplitude and energy interval attributes are numerous, chief of which are average
absolute amplitude, root-mean-square (rms) amplitude, and average energy. Average
absolute amplitude xa is the mean of the absolute values of the trace:

1 N
xa = |xn |. (27)
N n=1
Root-mean-square amplitude xrms is the square root of the average of the squared trace
values, or trace power:

N
1
xrms = x2 . (28)
N n=1 n
Average energy Ea is the square of the rms amplitude:
1 N
Ea = x2rms = x2 . (29)
N n=1 n
It may be preferred because it exhibits more contrast. For zero mean seismic data, average
energy equals the variance. Total energy E is related to average energy by E ¼ NEa.
Average peak value and average trough value are common map interval attributes, but
they have little value as trace attributes. Average peak value is most appropriate for a
narrow interval centered on a peak, and average trough value is most appropriate for a
narrow interval centered on a trough. For intervals that encompass both peaks and
troughs, rms amplitude is usually more appropriate. Apply these attributes only when
peaks or troughs are of specific interest.
Average reflection strength is offered occasionally as a map attribute. It serves as well
as rms amplitude, though it is less convenient to compute. Reflection strength is a complex
trace attribute, and it is explained in Chapter 4.
Frequency
Frequency interval attributes quantify various frequency properties, and include
zero-crossing frequency, average Fourier spectral frequency, rms spectral frequency,
rms frequency derived from autocorrelation, and peak spectral frequency.
Zero-crossing frequency fc estimates the average frequency of a waveform as half
the number of zero crossings Nc in a time interval divided by the interval length, Tl:
Nc
fc = . (30)
2Tl
Zero-crossing frequency is constrained to positive frequencies from zero to Nyquist fre-

quency. It is stable in long windows but noisy in short windows. Its frequency estimates
are crude and less reliable than those of competing measures. Avoid zero-crossing
frequency.
Average frequency fa is sometimes computed as the average Fourier spectral frequency

weighted by the power spectrum A 2( f ). Defining total spectral energy E as
E= A2 ( f )df , (31)
0
average spectral frequency is expressed as

1
1
fa = f A2 ( f )df = k f lf ; (32)
E
0
the brackets with subscript f denote weighted averaging in the frequency domain. As it
happens, average spectral frequency equals weighted average instantaneous frequency,
which is a complex trace attribute (Cohen, 1995, p. 9; Chapter 4).
Root-mean-square spectral frequency frms is the square root of the average squared
spectral frequency weighted by the power spectrum:

1
1
2
frms = f 2 A2 ( f )df = k f 2 lf . (33)
E
0
Root-mean-square frequency is always equal to or greater than average frequency. It serves

much the same purpose as average frequency, and it also has an equivalent complex trace
attribute.
An alternative and simpler estimate of rms spectral frequency is given in terms of
the autocorrelation of the seismic data at zero lag time. In terms of the continuous autocor-
relation function f(t), rms frequency is given by

1 f′′ (0)
frms = − , (34)
2p f(0)
where f′′ indicates the second derivative in time. This formula is derived noting that the
autocorrelation equals the inverse Fourier transform of the power spectrum, then taking
the second derivative in time of the transform relation and setting lag time t to 0. For dis-
crete seismic data xn, the autocorrelation fm is given by

N
fm = xn xn+m . (35)
n=1
Setting m to 0 yields the autocorrelation value at zero lag, f0, and setting it to 1 yields the
value at unit lag, f1. Frequency frms is then estimated as

1 f0 − f1
frms ≈ , (36)
pT 2f0
where T is the time sample period. This approximation is surprisingly accurate.

Peak frequency refers to the frequency of the strongest spectral component. This
depends on how spectral power is determined. Derived from Fourier spectra in a short
time window, peak frequency requires interpolation in much the same way as maximum
peak amplitude. Peak spectral frequency is noisier than average spectral frequency, and
lacks clear ties to geology.
Peak frequency is sometimes computed through maximum entropy decomposition,
also called Burg spectral decomposition or all poles spectral estimation (Claerbout,
1985, p. 133; Press et al., 1992 [reprinted 1995], p. 572). Maximum entropy spectral
decomposition is theoretically better than Fourier analysis for estimation of spectra of
short time series, but it is computationally expensive and does not generate more interpret-
able results.
Bandwidth
There are few bandwidth attributes, chief of which are effective bandwidth and spectral
bandwidth.
Effective bandwidth be is an empirical measure defined as the scaled ratio of the
zero-lag autocorrelation divided by the sum of the absolute values of the autocorrelation
at all lags:
f0
be = N−1 , (37)
2T m=−N+1 |fm |
where T is the sample period. This measure has units of hertz and a maximum value of
Nyquist frequency. The square of the autocorrelation values is sometimes used in place
of absolute values.
The mathematics of effective bandwidth look odd, but the underlying idea is easy to
grasp. The autocorrelation of broad bandwidth data is sharply peaked, with much of the
autocorrelation energy in the center value, fo. In this case, effective bandwidth is large
and approaches Nyquist frequency, 1/2T. In contrast, the autocorrelation of narrow band-
width data is spread out, so that little of its energy lies in the center value. In this case, effec-
tive bandwidth is small and approaches 0. Effective bandwidth tends to be on the order of the
average frequency.
Spectral bandwidth fb is commonly defined as the standard deviation of the spectral
frequency about the average spectral frequency fa:

1
1
fb =
2
( f − fa )2 A2 ( f )df = k( f − fa )2 lf . (38)
E
0
It is related to average spectral frequency and rms spectral frequency according to

fb2 = frms
2
− fa2 (39)
(Figure 6). Like average frequency and rms frequency, bandwidth can be computed in both
the frequency and time domains.
Arc length
Arc length is a mathematical attribute described as

the length of a seismic wiggle trace measured along its
wiggles and normalized by its time length (Figure 7).
This description makes sense on wiggle-trace plots,
but it is flawed as a general definition because amplitude
fa frms and time have different units. It suggests that arc length L
be approximated as
N−1

1
L= (xn+1 − xn ) + T ,
2 2 (40)
(N − 1)T n=1
where T is the sample period. The terms involving T

contribute no information. Discarding them leaves a
simpler formula with consistent units:
fb
1 N−1
L= |xn+1 − xn |. (41)
Figure 6. Average spectral N − 1 n=1
frequency fa, bandwidth fb, and
rms spectral frequency frms are Arc length increases with both amplitude and
related as the lengths of a right frequency. High arc length could indicate strongly
triangle. reflecting, moderately spaced bedding or moderately
reflecting, thinly spaced bedding. The interpretation of
arc length is thus ambiguous. It tends to be driven more by amplitude than by frequency
and typically resembles amplitude attributes. Avoid arc length.
Energy half-time
Energy half-time Eht is a relative measure of where the seismic energy is concentrated
in an analysis interval. It is defined as the time tc of the center of gravity of the trace power
in a time interval expressed as a percentage of the interval length tl:
tc
Eht = 100% · . (42)
tl
Time tc is the weighted average
N
tn x2n
tc = n=1
N , (43)
2
n=1 xn
where tn is time with respect to the start of the interval.

An energy half-time less than 50% indicates that the trace energy is more in the top half
of the interval, a value greater than 50% indicates that the energy is more in the bottom half
of the interval, and a value about 50% indicates that the energy is evenly distributed or con-
centrated near the interval center. In principle, energy half-time can track the extent of a
highly reflective unit juxtaposed against a weakly reflective unit, which could represent a
sand body bounded by shale.
0
For this purpose, the interval
should include only one domi-

nant energy zone. However,
energy half-time is more a
1
measure of relative amplitude
Amplitude change, with resolution set by
the interval length. Viewed in
this light, a value of less than
50% indicates that the ampli-
2
Time
tudes are decreasing with time,

and more than 50% indicates
that they are increasing.
Energy half-time occasionally
3
reveals features not seen with
other attributes.
Applied as a trace attribute,
energy half-time highlights
4
boundaries between reflections

and resembles a smooth version
of relative amplitude change,
which is introduced in Chapter
4. This becomes clear when it
5
is scaled to have zero mean and

displayed in monochrome.
Gallery of interval
6
attributes
Figure 7. The arc length of a waveform is said to be similar
to the length of a wavy piece of string, which can be measured Figure 8 compares ampli-
tude attribute maps. The maps
with a ruler after straightening. This idea makes little sense
because waveform amplitude and time have different units. are derived in an interval cen-
tered along the seismic horizon
of Figure 2 shifted down 100 ms to image channels where horizon interpretation is
more difficult. The amplitude attributes all look similar, and there is little reason to apply
more than one. For routine investigation of bright events, rms amplitude is best. Energy
has the same information as rms amplitude, but with increased contrast. Because the data
are fairly clean, the maximum and minimum values are only a little noisier than other
attributes. Arc length is included here because it closely resembles an amplitude attribute,
but it is also influenced somewhat by frequency.
Energy half-time, shown in Figure 9, indicates where in the interval the energy is
clustered, whether shallow, deep, or intermediate. It is best with clean data that have
well-defined features in a carefully defined interval. It is difficult to interpret, and conse-
quently has limited value as a map attribute.
a) b)
c) d)
High
Amplitude
e) f)
Low
g) h)
10 km
Figure 8. Amplitude attributes derived in a 40-ms (11 samples) window along the seismic horizon
of Figure 2 shifted down 100 ms. (a) Original seismic data extracted along the horizon; orange is
positive, blue is negative. (b) Root-mean-square amplitude. (c) Average absolute amplitude.
(d) Average peak amplitude. (e) Energy. (f) Maximum amplitude. (g) Minimum amplitude.
(h) Arc length. The color bar is reversed on the minimum amplitude map for easier comparison.
Data from the Taranaki Basin, offshore New Zealand.
Figure 10 compares amplitude attributes and energy half-time computed as trace attri-
butes on a seismic line. Again, the amplitude attributes are comparable. Displayed on ver-
tical sections in grayscale, energy half-time appears illuminated to highlight boundaries
between seismic reflections.
Shallow
Deep
10 km
Figure 9. Energy half-time for the same horizon and interval as the amplitude attributes shown in
Figure 8. Blue and green indicate that the concentration of energy in the interval is shallow, yellow
indicates that it is centered, and orange and red indicate that it is deep.
a) b)
c) d) High
Amplitude
Low
e) f)
Figure 10. Amplitude attributes derived from a seismic line offshore Australia. (a) Original
seismic data. (b) Root-mean-square amplitude. (c) Average absolute amplitude. (d) Average energy.
(e) Arc length. (f) Energy half-time. The attributes are derived in a 60-ms (15 samples) window.
Figure 11 compares four frequency attribute maps. They differ in detail and scale, but
they show much the same overall features. Zero-crossing frequency is the least satisfactory
measure, and average spectral frequency is arguably the best. Frequency attributes are
applied in stratigraphic analysis and in attenuation studies.
Effective bandwidth and spectral bandwidth produce roughly comparable images,
as seen in Figure 12, though their scales differ considerably. Spectral bandwidth is
a) b)
High
Frequency
c) d)
Low
Figure 11. Frequency attributes derived from the seismic line shown in Figure 10.
(a) Zero-crossing frequency, scaled from 15 to 60 Hz. (b) Autocorrelation rms frequency, scaled
from 15 to 60 Hz. (c) Spectral average frequency, scaled from 5 to 65 Hz. (d) Spectral peak
frequency, scaled from 0 to 60 Hz. The attributes are derived in a 60-ms (15 samples) window.
a) b)
High
Frequency
Low
Figure 12. Bandwidth attributes derived from the seismic line shown in Figure 10. (a) Effective
bandwidth, scaled from 15 to 50 Hz. (b) Spectral bandwidth, scaled from 5 to 35 Hz. Both attributes
are derived in a 60-ms (15 samples) window.
preferable because it provides a smoother image and is easier to understand and relate to
frequency attributes.
Summary
Maps are the most common form of seismic attributes because they are easy to apply
and convenient to interpret. Attribute maps derive either directly from interpreted seismic
horizons, or from seismic data in horizon-guided intervals. Attribute maps that derive from
seismic horizons are necessarily structural. They are strongly influenced by the quality of
the horizon interpretation, so they highlight flaws in the horizon as well as structural details.
Attributes derived in an interval are called interval attributes. Interval attributes are the
most numerous type of seismic attributes, and are common both as map attributes and as
trace attributes. Many interval attributes measure statistics of the seismic data such as
averages, variances, counts, and extrema. Others record statistics of the autocorrelation
or Fourier transform of the data. The most useful interval attributes measure amplitude, fre-
quency, bandwidth, and other geophysical properties. Interval attributes are numerous in
part because they are easy to invent. Many are duplicates, or have narrow utility, or lack
useful meaning. Those that count occurrences, such as number of peaks, or attributes
that measure extrema, such as the largest value, are inherently noisy and should be
employed only when specifically needed.
Interval attributes are one-dimensional in that they measure properties in vertical inter-
vals of data. Two- and three-dimensional seismic properties are more difficult to quantify
and are rarely offered as map attributes in commercial software. This is an area for future
improvement.
Frequency and bandwidth attributes are readily derived through autocorrelations and
spectral transforms. However, they are more commonly derived through the complex
seismic trace, the topic of the next chapter.

Chapter 4
Complex Seismic Trace Analysis
Introduction
Complex seismic trace analysis is the most widely applied and most versatile method
for computing seismic attributes. In its original 1D form, complex trace analysis
operates on individual traces to produce trace attributes (Taner et al., 1979). Extended
to 3D, complex trace analysis operates across traces as well as down to produce volume
attributes.
Complex seismic trace analysis treats a seismic trace as the product of two separable
attributes, instantaneous amplitude and cosine of the instantaneous phase. All other
complex trace attributes derive from amplitude and phase, notably frequency, bandwidth,
dip, azimuth, and relative amplitude change. Instantaneous frequency appears noisy and
is prone to spikes; this is characteristic of instantaneous attributes that involve differen-
tiation. Nonlinear filtering or weighted averaging reduces the apparent noise and
removes spikes. Filtered or averaged attributes gain intuitive meaning as time-variant
Fourier spectral averages.
This chapter reviews the basic ideas and attributes of complex seismic trace analysis. It
investigates the nature of the spikes in instantaneous attributes and their removal by
response attributes and weighted average attributes. It introduces the phase and group
wavenumber vectors to extend 1D complex trace analysis to 3D. Throughout, it emphasizes
real mathematics and simple illustrations to make clear that the method is not inherently
complex, either mathematically or conceptually.
1D complex seismic trace analysis

The essence of complex seismic trace analysis lies in the separation of amplitude infor-
mation from phase information in seismic data. This idea is inspired by the example of
Fourier analysis.
Fourier frequency transforms are represented either in Cartesian form as cosine and
sine spectra, or in polar form as amplitude and phase spectra. Cosine and sine spectra
are natural products of the Fourier transform, but amplitude and phase spectra are better
suited for physical studies. Cosine and sine spectra X( f ) and Y( f ) are related to amplitude
and phase spectra A( f ) and u( f ) by
X( f ) = A( f ) cos u ( f ), and (1)
Y( f ) = A( f ) sin u ( f ). (2)
45
Just as Fourier transforms are represented in terms of amplitude and phase spectra,
so too can seismic traces be represented with analogous amplitude and phase attributes.
To accomplish this, complex trace analysis introduces a quadrature trace derived from
the seismic trace (Cohen, 1995; Vakman, 1998). The seismic and quadrature traces are
analogous to the cosine and sine spectra in Fourier analysis, with the essential difference
that the quadrature trace depends on the real trace, whereas the cosine and sine spectra
are independent. A further difference, implied in the derivation of the quadrature trace,
is that generally the seismic trace should not have a dc component. Complex trace
amplitude and phase attributes prove as useful as their Fourier counterparts.
In the following development, a subscript i distinguishes instantaneous phase ui from
Fourier spectral phase u. A subscript i also distinguishes instantaneous frequency, band-
width, and root-mean-square (rms) frequency from their Fourier spectral counterparts.
Foundation
A seismic trace in time x(t) can be expressed as the product of two separate attributes,
instantaneous amplitude a(t) and the cosine of the instantaneous phase cos ui(t):
x(t) = a(t) cos ui (t). (3)
This equation does not define the amplitude and phase attributes because it relates a
single known function to two unknown functions. This difficulty is met by inventing
the quadrature trace y(t), expressed as the product of the instantaneous amplitude with
the sine of the instantaneous phase,
y(t) = a(t) sin ui (t), (4)
and defined as the Hilbert transform of the seismic trace,
y(t) = h(t) ∗ x(t), (5)

Quadrature trace
where h(t) is the Hilbert transform

y(t)
operator (Appendix B). The “magic”
a(t)
of complex trace analysis lies in this
step. It is suggested by the nature of
the Hilbert transform, which changes
θi (t) cosine waves into sine waves, effecting
Seismic a 2908 phase rotation of the seismic
x(t) trace
trace. Fortunately, complex trace attri-
butes can be understood without
delving into the esoteric mystery of
Figure 1. At an instant in time, the seismic and the quadrature trace.
quadrature traces represent a point in Cartesian Instantaneous amplitude and phase
coordinates, and the amplitude and phase attributes derive from the seismic and quadrature
represent the point in polar coordinates. traces through the Cartesian to polar
Chapter 4: Complex Seismic Trace Analysis 47
coordinate transform,

a(t) = x2 (t) + y2 (t), and (6)

y(t)
ui (t) = arctan (7)
x(t)
(Figure 1). Instantaneous amplitude and phase are the basic complex trace attributes from
which all others derive.
Amplitude
Instantaneous amplitude is an amplitude measure that is independent of the polarity
or overall phase of the seismic data. More commonly called “trace envelope” or “reflec-
tion strength,” it is possibly the most widely applied seismic attribute. At a given time,
it represents the magnitude of the sinusoid that best matches the seismic trace in a
small window about that time. More precisely, it equals the maximum value that the
trace can attain through constant phase rotation. Hence instantaneous amplitude is
invariant under a phase rotation of the trace and bounds all possible constant phase
rotations (Figure 2). Squared, it becomes instantaneous power, which is used as a
weighting function for average attributes. The logarithm of the instantaneous amplitude
is employed where amplitude contrasts are large, such as on ungained shot gathers.
Removing the dc or low-frequency components from instantaneous amplitude produces
the perigram, an amplitude-demodulated form of the seismic trace (Gelchinsky et al.,
1985).
Like all amplitude attributes, instantaneous amplitude highlights bright spots, dim
spots, and amplitude anomalies in general.
Phase
Instantaneous phase is an angular measure in degrees of relative position on a sinusoi-
dal waveform (Figure 3). At a given time, it represents the phase of the sinusoid that best
matches the seismic trace in a small window about that time. Thus instantaneous phase is
08 at peaks, 1808 at troughs, +908 at downgoing zero crossings, and 2908 at upgoing zero
crossings (Figure 4). Instantaneous phase is mathematically discontinuous and resembles
a sawtooth function, wrapping around 21808 and +1808. In color displays, the sawtooth
effect is remedied by employing a circular color bar.
Cosine of the phase is sometimes employed in place of instantaneous phase because it
has no discontinuities and looks like highly gained seismic data (Figure 5). Indeed, cosine
of the phase is the ultimate automatic gain in that it removes all amplitude information.
Seismic reflections are easier to follow on instantaneous phase or cosine of the phase
because they lack the amplitude contrasts that sometimes mask reflection continuity on
standard displays of seismic data.
a)
b)
c)
Figure 2. The trace envelope bounds all phase rotations of the seismic trace. At a given time, it
is the maximum value that a seismic trace can attain through phase rotation. (a) A seismic trace
(light line) and its envelope (heavy line). (b) The trace rotated in phase by 08, 458, 908, 1358,
1808, 2258, 2708, and 3158 (light lines). (c) Enlargement of the center portion of (b). The envelope
is shown as heavy lines above the rotated traces as well as reversed in sign below them.
Frequency
Instantaneous frequency fi(t) is the time derivative of the instantaneous phase scaled to
units of hertz:
1 d
fi (t) = ui (t). (8)
2p dt
At a given time, instantaneous frequency represents the frequency of the sinusoid that best
matches the seismic trace in a small window about that time.
Figure 3. (a) Instantaneous phase and (b) cosine of the phase (dark line) for the seismic trace of
Figure 2a (light line).
Figure 4. A seismic trace (blue) and its instantaneous phase (red). Instantaneous phase is 08 at
peaks, 1808 at troughs, +908 at downgoing zero crossings, and 2908 at upgoing zero crossings.
This accords with the phase sign convention (see Figure 4 in Chapter 1).
Equation 8 is unsuitable for computing instantaneous frequency because instantaneous

phase, being discontinuous, cannot be continuously differentiated. This equation could
be applied if discontinuities were removed first by unwrapping the phase, but this
process is cumbersome. Instead, the difficulty is sidestepped. Substituting the definition
of instantaneous phase into equation 8 gives

1 d y(t)
fi (t) = arctan , (9)
2p dt x(t)
Figure 5. Cosine of the phase acts like a strong automatic gain. (a) Cosine of the phase applied
to the seismic line of Figure 10a in Chapter 3. (b) Strong automatic gain applied to the same
data. The gain window is 28 ms long (7 samples). The two sections are nearly identical.
which expands to the practical formula for instantaneous frequency,
1 x(t)y′ (t) − x′ (t)y(t)

fi (t) = . (10)
2p x2 (t) + y2 (t)
Differentiation is accomplished through a derivative filter (Appendix C). Alternatively,

instantaneous frequency is computed through an efficient difference approximation
(Appendix D). Figure 6a shows the instantaneous frequency of a seismic trace.
Instantaneous frequency often looks noisy because differentiation boosts high frequen-
cies, which tend to be less coherent, and suppresses low frequencies, which tend to be more
coherent. Further, it is prone to large and confusing spikes. These problems are common to
instantaneous attributes that involve differentiation. They are remedied through nonlinear
filtering or weighted averaging.
Relative amplitude change and bandwidth

The amplitude counterpart to instantaneous frequency is relative amplitude change
s(t), defined as the time rate of change of the logarithm of the amplitude:
d a′ (t)
s(t) = ln a(t) = . (11)
dt a(t)
Like instantaneous frequency, relative amplitude change can be computed through an effi-
cient approximation (Appendix D). Unlike instantaneous frequency, it is a directional attri-
bute that appears illuminated along the time axis when displayed in monochrome. Relative
amplitude change reveals detail hidden in the amplitudes and highlights zones of reflection
interference, which cause spikes.
Figure 6. (a) Instantaneous frequency, (b) instantaneous relative amplitude change, and
(c) instantaneous bandwidth for the seismic trace of Figure 2a.
Instantaneous bandwidth bi(t) is the absolute value of relative amplitude change scaled
to have units of hertz:

1 1 d
bi (t) = |s(t)| = ln a(t). (12)
2p 2p dt
Instantaneous bandwidth is an instantaneous estimate of spectral bandwidth based solely

on amplitude change. In accord with intuition, it increases with relatively greater amplitude
change. In conflict with intuition, it vanishes at envelope peaks.
Figure 6b and 6c compares relative amplitude change and bandwidth for a
seismic trace.
Root-mean-square (rms) frequency

Instantaneous rms frequency firms(t) is the square root of

the sum of the squares of instantaneous frequency and instan-
taneous bandwidth:
|fi| firms
firms (t) = fi2 (t) + b2i (t). (13)
Instantaneous rms frequency is an estimate of the rms spectral

frequency at an instant in time. It is always positive and equal
ϕ to or greater than instantaneous frequency. Instantaneous rms
frequency and quality factor contain the same information as
bi
instantaneous frequency and bandwidth (Figure 7).
Figure 7. The relation
between instantaneous Quality factor
frequency fi, instantaneous
bandwidth bi, and
Instantaneous quality factor q(t) is the scaled ratio of
instantaneous rms instantaneous frequency to relative amplitude change,
frequency firms.
p fi (t)
Instantaneous quality factor q(t) = − (14)
is related to angle w by s(t)
q ¼ 12 tan w. Compare with
Figure 6 in Chapter 3. (Tonn, 1991); sometimes it is defined to be strictly positive.
Instantaneous quality factor is more interesting than useful.
Its connection with rock property Q is tenuous, and its value in highlighting stratigraphy
is slight.
The definition of instantaneous quality factor is suggested by seismic wave attenuation
theory. The instantaneous amplitude a(t) of a monochromatic plane wave of frequency f
propagating in an attenuating homogeneous medium with quality factor Q is given
closely by

p ft
a(t) = ao exp − , (15)
Q
where ao is the initial amplitude (Johnston, 1981, p. 2). Taking the logarithm followed by
the time derivative,
d pf
ln a(t) = − . (16)
dt Q
The left-hand side is the relative amplitude change s. Rearranging,

pf
Q=− . (17)
s
Equation 17 suggests equation 14, the definition of instantaneous quality factor.
Frequency change and amplitude acceleration

Instantaneous frequency change and amplitude acceleration employ second derivatives

and have units of hertz per second. They reveal subtle details but are noisy. They are
seldom used.
Instantaneous frequency change a(t) is the time rate of change of instantaneous
frequency:
d d2
a(t) = 2p fi (t) = 2 ui (t). (18)
dt dt
It serves as a rough inverse measure of the change in apparent reflection spacing.
Instantaneous amplitude acceleration b(t) is the time rate of change of relative
amplitude change:
d d2
b(t) = s(t) = 2 ln a(t). (19)
dt dt
Oliveros and Radovich (1997) call this “instantaneous vertical discontinuity.”
Complex trace
Complex notation often is preferred to real notation because it is succinct and simplifies
mathematics. Its basis is the complex or analytic seismic trace z(t), defined as
z(t) = x(t) + iy(t) = a(t) exp iui (t), (20)
where i is the square root of 21. In this context, the seismic and quadrature traces are called
the “real” and “imaginary” traces. Instantaneous amplitude and phase are expressed in
terms of the complex trace as

a(t) = |z(t)| = z∗ (t)z(t), and (21)
ui (t) = arg z(t) = Im[ln z(t)]; (22)
the asterisk denotes the complex conjugate. Formulas 6 and 7 for amplitude and phase
follow by substituting x(t) + iy(t) for z(t).
Instantaneous frequency is given in terms of the complex trace as
′
1 z (t)
fi (t) = Im . (23)
2p z(t)
The practical formula for instantaneous frequency, equation 10, follows by multiplying
both numerator and denominator of the term within the brackets by z∗ (t) and substituting
x(t) + iy(t) for z(t).
In terms of the complex trace, relative amplitude change is given by

′
z (t)
s (t) = Re , (24)
z(t)
instantaneous rms frequency by

1 z′ (t)
firms (t) = , (25)
2p z(t)
and instantaneous quality factor by
′
1 z (t)
q(t) = − tan arg . (26)
2 z(t)
Spikes
Spikes are ubiquitous in complex trace attributes that involve differentiation, such as
instantaneous frequency and relative amplitude change (Figure 8). Attribute spikes are
anomalously large but transient values, positive or negative. They coincide with envelope
minima caused by destructive reflection interference, discontinuities, and noise. Spikes
have value as discontinuity indicators (Hardage, 1987, p. 217; Hardage et al., 1998), but
they confound quantitative interpretation. They are removed readily through nonlinear
filtering or amplitude weighted averaging.
Because spikes in instantaneous frequency are caused by phase discontinuities that cor-
respond to amplitude minima, instantaneous amplitude and phase must be partly related.
Destructive interference
Destructive reflection interference is the major cause of attribute spikes on seismic data
with strong signal and low noise. Spikes caused by interference occur between reflections
at envelope minima and form trends that parallel the reflections. They are sensitive to the
Figure 8. A comparison of instantaneous frequency (red), relative amplitude change in units of Hz

(blue), and the trace envelope, rescaled for comparison (gray). Spikes in instantaneous frequency
and relative amplitude change coincide with envelope minima.
reflection spacing and the seismic wavelet; small changes in the spacing or the wavelet can
cause large changes in the spikes. Destructive interference that causes a jump in instan-
taneous phase slightly less than +1808 results in a positive spike in instantaneous fre-
quency. Interference that causes a jump in phase slightly less than 21808 results in a
negative frequency spike. Complete destructive interference causes a jump in instan-
taneous phase of 1808, resulting in a frequency spike that approaches a delta function.
Constructive reflection interference does not produce spikes.
Figures 9 and 10 illustrate how interference causes frequency spikes. They show syn-
thetic data and their trace envelope and instantaneous frequency for two wedge models
with a zero-phase 25-Hz Gabor wavelet (similar figures are given by Robertson and
Fisher, 1988). The model of Figure 9 has equal magnitude but opposite polarity reflections,
and that of Figure 10 has equal magnitude and equal polarity reflections. The envelopes
of the two reflection wavelets are equal at the center of the wedge. Where interference is
Figure 9. Synthetic seismic data (top) with derived trace envelope (center) and instantaneous
frequency (bottom). The synthetic data derive from a wedge model convolved with a Gabor wavelet.
The model has equal magnitude but opposite polarity reflections, and the wavelet has a 25-Hz center
frequency and 20-ms standard deviation in time. Instantaneous frequency maxima coincide with
envelope minima and are marked by red lines.
Figure 10. Synthetic seismic data (top) with derived trace envelope (center) and instantaneous
frequency (bottom). The synthetic data derive from a wedge model convolved with the same 25-Hz
Gabor wavelet used in Figure 9. The model has equal magnitude and equal polarity reflections.
Instantaneous frequency maxima coincide with envelope minima and are marked by red lines.
completely constructive, the trace envelope is twice that of either reflection’s envelope, and
the instantaneous frequency is the center frequency of the wavelet, which here is 25 Hz.
Where interference is completely destructive, the envelope is transiently zero and instan-
taneous frequency spikes to infinity. Spikes can occur only if the reflections are separate
and distinct. They cannot occur if the reflection spacings are less than the limit of resol-
ution. As explained in Chapter 7, this limit is roughly one quarter the period of the
center frequency, which is 10 ms here. This analysis holds exactly for all symmetric wave-
lets that by themselves do not generate spikes, and it holds approximately for uncompli-
cated wavelets in general.
Spikes caused by reflection interference have scant value for seismic interpretation.
They cannot indicate thin beds or pinch-outs because they do not occur between closely
spaced reflections, and they are unreliable as geological markers because they depend on
the seismic wavelet and are sensitive to small changes in reflection spacing. Spikes
simply indicate where the dominance of one reflection ends and another begins.
Faults
Reflection discontinuities due to faults are a key cause of attribute spikes. In con-
trast to reflection interference, which causes spike trends that parallel reflections,
faults cause spike trends that cut across reflections. Several discontinuity attributes
employ spikes to detect discontinuities (Luo et al., 1996; Oliveros and Radovich,
1997; Hardage et al., 1998; see Chapter 6). Because faults tend to be relatively vertical,
they are better detected by instantaneous attributes computed horizontally, like horizon-
tal wavenumber discussed below, than by those computed vertically, like instantaneous
frequency.
Response attributes and average attributes

Instantaneous attributes are made more interpretable through filtering that removes
spikes and smooths the attribute. Three approaches are common: median filtering, selection
at envelope peaks, and averaging weighted by the envelope or envelope squared. Median
filtering effectively removes isolated spikes. It is best applied as a 3D filter to achieve
lateral smoothing as well as vertical. Selection at envelope peaks and weighted averaging
both suppress spikes because spikes occur at envelope minima.
Selection at envelope peaks produces “response” attributes, sometimes called “wave-
let” attributes. These include response phase and response frequency as well as apparent
polarity and sweetness. The chief weighted average attributes are average frequency and
bandwidth.
Response phase and frequency

Response attributes derive from instantaneous attributes through a nonlinear process.
In each interval on a seismic trace bounded by successive envelope troughs, the value of
the instantaneous attribute at the time of the envelope peak is selected and assigned to
the entire interval (Figure 11). Because the time of the envelope peak rarely coincides
with a sample point, the selection process requires interpolation. The result is characteristi-
cally blocky.
Assuming constant phase wavelets and isolated reflections free of noise, response
phase equals the phase of the seismic wavelet in the data, and response frequency equals
the average spectral frequency weighted by the amplitude spectrum of the wavelet
(Bodine, 1984; Robertson and Fisher, 1988). These relations seem to give these attributes
inherent and appealing meaning, but they almost never hold because seismic reflections are
rarely isolated and noise is always present (White, 1991). Response phase records apparent
phases of reflections, which can help in reflection tracking. Response frequency offers a
cleaner version of instantaneous frequency, and resembles a blocky version of weighted
average frequency (Figure 12).
Instantaneous bandwidth has no corresponding response attribute because it is
always zero at envelope peaks. A meaningful measure of response bandwidth can
be defined in terms of the scaled square root of the instantaneous amplitude acceleration
Figure 11. A comparison of response frequency (red), instantaneous frequency (gray), and the
envelope (black). Blue arrows mark envelope peaks, and dashed lines mark envelope minima.
Within an interval bounded by successive envelope minima, response frequency is constant and
equal to the instantaneous frequency at the time of the envelope peak. Response phase is defined
similarly.
Figure 12. (a) Response phase (red) compared with instantaneous phase (gray). (b) Response
frequency (red) and weighted average frequency (blue), computed in a 52-ms Hamming window,
compared with instantaneous frequency (gray). These attributes derive from the seismic trace of
Figure 2a.
at the envelope peak, which, ideally, records the standard deviation of the wavelet ampli-
tude spectrum about the response frequency. This measure has not found practical
application.
Apparent polarity
Apparent polarity is the sign of the seismic data at the envelope peak scaled by the
value of the envelope peak. It is not usually recognized as a response attribute, though it
was the first one to be invented.
For zero-phase seismic data, apparent polarity correctly indicates the polarity of well-
separated reflections. It is unreliable where reflections interfere and unstable for thin-bed
reflections, which have apparent phase of +908. Response phase is a more reliable substi-
tute. Figure 13 demonstrates this with synthetic data; the same problem occurs on real
seismic data.
Sweetness
Sweetness is an empirical attribute designed to identify “sweet spots,” places that are
oil and gas prone. Sweetness sr(t) is defined as response amplitude ar(t) divided by the
square root of response frequency fr(t),
ar (t)
sr (t) = √ ; (27)
fr (t)
Figure 13. Illustration of the instability of apparent polarity. The synthetic data have three
reflections and a small level of random noise. Apparent polarity correctly identifies the polarity of
the top and bottom reflections. The middle reflection is a composite of two equal and opposite
reflections 4 ms apart. It represents a thin-bed reflection and resembles a single reflection with 908 of
phase. The noise causes its apparent polarity to flip randomly. In contrast, response phase is
insensitive to the noise and correctly estimates the apparent phases of all three reflections. Every
20th trace is overlain in wiggle format. Red is positive polarity, blue is negative polarity, and yellow
is 908 phase.
response amplitude is the value of the envelope at the envelope peak (Oliveros and Rado-
vich, 1997). This definition is motivated by the observation that, in young clastic sedimen-
tary basins, sweet spots tend to be characterized seismically by strong amplitudes and low
frequencies. Reflection strength and instantaneous frequency often are substituted in place
of response amplitude and response frequency to produce a more variable sweetness
measure.
Sweetness closely resembles reflection strength. Sweetness anomalies of interest
are those that are more pronounced than the corresponding reflection strength anomalies.
Hart (2008a) finds that sweetness is particularly useful for channel detection.
Average frequency
Average complex seismic trace attributes are generated from instantaneous attributes
through weighted averaging in a running window. The weighting function is nearly
always the envelope, which is reflection strength, or the envelope squared, which is instan-
taneous power. These two weighting functions produce closely similar results, but weight-
ing by instantaneous power offers the theoretical advantage that the attributes approximate
time-variant spectral averages. Several time-frequency relations equate average attributes
weighted by instantaneous power to average spectral quantities. The chief relations are Par-
seval’s relation and the first and second moment formulas (Cohen, 1995). As a result,
average frequency, rms frequency, and bandwidth can be computed either in the time
domain through complex trace analysis, or in the frequency domain through spectral
formulas.
Average instantaneous frequency fa(t), weighted by instantaneous power a 2(t) in a
window w(t), is given by
1
fi (t)a2 (t)w(t − t)d t
fa (t) = k fi (t)lw = −1 1 ; (28)

a2 (t)w(t − t)d t
−1
the brackets with subscript w denote weighted averaging in a window (Figure 12b).
Average frequency has units of hertz and represents a time-varying estimate of the
average Fourier spectral frequency.
Similarly, rms frequency frms(t) is
1
2
firms (t)a2 (t)w(t − t)d t
−1
2
frms (t) = k firms
2
(t)lw = . (29)

1
a2 (t)w(t − t)d t
−1
The rms frequency attribute is a time-varying estimate of the local rms Fourier spectral
frequency.
Bandwidth
Bandwidth b(t) combines instantaneous bandwidth and the variance of instantaneous

frequency through weighted averaging according to

1
(( fi (t) − fa (t))2 + b2i (t))a2 (t)w(t − t)d t
b2 (t) = −1 . (30)

1
a (t)w(t − t)d t
2
−1
Bandwidth has units of hertz and represents a time-varying estimate of the standard devi-
ation of the Fourier spectral frequency. Changes in both frequency and amplitude contrib-
ute to bandwidth. This accords with common experience. For example, vibroseis sweeps
that involve only frequency changes have bandwidth, as do sharp pulses that involve
only amplitude changes. Like average frequency, bandwidth finds application in strati-
graphic analysis. It is sometimes effective at revealing channels and channel systems.
Bandwidth becomes instantaneous bandwidth when the length of the analysis window
is reduced to zero. This provides the rationale for considering instantaneous bandwidth to
be a measure of bandwidth. Figure 14 compares the windowed and instantaneous band-
widths of a seismic trace.
Bandwidth is related to rms frequency and average frequency by
b2 (t) = frms
2
(t) − fa2 (t). (31)
This is the average attribute version of the relation illustrated in Figure 7 for instantaneous
attributes.
Thin-bed indicator
Suggesting that thin beds cause spikes in instantaneous frequency, Taner (2000) pro-
poses a thin-bed indicator j(t) as the difference between instantaneous frequency and
Figure 14. Bandwidth (red), computed in a 54-ms (27-point) Hamming window, compared with
instantaneous bandwidth (blue) for the seismic trace of Figure 2a. Bandwidth is more than an
average of instantaneous bandwidth because it includes the variance of the instantaneous frequency.
weighted average frequency:

j(t) = fi (t) − fa (t). (32)
The thin-bed indicator discards the background trend of the instantaneous frequency but
retains the spikes. However, as noted above, true thin beds do not generate spikes, though
thicker beds can. Zeng (2010) claims that a sequence of thin beds can give rise to spikes,
but such spikes result from the sequence and do not identify any particular thin bed. Thus
the thin-bed indicator cannot indicate thin beds or pinch-outs. It has little value.
Average quality factor

An average quality factor attribute qa(t) is defined as the ratio of the average
frequency to twice the bandwidth:
fa (t)
qa (t) = . (33)
2b(t)
This definition is similar to that for the quality factor of a filter or an electric
circuit (Close, 1966, p. 296). It is an average version of the instantaneous quality factor
defined above.
Like its instantaneous counterpart, average quality factor is largely unrelated to signal
attenuation, and its connection to rock properties is tenuous. Instead, quality factor finds
application as a qualitative measure to distinguish spectral anomalies. Zones where band-
width and average frequency change separately are characterized either by relatively low or
high quality factors. Like average frequency and bandwidth, quality factor finds modest
application in stratigraphic analysis.
3D complex seismic trace analysis

Complex seismic trace analysis is extended to 3D by measuring lateral as well
as vertical variations in seismic data. As in 1D analysis, 3D complex trace analysis
starts with the Hilbert transform. A standard 1D Hilbert transform in time suffices for
zero-mean seismic traces, which is fortunate because 3D Hilbert transformation is
awkward and costly. A 3D gradient computation replaces time differentiation. The gradient
of the instantaneous phase produces the phase vector, and the gradient of the logarithm of
the instantaneous amplitude produces the group vector. These vectors provide a basis for
computing 3D attributes, principally slope, azimuth, and relative amplitude change.
Because reflections are better defined by phase than by amplitude, the phase vector is
used for most attributes, whereas the group vector finds more limited application.
In the following development, assume the data to be in time and omit the independent
spatial variables x, y, and t. The 3D gradient operator ∇ is
∂ ∂ ∂
∇= x̂ + ŷ + t̂, (34)
∂x ∂y ∂t
where the circumflex denotes

a unit vector in a given direc-
tion. Depth data are handled by

replacing the time variable t in
the gradient operator with the
depth variable z.
Phase vector
The instantaneous phase
vector ki is the 3D counterpart
to instantaneous frequency. It
is defined as the gradient of
the instantaneous phase divided
by 2p,
1
ki = ∇ui . (35)
2p
Figure 15. For seismic data in depth, or on displays of
At any point in a seismic seismic data in time that have no vertical distortion, the 2D
volume, ki points in the direction phase vector (red arrows) is everywhere orthogonal to the
of greatest increase in phase. For local line of constant phase and points downward in the
time data, ki is a mix of horizon- direction of greatest increase in phase. In 3D, the phase
tal wavenumbers and vertical vector is everywhere orthogonal to the local surface of
frequency, and thus it has hori- constant phase. Whereas this concept of orthogonality aids
zontal units of inverse meters intuition and mathematics, it is generally invalid for seismic
or feet, and vertical units of data in time. However, it is not necessary for deriving
hertz. For depth data, the phase reflection slope and azimuth.
vector is a pure wavenumber
vector with units of inverse meters or feet. Because depth data have an orthogonal coor-
dinate system, the phase vector is orthogonal to the local surface of constant phase.
This orthogonality is a mathematical and conceptual advantage, and for this reason
many discussions of dip and azimuth attributes adopt orthogonal coordinate systems.
However, an orthogonal coordinate system is not necessary for the computation of slope
and azimuth, so for these attributes time data suffice (Figure 15).
The instantaneous phase vector is expressed in terms of wavenumber components
kxi and kyi in the x and y directions, and instantaneous frequency fi in the t direction, as
ki = kxi x̂ + kyi ŷ + fi t̂. (36)
Instantaneous wavenumber kxi is defined as
1 ∂ui
kxi = . (37)
2p ∂x
Like instantaneous frequency, instantaneous wavenumber is not computed directly from

instantaneous phase because instantaneous phase is not continuously differentiable.
Instead, wavenumber kxi is computed through
∂v ∂u
1 u ∂x − v ∂x
kxi = , (38)
2p u2 + v2
where u is the real 3D seismic data and v is the corresponding quadrature 3D data. Wave-
number kyi is computed similarly, as is vertical wavenumber kzi for depth data.
Wavenumbers kxi and kyi are negative when measured in a downdip direction and
positive when measured in an updip direction. In accordance with the dip sign convention,
the sign of wavenumber kxi is opposite to the sign of the slope along the x axis px. Similarly,
the sign of wavenumber kyi is opposite to the sign of the slope along the y axis py. This must
be taken into account in the attribute formulas.
Slope and azimuth

The instantaneous phase slope of a reflection in the x direction, px, is
kxi
px = − , (39)
fi
which has units of slowness; the minus sign is needed to conform to the dip sign convention
(Figure 16). The corresponding slope in the y direction py is
kyi
py = − . (40)
fi
x In accordance with equation 3 in

Chapter 3, instantaneous slope magni-
1/kx tude p is

kxi2 + kyi2
1/fi p= . (41)
Ref
lect
| fi |
ion
Instantaneous azimuth f quanti-
fies the dip direction. It is the angle
t in degrees from geographic north of
the downdip direction of the reflec-
Figure 16. Reflection slope in 2D is a function of tion, and is given by
wavenumber component kx and instantaneous
frequency fi. Here, the reflection slope is positive
kyi
and wavenumber kx is negative. In 3D, slope f= arctan + 180◦ + f0 , (42)
includes component ky. kxi
where f0 is the angle between the y axis N y

of the seismic survey and geographic
–kx
north (Figure 17). The added 1808
refers azimuth to the downdip direction, ϕ Downdip direction
ϕ0
and is necessary because wavenum-
bers kxi and kyi are negative in a
–ky
downdip direction. Interestingly, this
equation determines azimuth solely W E
with horizontal measures.
Wavelength and dip x

S
Wavelength and dip can be derived
only from depth data, or approximately Figure 17. Reflection azimuth f is a function of
derived from time data by converting the wavenumbers in the x and y directions, kx and ky.
The negative of the wavenumbers must be taken to
the instantaneous frequency to vertical
make azimuth positive in the downdip direction.
wavenumber, with a suitable velocity Angle f0 refers the x-y axes of the seismic survey
function v(t), according to to geographic north. Compare with Figure 8 in
2 fi (t) Chapter 1.
kzi (z) = . (43)
v(t)
Instantaneous wavelength l is the
inverse of the wavenumber magni-
tude |ki|,
x
1 1
l= = , (44)
|ki | k + k2 + k2
2
xi yi zi
and has units of meters or feet. Instan- 1/kx

taneous wavelength is the distance γ
along one cycle of a 3D sinusoidal λ
waveform measured perpendicularly 1/kz
to planes of constant phase (Figure
18). It serves as a rough measure of
apparent reflection spacing that is
independent of dip.
Instantaneous dip g is the angle in
degrees between the reflection and the z
horizontal. In 3D, instantaneous dip is
unsigned and given by Figure 18. For depth data, 2D instantaneous
wavelength l and dip g are functions of the
⎡⎤ wavenumbers in the x and z directions, kx and kz; l is
⎣ kxi + kyi ⎦
2 2 the perpendicular distance between lines of constant
g = arctan . (45) phase. Similarly, in 3D the wavelength is the distance
|kzi | between planes of constant phase.
In 2D, dip is signed according to the dip sign convention of Chapter 1, and becomes

kxi
g = −arctan . (46)
kzi
Average phase attributes
Average phase attributes in 3D cannot be derived correctly as scalar averages of instan-
taneous phase attributes, but instead must be computed through vector averaging
(Figure 19). As a consequence, the average phase attributes employ the same formulas
as their instantaneous counterparts with the difference that weighted average vector com-
ponents replace instantaneous components.
The weighted average phase vector in a 3D window w, kki lw , is
kki lw = kkxi lw x̂ + kkyi lw ŷ + k fi lw t̂. (47)
The vector components kkxi lw , kkyi lw , and k fi lw are weighted averages in the x, y, and t
directions. Component kkxi lw is derived in the window w according to

wkxi a2 dv
kkxi lw = v
, (48)
2
wa dv
v
where dv refers to an elemental volume. The components in the y and t directions are
defined similarly.
The formulas for average slope pa and azimuth fa thus become

kkxi l2w + kkyi l2w
pa = , and (49)
|k fi lw |
Figure 19. The need for vector averaging instead of scalar averaging is illustrated by two
azimuths, +1798 and 21798, which represent nearly the same direction. Their vector average is
1808, which reasonably lies midway between them, but their scalar average is 08, which
unreasonably represents the opposite direction.

kkxi lw
fa = arctan + 180◦ + f0 . (50)
kkyi lw
Group vector
The 3D counterpart to 1D instantaneous relative amplitude change is the instantaneous
group vector g, defined as
1 s
g= ∇ ln a = gx x̂ + gy ŷ + t̂, (51)
2p 2p
where gx, gy, and s are the relative amplitude changes along the x, y, and t axes. Component
gx is given by
1 ∂
gx = ln a. (52)
2p ∂x
Component gy is defined similarly. At any point in a seismic volume, g points in the direc-
tion of increasing amplitude. Thus g points downward above a reflection peak and points
upward beneath it (Figure 20). For depth data, the group vector is orthogonal to the local
surface of constant amplitude.
The components of the group vector serve as directional discontinuity measures, as
discussed in Chapter 6, but otherwise the group vector finds little application. Like the
phase vector, it yields estimates of
slope and azimuth. However, the group
x
vector must be adjusted to obtain con-
sistent estimates of azimuth or to com-
pute vector averages. One adjustment gx
reverses the direction of the group
g
vector where its vertical component σ
is negative. Alternatively, the gradient 2ππ
squared tensor can be employed in
place of the gradient to reduce vector σ
directions to orientations. 2ππ
g
gx
Gallery of complex trace

attributes t
A gallery of examples illustra-
Figure 20. The 2D instantaneous group vector g
tes complex seismic trace attributes
with components gx and 2sp for a dipping reflection.
applied to a seismic line from a 3D Trace envelopes are shown as filled lines and
survey in the Gippsland Basin, off- reflection wavelets as dashed lines. The group
shore southeast Australia. All attribu- vector points in the direction of increasing amplitude,
tes derive from the seismic line of which is downward above a reflection peak but
Figure 21. upward beneath it.
Figure 21. A seismic line and its amplitude and phase attributes. This is the same seismic line
shown in Figure 10 of Chapter 3. (a) Original seismic data. (b) Reflection strength.
(c) Instantaneous phase. (d) Response phase. (e) Cosine of the instantaneous phase. (f) Apparent
polarity. The data are from the Gippsland Basin, offshore southeast Australia.
1D attributes
Figure 21 compares amplitude and phase attributes: reflection strength, instantaneous
phase, response phase, cosine of the phase, and apparent polarity. Reflection strength high-
lights amplitude contrasts, and instantaneous phase and cosine of the phase reveal details in
reflection continuity. Response phase and apparent polarity appear blocky and are applied
when reflection phase or polarity are of interest.
Figure 22 compares frequency and bandwidth attributes, with the original seismic data
overlain to relate the attributes to the reflections. Average frequency is more interpretable
and smoother than both instantaneous frequency and response frequency. Instantaneous
Figure 22. Frequency and bandwidth attributes for the seismic line of Figure 21a. (a)
Instantaneous frequency. (b) Average frequency. (c) Instantaneous rms frequency. (d) Average rms
frequency. (e) Response frequency. (f) Bandwidth. Average frequency, average rms frequency, and
bandwidth are computed in a window of 11 samples (44 ms). The five frequency attributes are
scaled from 0 to 75 Hz, but bandwidth is scaled from 0 to 40 Hz. Every 5th original seismic
trace is overlain in variable area format.
rms frequency resembles instantaneous frequency, but with higher values. Bandwidth
complements average frequency because it incorporates both amplitude and frequency
changes.
Figure 23 shows relative amplitude change in time, sweetness, thin-bed indicator, and
average quality factor. Average quality factor is preferred to instantaneous quality factor
because it is more stable and clearer. Likewise, average sweetness is preferred to sweetness
Figure 23. Miscellaneous complex trace attributes for the seismic line of Figure 21a.
(a) Relative amplitude change. (b) Sweetness. (c) Thin-bed indicator; blue indicates negative
frequency spikes and yellow indicates positive spikes. (d) Average quality factor. Relative
amplitude change is instantaneous, but the other three attributes are derived in a window of 11
samples (44 ms). In (b) and (d), every 5th original seismic trace is overlain in variable area format.
defined as a response attribute because it is somewhat cleaner and not blocky. Quality
factor and sweetness incorporate both amplitude and frequency information. Quality
factor does not resemble other attributes, but sweetness resembles reflection strength.
Relative amplitude change has ridges where reflections interfere, which sometimes
indicate channels or faults. These ridges appear illuminated and coincide with the fre-
quency spikes that are highlighted by the thin-bed indicator. Thus relative amplitude
change and the thin-bed indicator provide much the same information, even though rela-
tive amplitude change is derived from amplitudes and the thin-bed indicator is derived
from phases.
3D attributes
Figure 24 compares instantaneous and average wavelengths, dips, and azimuths
derived from seismic data in time for which the frequency component of the phase
vector has been converted to a vertical wavenumber. The instantaneous attributes are
noisy and inferior to the average attributes. Because dips are not strong in these data,
average wavelength resembles average frequency. Figure 25 shows horizontal relative
Figure 24. Phase vector attributes for the seismic line of Figure 21a. (a) Instantaneous azimuth.
(b) Average azimuth. (c) Instantaneous dip. (d) Average dip. (e) Instantaneous wavelength.
(f) Average wavelength. The average attributes are computed in a window of 3 lines by 3 traces
by 5 samples. Every 5th seismic trace is overlain in variable area format.
amplitude change. Horizontal relative amplitude change resembles vertical relative

amplitude change, though they differ in scale and computation direction. Vertical relative
amplitude change highlights horizontal features and is insensitive to vertical features,
whereas horizontal relative amplitude change highlights vertical features and is insensitive
to horizontal features. As a result, horizontal relative amplitude change is more effective at
revealing faults and channels.
Summary
The essence of complex seismic trace

analysis lies in the separation of the ampli-
tude information from the phase infor-
mation in seismic data. This separation
produces two fundamental attributes, in-
stantaneous amplitude and instantaneous
phase, from which all other complex trace
attributes derive, chiefly frequency, band-
width, slope, azimuth, and relative ampli-
tude change. At a given time on a seismic
Figure 25. Horizontal relative amplitude trace, instantaneous amplitude, phase, and
change for the seismic line of Figure 21a; frequency describe a sinusoid that locally
compare with vertical relative amplitude change matches the trace at that time. Similarly, at
shown in Figure 23a. The yellow arrow indicates a given point in a seismic volume, instan-
the direction of computation. Amplitude taneous amplitude and phase, along with
change reveals detail hidden in the amplitudes. the instantaneous phase vector, describe a
3D sinusoid that locally matches the data
at the point. Figure 26 summarizes the
relationships between the key 1D complex
seismic trace attributes, and Figure 27 sum-
marizes the relationships between the key
3D attributes.
There are striking parallels and strong
ties between complex trace analysis and
Fourier analysis. Both employ similar math-
ematics to separate amplitude and phase
information. Further, many average com-
plex trace attributes equal Fourier spectral
averages. In particular, average instan-
taneous frequency equals an average
Fourier spectral frequency, which justifies
its use in attenuation studies. Nonetheless,
there are significant differences between
complex trace analysis and Fourier analysis.
The chief difference is that complex seismic
trace analysis introduces a quadrature trace
derived from the real seismic trace. Its
Figure 26. Chart relating the principal 1D
definition is reasonable but not obvious,
instantaneous complex seismic trace attributes
in time; h(t) is the Hilbert transform operator because it has no general physical signifi-
and d/dt represents differentiation. Amplitude cance. In contrast, Fourier cosine and sine
and phase are the fundamental attributes from spectra are independent functions with
which all others derive. physical significance.
In spite of its name, complex seismic

trace analysis can be developed without
resort to complex mathematics; the eso-

teric ideas of the complex trace should not
obscure the prosaic meaning of the attributes.
Complex mathematics facilitates the deri-
vation of many attributes, but it is not required.
Most instantaneous attributes require
differentiation. As a result, they tend to look
noisy and are prone to spikes. Spikes occur
at envelope minima and are caused by reflec-
tion interference, discontinuities, and noise.
They confuse seismic interpretation, but
they are easily removed through nonlinear fil-
Figure 27. Chart relating the principal 3D tering or weighted averaging. Spikes nonethe-
instantaneous complex trace attributes in less have value as discontinuity indicators.
depth; h(t) is the standard Hilbert transform The extension of complex seismic trace
operator in time, and ∇(x,y,t) is the 3D analysis to 3D is based on the phase and
gradient operator.
group vectors, from which nearly all 3D
complex trace attributes derive. The phase
vector treats reflections as surfaces of constant phase and points in the direction of
increasing phase, which is downward except possibly at spikes. The group vector treats
reflections as surfaces of constant amplitude and points in the direction of increasing ampli-
tude, which is upward as often as downward. Both phase and group vectors record reflec-
tion orientations and spacings, but because seismic reflections are better defined by phase
than by amplitude, the phase vector is preferred. The components of the group vector are
relative amplitude changes, which serve as high-resolution directional discontinuity
measures.
Complex seismic trace analysis is one of several competing methods for quantifying
reflection orientations. The other methods, and their application to structural and strati-
graphic attributes, form the topic of the next chapter.

Chapter 5
Structural and Stratigraphic Attributes
Introduction
Structural and stratigraphic seismic attributes quantify seismic properties associated
with geological structure and stratigraphy. Structural attributes quantify properties of
faults, folds, and diapirs, and include dip, slope, azimuth, seismic shaded relief, and curva-
ture. Stratigraphic attributes quantify reflection patterns related to stratigraphy. The pat-
terns considered here are 3D and derive from the ideas of seismic stratigraphy. They
include amplitude variance, reflection spacing, parallelism, and divergence. Discontinuity
attributes highlight elements of both geological structure and stratigraphy and are treated
separately in Chapter 6.
Dip and azimuth record seismic reflection orientations. They are computed through
various methods, primarily dip scanning, complex seismic trace analysis, the plane-wave
destructor, and the gradient squared tensor. For qualitative seismic interpretation, dip
and azimuth are better combined as apparent dip or as seismic shaded relief, which
resemble apparent topography.
Curvature is the rate of change of reflection orientation. Because reflection orientation
can change greatly in different directions, curvature is a complicated property that gives
rise to a number of seismic attributes. In practice, only a few curvature attributes prove
helpful. Curvature complements discontinuity by revealing different and more detailed
structural features.
Three-dimensional stratigraphic attributes are relatively uncommon. In spite of keen
interest for such attributes, their development has lagged that of structural attributes
because 3D stratigraphic properties are harder to quantify than structural properties.
Further, stratigraphic attributes are better interpreted as a set according to principles
of seismic stratigraphy, which is more challenging than interpretation of individual
attributes.
Dip and azimuth

Dip and azimuth record reflection orientations and are the basic 3D seismic attributes.
Their value as individual attributes is limited, but when they are combined as apparent dip
or shaded relief, they produce geologically intuitive displays. More importantly, dip and
azimuth provide the foundation for many 3D attributes, such as curvature and parallelism,
as well as for processes that follow reflections, such as coherency filtering and automatic
volume flattening.
75
Dip and azimuth are computed through dip

scanning, complex seismic trace analysis, the
plane-wave destructor, and the gradient squared

tensor. Most modern applications implement either
the plane-wave destructor or the gradient squared
tensor because they are faster than dip scanning or
complex seismic trace analysis.
Complex seismic trace analysis computes
slope, dip, and azimuth through instantaneous fre-
quency and horizontal wavenumbers, as described
in Chapter 4. The other methods find the slopes in
Figure 1. In a small analysis the x and y directions, px and py, and transform
window, seismic reflection data them to slope magnitude, dip, and azimuth accord-
approximate a plane wave, which has ing to equations 3, 4, and 5 in Chapter 3. They are
constant slope. In this 2D example, founded on the observation that, locally, seismic
slope px Dt/Dx. data tend to approximate a plane wave with con-
stant slopes (Figure 1). Thus, within a small analysis
window, 3D seismic data in time u(x, y, t) approximate
u(x, y, t) ≈ u(t − px x − py y). (1)
For depth data, depth variable z replaces time variable t.

Though there is a distinction between dip and slope, as described in Chapter 1, it
is common to refer to both simply as dip. For example, the method of dip scanning to
find dip and azimuth actually scans along slopes to find the reflection slopes.
Dip scanning
Dip scanning is a brute-force method for determining slope and azimuth. It estimates
the reflection slopes as those slopes along which the semblance S of the seismic data is
maximum (Marfurt et al., 1998). These estimates are limited to a set of trial slopes that
are chosen in a compromise between computational efficiency and resolution.
Semblance S computed along trial slopes px and py for discrete seismic data u(x, y, t)
is given by
2

N
M
L
u(xj , yk , ti − px Dx − py Dy)
i=1 j=1 k=1
S( px , py ) = , (2)
N
M
L
ML u2 (x j, yk , ti − px Dx − py Dy)
i=1 j=1 k=1
where i, j, and k are sample indices for t, x, and y, N is the number of time samples per trace,
M is the number of traces per line in the x direction, and L is the number of traces per line in
the y direction. The reflection slopes are taken to be those values of px and py that maximize
Chapter 5: Structural and Stratigraphic Attributes 77
the semblance. Accurate estimates require interpolation or resampling in time. Marfurt

et al. (1998) substitute complex traces in place of real traces to obtain somewhat more
stable estimates of semblance, which roughly doubles the computational cost.
Plane-wave destructor
The plane-wave destructor estimates reflection slopes through a least-squares fit to the
wave equation (Claerbout, 1992, p. 94). Consider 3D seismic reflections with slope px in
the x direction. The reflections can be nearly “destroyed” by a 2D differential operator
in the x-t plane according to

∂ ∂
+ px u(x, y, t) ≈ 1x , (3)
∂x ∂t
where 1x is an error term. The square of 1x is the error energy. The more closely the
reflection slope matches px, the smaller the error energy. Represent the discrete seismic
data u(x, y, t) as a set of N data points, ui (i ¼ 1, 2, . . . N ), where index i refers to a
unique sample location in a 3D analysis window. The total error energy Ex in the
window is the sum of the N individual error energies 12xi:

N N
2
∂ui ∂ui
Ex = 12xi = + px . (4)
i=1 i=1
∂x ∂t
The slope of the reflections in the x direction is taken to be that value of px that produces
the smallest total error energy Ex. To find this slope, take the derivative of Ex with respect
to px and set it to zero:
N
∂Ex ∂ui ∂ui ∂ui
=2 + px = 0. (5)
∂ px i=1
∂t ∂x ∂t
Solving for px,

N ∂u ∂u
i i
i=1 ∂t ∂x
px = − 2 . (6)
N ∂ui
i=1 ∂t
Similarly, the slope in the y direction, py, is

N ∂u ∂u
i i
∂t ∂y
py = − i=1 2 . (7)
N ∂ui
i=1 ∂t
Gradient squared tensor

Like the plane-wave destructor, the gradient squared tensor determines reflection
slopes from gradients of the seismic data (Randen et al., 2000). Let vector gx be the gradient
operator in the x direction. Gradient operators in the gradient squared tensor typically
combine differentiation with smoothing by a Gaussian filter. The gradient in the x direction
of the seismic data u is the vector Gx given by
Gx = gx ∗ u(x, y, t). (8)
Vector gradients Gy and Gt in the y and t directions are defined similarly. The full gradient
G of the seismic data is the sum of Gx, Gy, and Gt. The gradient structure tensor is the matrix
formed as the outer product of the full gradient with itself:

G2
x Gx Gy Gx Gt
G · T
G = Gy Gx

G2y Gy Gt .

(9)
Gt Gx Gt Gy G2t
The matrix elements may be averaged locally to produce more stable estimates. The
eigenvector e1 associated with the largest eigenvalue of the matrix is in the direction
of the strongest gradient, which is normal to the dominant reflections. Constraining eigen-
vector e1 to point downward and representing its components in x, y, and t as ex, ey, and et,
slopes px and py are given by
ex
px = , and (10)
et
x
␥ ey
py = (11)
et
n
tio (Figure 2). The gradient squared tensor is widely
ec
Refl employed in tools for seismic image processing,
such as coherency filtering. For this purpose, the
␥
vertical axis is usually labeled z, regardless of
et the data domain, and the trace and sample spa-
cings are set equal to 1.
ex e1
t
Exaggerated slope and dip
Figure 2. Reflection slope px and dip
g are found from the components ex and Reflection slopes and dips are sometimes
ey of eigenvector e1, which is normal to exaggerated to improve the contrast in a display.
the reflection. If the magnitude of Slope p is exaggerated by scaling with a factor
eigenvector e1 is 1, then dip x, which typically ranges from 5 to 20. For data
g ¼ arccos et. in depth, the corresponding exaggerated dip gx
is the arctangent of the exaggerated slope:

gx = arctan [xp]. (12)
Apparent slope and dip

Apparent slope combines dip and azimuth to produce images that appear more
natural than slope magnitude or azimuth taken separately (Marfurt and Kirlin, 2000). It
is computed through directional derivatives and therefore is signed. Like all direc-
tional attributes, it looks like illuminated apparent topography when displayed in
monochrome.
Slope components px and py are apparent slopes in the x and y directions. An apparent
slope pu in an arbitrary direction in the x-y plane is defined as a combination of these com-
ponents. Define slope vector p as the vector sum,
p = px x̂ + py ŷ, (13)
where x̂ and ŷ are unit vectors in the x and y directions. Express an arbitrary direction in the
x-y plane as the unit vector û,
û = cos w · x̂ + sinw · ŷ, (14)
where w is the angle measured counterclockwise from the x-axis to the direction of û
(Figure 3). The dot product of p with û yields the apparent slope pu in the direction û:
pu = p · û = p cos w + p sinw.
x y (15)
For depth data, apparent dip gu in direction û is the arctangent of the apparent slope pu.
Following the sign convention for reflection dip, apparent slope and dip are positive for
reflections that dip downward in the direction of com-
putation, and are negative for reflections that dip
upward.
Apparent dips equal true dips for reflections whose
azimuths parallel the computation direction, but are
zero for reflections whose azimuths trend perpendicu-
lar to the computation direction. Apparent dip acts as
a directional filter, revealing faults, anticlines, and
other features that trend perpendicular to the compu-
tation direction, while hiding features that parallel it.
Reflection dip and azimuth can also be combined
through a 2D color bar or through volume blend-
ing to produce “dip-azimuth” (Marfurt et al., 1998). Figure 3. Unit vector û defines an
Seismic shaded relief, reviewed below, presents the arbitrary direction in the x-y plane.
same information as dip-azimuth in a more natural, The angle w is measured counter-
qualitative way. clockwise from the x-axis to û.
Figure 4. A comparison of reflection slopes for the seismic line of Figure 21a in Chapter 4. The
slopes are computed through (a) dip scanning, (b) complex seismic trace analysis, (c) the plane-
wave destructor, and (d) the gradient squared tensor. All four methods employ an analysis window
of 3 lines by 3 traces by 7 samples. Every 5th trace of the seismic data is overlain in variable area
format.
Examples of dip and azimuth

Figures 4 and 5 compare the reflection slopes and azimuths computed through dip
scanning, complex seismic trace analysis, the plane-wave destructor, and the gradient
squared tensor. Though the four methods differ greatly, their results are closely compar-
able. Attributes derived through dip scanning are quantized and therefore less accurate
than those of the other methods, which differ only in detail. In these tests, run times for
dip scanning and complex trace analysis are roughly comparable, but the plane-wave
destructor is three times faster, and the gradient squared tensor is twice as fast.
Seismic shaded relief

Shaded relief maps of digital terrain and bathymetric data look like illuminated topo-
graphy. Such maps appear more natural than contour maps and are easier to comprehend.
Shaded relief maps of gravity, magnetic, and other geophysical data show apparent illumi-
nated topography. They aid geological intuition because apparent topography often
suggests true geology.
Figure 5. A comparison of reflection azimuths for the seismic line of Figure 21a in Chapter 4. The
azimuths are computed through (a) dip scanning, (b) complex seismic trace analysis, (c) the
plane-wave destructor, and (d) the gradient squared tensor. All four methods employ an analysis
window of 3 lines by 3 traces by 7 samples. Every 5th trace of the seismic data is overlain in variable
area format.
Shaded relief techniques are readily adapted to 3D seismic data to produce a shaded
relief seismic attribute that resembles apparent illuminated topography when viewed
along horizontal slices or horizons. Illumination acts as a directional filter that enhances
features perpendicular to the illumination direction and suppresses features parallel to it.
Seismic shaded relief is therefore directional and, like all directional attributes, should
be created in pairs with orthogonal illumination directions so as to capture all features
(see Brown, 2011, p. 265). Shaded relief produces intuitively interpretable images that
reveal anticlines, synclines, folds, domes, basins, faults, and channels.
Gersztenkorn (2012) describes an attribute, based on the continuous wavelet transform,
that closely resembles seismic shaded relief with possibly superior resolution.
Illumination models
To adapt the concept of shaded relief to 3D seismic data, treat each point in the seismic
data volume as lying on a surface whose orientation is defined by the local seismic reflec-
tion dip and azimuth. Illuminate all reflection surfaces simultaneously from a single light
source, the “sun.” Let the sun be distant so the illumination direction and intensity are
constant everywhere (Figure 6). The

surfaces reflect light according to an
illumination model. Consider three

illumination models: a model for dull sur-
faces; a model for shiny surfaces; and a
model for surfaces intermediate between
dull and shiny, the Phong model (Nikolai-
dis and Pitas, 2001).
The following development treats
seismic reflections like topographic sur-
faces in a conventional 3D space with
the z-axis positive upward, and with the
sun above the surfaces. This approach
to illumination is familiar and aids under-
standing (see Horn, 1981). In practical
application to seismic data in time, the
time axis is converted roughly to depth
using a constant velocity. Reflection
dips are usually exaggerated to produce
sufficient contrast, and reflection orien-
tations are quantified by unit vectors
that point upward, instead of downward
as is customary. These adjustments
Figure 6. Seismic shaded relief illuminates are justifiable because seismic shaded
subsurface geological structure. (a) A geological relief is purely qualitative.
cross section with flat layers overlaying
Dull surfaces reflect light diffusely.
anticlinal layers. (b) The cross section as it would
appear if imaged by seismic data and converted
By this model, the illumination Id of a
to seismic shaded relief. At any point on a layer surface is proportional to the energy
surface, the illumination is a function of the density of the light incident upon it.
angle of incidence of the light upon the Hence
reflection surface.
Id = n̂ · ŝ = cos u, (16)
where n̂ is the upward-pointing surface normal, ŝ is the direction to the sun, and u is the
angle between n̂ and ŝ (Figure 7). Shiny surfaces reflect light specularly. By this model,
the illumination Is of a point on a surface depends strongly on the angle f between the
direction of the reflected light û and the direction to the observer v̂ (Figure 8). It is
expressed as
Is = [û · v̂]b = cosb f. (17)
Exponent b governs the shininess of the surface. Larger values of b make the surface appear
shinier. Typically, b is set to 2.
Surfaces intermediate between dull and shiny are modeled by averaging the illumi-
nations for diffuse and specular reflections.
Negative illuminations imply shadows. Shadows are more prevalent with low sun
elevations. For specular reflections, set the sign of the illumination to the sign of cos f.
Illumination formulas
To illuminate seismic data, express illumination model equations 16 and 17 in terms of
reflection dip and azimuth and sun elevation and azimuth. Let gr be reflection dip magni-
tude, fr be reflection azimuth, gs be the angle
between the vertical and the sun direction
(908 minus the sun elevation), and fs be sun
azimuth. The practical formula for illumina-
tion according to the diffuse reflection model,
equation 16, becomes
Id = sin gr sin gs cos (fr − fs )

+ cos gr cos gs . (18)
Figure 7. The illumination model for light To convert illumination equation 17 for
reflected diffusely. Unit vector n̂ is the shaded relief of shiny surfaces into a practical
surface normal, ŝ points to the light source, formula, express vector û in terms of vectors n̂
and u is the angle between them. Illumination and ŝ (refer to Figure 8):
is proportional to cosu.
û = (ŝ · n̂)n̂ + (ŝ × n̂) × n̂. (19)
Insert this into equation 17 to obtain
Is = ([(ŝ · n̂)n̂ + (ŝ × n̂) × n̂] · v̂)b .

(20)
Structural attributes are usually

interpreted along horizontal slices
instead of on vertical sections.
When viewing a horizontal slice,
the observer looks down the z-axis,
which implies that v̂ ¼ ẑ. In this
case, equation 20 reduces to
Figure 8. The illumination model for light reflected Is = (2Id cos gr − cos gs )b . (21)
specularly. Unit vector n̂ is the surface normal, ŝ points
to the light source, û is the direction of reflected light, The practical formula for Phong
v̂ points to the observer, and u is the angle between û and illumination Ip is the weighted
v̂. Illumination is a function of cosu. average of diffuse and specular
illumination,
Ip = wId + (1 − w)Is , (22)
where w is a weight between 0 and 1. Smaller

weights produce shinier surfaces.
Examples of seismic shaded relief

Seismic shaded relief is displayed with a
gray or monochrome color scale. High values
are assigned light shades and low values are
assigned dark shades. The apparent topography
on a horizontal slice through seismic shaded
relief closely matches the subsurface topo-
graphy imaged by the seismic data, especially
where reflections exhibit similar folding. Con-
ceptually, the attribute reconstructs subsurface
topography around a slice given dips and azi-
muths along the slice, much like a geologist
mentally reconstructs eroded topography given
strike and dip measurements along an erosional
surface. The overall geological structure on a
slice through seismic shaded relief resembles
the structure on an intersecting interpreted
horizon (Figure 9). The correspondence is often
close even if the horizon encompasses a large
vertical range, or where the structure is compli-
cated. This justifies employing seismic shaded
relief for reconnaissance of subsurface structure.
However, details seen on a seismic shaded relief
horizontal slice, such as small channels or minor
faults, match details on a horizon only where the
Figure 9. A time slice through a seismic horizon is close to the slice.
shaded relief volume shows the same Figure 10 shows seismic shaded relief along
general geological structure as an a time slice with large faults and a prominent
interpreted seismic horizon. (a) Seismic igneous intrusion. Rotating the direction of illu-
horizon. The horizon times span nearly mination 908 changes the appearance substan-
800 ms and range from 535 ms (red) to tially. Shaded relief for a dull surface looks
1354 ms (dark blue); bright yellow
more natural than for a shiny surface, but a
corresponds to 840 ms. The yellow arrow
indicates the direction of illumination.
shiny surface accentuates subtle features.
(b) Seismic shaded relief time slice at Seismic shaded relief finds application in
840 ms. (c) Original seismic data at 840 ms. data reconnaissance and in volume blending as
Data from the Taranaki Basin, offshore a background attribute to highlight geological
New Zealand. structure.
Figure 10. Seismic shaded relief along a time slice (1864 ms). (a) Original seismic data.
(b) Seismic shaded relief for a dull surface. (c) Seismic shaded relief for a dull surface with
illumination orthogonal to (b). (d) Seismic shaded relief for a shiny surface. Yellow arrows indicate
directions of illumination. Data from the Taranaki Basin, offshore New Zealand.
Volume curvature
Volume curvature attributes quantify changes in reflection dip and azimuth along
reflections in 3D seismic data (Al-Dossary and Marfurt, 2006). There are many curvature
attributes, but only most positive and most negative curvatures have found much appli-
cation as volume attributes. Curvature attributes complement discontinuity attributes by
revealing structural details, including faults, sags, bumps, and channel edges. Because cur-
vature attributes involve second derivatives, they tend to be noisy and often enhance acqui-
sition footprints. For this reason, it is best to derive curvature from coherency-filtered
seismic data, or employ large computation windows.
Volume curvature attributes are computed much like horizon curvature attributes.
Reflection slopes are found through any suitable method and then employed in the
equations for the horizon curvature coefficients as described in Chapter 3. The only differ-
ence is that horizon curvature is derived directly from the horizon time or depth, but
volume curvature is derived from the reflection slopes. For well-picked horizons, the
horizon curvature closely matches the corresponding volume curvature extracted along
the horizon.
Volume curvature also can be computed through curve-fitting methods that bypass the
direct computation of slopes (Klein et al., 2009).
Phase and amplitude curvature

The standard approach for deriving curvature attributes treats seismic reflections as
surfaces of constant phase with variable depth. An alternative approach derives curvature
attributes from surfaces of constant depth but variable phase (Taner, 2000). These two
approaches produce comparable results and are equivalent where successive reflections
have similar folding.
A horizontal slice through a 3D seismic volume can be treated as a surface with con-
stant time or depth but variable instantaneous phase uı(x, y). It is approximated by the
second-order polynomial
ui (x, y) = Cxx x2 + Cyy y2 + Cxy xy + Cx x + Cy y + C0 . (23)
The coefficients are defined as before in Chapter 3, with instantaneous phase ui repla-
cing the time or depth variable z. “Phase curvature” attributes follow from the same
equations used for standard curvature.
Chopra and Marfurt (2011) introduce amplitude curvature attributes based on
second derivatives of the “coherent amplitude.” Their amplitude curvature attributes
appear to provide better resolution than standard curvature attributes. Similar amplitude
curvature attributes are derived more simply by applying standard curvature algorithms
to the perigram of the seismic data. Alternatively, an amplitude curvature can be derived
like phase curvature, by taking derivatives of the logarithm of the instantaneous ampli-
tude. Oliveros and Radovich (1997) develop similar attributes, which they apply as dis-
continuity measures. The meaning of amplitude curvature attributes has yet to be fully
explored.
Examples of volume curvature

Most positive curvature records the most positive rate of change of the reflection
dip and highlights anticlinal tops and reflection bumps (see Figure 9 in Chapter 1;
Figure 11). It is closely related to most negative curvature, which records the most negative
rate of change of dip and highlights synclinal bottoms and reflection sags. In the example
of Figure 11, bumps and sags are large and easily identified, but curvature also reveals
smaller features that are difficult to detect otherwise.
Figure 12 shows a time slice of seismic data with a channel, which becomes distinct
with seismic shaded relief. Comparing positive and negative curvatures with the shaded
relief, it is seen that channel edges have positive curvature and channel interiors have nega-
tive curvature. This pattern is typical. However, where differential compaction occurs, as in
the basins of western Canada, channels may acquire a mounded appearance like eskers so
Figure 11. Reflection bumps create positive curvature and sags create negative curvature. (a) A
close-up of seismic data in a vertical view. The yellow oval identifies a reflection bump and the blue
oval identifies a reflection sag. (b) Strong values of most positive curvature (orange) overlain on the
seismic data. (c) Strong values of most negative curvature (blue) overlain on the seismic data.
Figure 12. Channel edges have positive curvature and channel interiors have negative curvature.
(a) A close-up view of a time slice through seismic data with channels. (b) Seismic shaded relief
reveals the channel geometry. (c) Strong values of most positive curvature (orange) overlain on
shaded relief. (d) Strong values of most negative curvature (blue) overlain on shaded relief.
that the curvatures are reversed, with positive curvature in the interior and negative curva-
ture on the edges (Chopra and Marfurt, 2012).
Normal faults often exhibit positive curvature on the upthrown side and negative cur-
vature on the downthrown side (Figure 13). Curvature attributes reveal faults, but tend to
show them offset slightly from their true locations, which are better determined by discon-
tinuity attributes.
Figure 13. Faults tend to have positive curvature on their upthrown sides and negative
curvature on their downthrown sides. (a) Vertical section through seismic data with a large fault.
(b) Strong values of seismic discontinuity (black) overlain on the seismic data. (c) Strong values
of most positive curvature (orange) overlain on the seismic data. (d) Strong values of most
negative curvature (blue) overlain on the seismic data. Data from the Taranaki Basin, offshore
New Zealand.
Figure 14 compares most positive and most negative curvatures applied to seismic
data with a great deal of structure; a discontinuity attribute again serves as a reference.
The three attributes show similar features. The curvature attributes reveal finer detail
than discontinuity but are noisier. Discontinuity is best for imaging faults or channels.
Curvature is best for identifying where fracturing most likely occurs, such as anticlinal
tops and synclinal bottoms. Reflection bumps and sags also identify possible fracture
zones. Exercise caution when interpreting details in curvature because they are not well
understood. Here, “details” refers to features that are roughly at the scale of the operator
size of the curvature attribute. Details in most positive curvature do not correspond to
details in most negative curvature in the same way as larger features, which complicates
their interpretation.
Figure 14. Comparison of discontinuity and curvature attributes along a time slice (1028 ms)
through seismic data with channels, faults, and diapirs. (a) Original seismic data. (b) Discontinuity.
(c) Most positive curvature. (d) Most negative curvature. Data from the Taranaki Basin, offshore
New Zealand.
Stratigraphic attributes
Three-dimensional stratigraphic attributes quantify stratigraphic reflection patterns in
seismic data (Randen et al., 2000; van Hoek et al., 2010). In seismic stratigraphy, these pat-
terns are termed “seismic facies parameters,” and include amplitude variance, reflection
spacing, parallelism, continuity, divergence, waviness, and reflection-free. Taken together,
seismic facies parameters define a “seismic facies.” A seismic facies is a recognizable and
quantifiable pattern in seismic data with geological significance. More specifically, a
seismic facies may represent a particular environment of deposition, which suggests a lith-
ology. For example, high-energy environments are sand prone, and low-energy environ-
ments are shale prone. Seismic stratigraphy is the identification, delineation, and
stratigraphic interpretation of seismic facies (Mitchum et al., 1977). The principles of
seismic stratigraphy provide the geological basis for designing and interpreting strati-
graphic attributes.
Most stratigraphic attributes record variations in reflection orientation. They are com-
puted in broad 3D analysis windows that ideally follow structure. Like seismic shaded
relief, they treat each point in the seismic data as lying on a reflection surface whose orien-
tation is quantified as a unit vector perpendicular to the reflection (Dalley, 2008; Figure 15).
In practical computation, it may be

necessary to exaggerate the reflec-
tion dips to produce sufficient

contrast in the attributes. Three-
dimensional stratigraphic properties
have attracted less attention than
structural properties because they
are more difficult to quantify and
interpret. The empirical formulas
given below suggest how 3D strati-
graphic patterns might be measured
Figure 15. A vector field comprises vectors that are qualitatively.
everywhere perpendicular to seismic reflections.
Their directions yield reflection dip and azimuth. The
Reflection amplitude variance
vector magnitudes often are set to 1, but they could
represent reflection spacing, as in this figure, or someReflection amplitude variance is
other property. The concept of a vector field is the relative degree to which seismic
natural for seismic data in depth. For qualitative amplitudes vary locally. It is quanti-
analysis, it can be adapted for seismic data in time,
fied as the variance in amplitude
such as by treating the data as an image with unit
normalized by the square of the
spacings between traces and samples.
average amplitude, where amplitude
is measured by reflection strength or
a similar attribute. Normalized amplitude variance VN for an amplitude attribute a is
s 2a k(a − kal)2 l
VN = = , (24)
m2a kal2
where sa is the standard deviation of the amplitude attribute and ma is its average value.
This is given a more convenient range by taking its square root and multiplying by 100
to produce a scaled coefficient of variation s̃ :

s̃ = 100 VN . (25)
Presented in this way, amplitude variance is 0 for perfectly uniform amplitudes, about 10 to
40 for moderately varying amplitudes, and up to 100 or more for highly varying amplitudes.
Unlike amplitude itself, amplitude variance is quantitatively comparable between different
data sets. For good-quality data, high amplitude variance indicates greatly varying reflec-
tivity and suggests a high-energy depositional environment; low amplitude variance indi-
cates uniform reflectivity and suggests a low-energy depositional environment.
Reflection spacing
Reflection spacing R is the distance between successive reflections measured perpen-
dicularly to the reflections. The reality of reflection interference and limited resolution
make this concept vague and only loosely related to the stratigraphic property of interest,
bed thickness, which itself lacks precise definition. Nonetheless, reflection spacing retains
qualitative value for characterizing seismic data. Suitable measures include phase wave-
length and a wavelength based on both amplitude and phase changes,
1
R = , (26)
k2i + g2i
where ki and gi are the instantaneous phase and group wavenumber vectors. This measure is
akin to rms frequency.
Schmidt et al. (2013) offer an alternative measure of bed thickness, which they describe
as the time derivative of the seismic data converted to relative geological time.
Reflection parallelism
Reflection parallelism measures how parallel the reflections in a sequence are to
each other. It is quantified in terms of the variance in the reflection orientations within a
window (Taner, 2000; de Rooij and Tingdahl, 2002; Figure 16). Parallel reflections
exhibit uniform dip and azimuth and suggest a low-energy depositional environment.
Nonparallel reflections exhibit significant variance in dip and azimuth and suggest a
high-energy depositional environment.
An empirical measure of parallelism P is defined as the square of the average cosine
between the unit vectors for the reflection orientation r̂ at a point and the average local
reflection orientation r̂a . Hence
P = 100 kr̂ · r̂a l2 . (27)
The values are scaled to fall into

the convenient range of 0 to 100.
Large values signify parallel reflec-
tions, and small values signify non-
parallel reflections.
Reflection divergence
Reflection divergence is the
degree to which reflections in a
sequence diverge consistently. Diver-
gent reflections are characterized by Figure 16. A sequence of reflections with various
constant azimuth, increasing dip orientations in a vertical window. The large arrow
with time, and decreasing dip in the represents the average reflection orientation.
azimuth direction (Figure 17). At Parallelism quantifies how much reflection orientations
the base of a divergent sequence, vary from the average.
reflections converge and pinch out.

Divergence is characteristic of
progradation and suggests a higher-

energy depositional environment.
In the absence of postdepositional
structural deformation, the azimuth
of divergent reflections is the direc-
tion of sediment transport.
Reflection divergence D can be
quantified by
D=c k∂p∂t l · W(f), (28)
Figure 17. Divergent reflections in a vertical where p is reflection slope, c is a

window. Divergence requires a large analysis window.constant that scales divergence to
an appropriate range, and W(f) is
a weighting function of azimuth f. The weighting function favors sequences with constant
azimuth, and is given by
W(f) = kcos (f − fc )l, (29)
where fc is the average azimuth in the analysis window. This formula for divergence con-
siders only vertical changes in reflection dip and neglects lateral changes, which tend to be
smaller. Depending on the value of the scalar c, it has dimensions of unit change in slope
per unit interval in time.
Other reflection patterns

Other reflection patterns include reflection continuity, reflection length, chaotic pat-
terns, hummocky reflections, sigmoidal patterns, and reflection-free.
Stratigraphic reflection continuity is the degree to which reflections have constant
phase and amplitude along individual reflections (Mitchum et al., 1977; Sheriff,
1980, p. 86). This is not necessarily the same concept as reflection continuity in the
context of fault detection. Reflection continuity could be measured similarly for both
purposes by using broad analysis windows for stratigraphic continuity, and narrow
windows for fault detection. Like parallelism, continuous reflections suggest a low-
energy depositional environment, and discontinuous reflections suggest a higher-
energy environment.
Zones of chaotic reflections are characterized by short, nonparallel, discontinuous
reflections of moderate amplitude. Taner (2000) defines a chaotic bedding indicator as
the product of the average deviation in reflection dip with the reflection continuity.
More directly, van Hoek et al. (2010) define chaos as the opposite of parallelism.
Randen et al. (2000) define a “chaos attribute” that is more a measure of dip-corrected
discontinuity.
Hummocky or wavy reflections are characterized by relatively regular lateral changes

in reflection orientation with constant orientation in time. A sigmoidal reflection pattern
comprises a reflection sequence, bounded by unconformities at its top and bottom,

whose reflections diverge updip and converge downdip. Both patterns are difficult to quan-
tify as attributes.
Reflection-free zones are caused by a lack of reflectivity and not a lack of signal.
They are characterized by discontinuous weak amplitudes and resemble zones of
random noise. They are recognized as reflection-free if the surrounding data have well-
defined reflections. Reflection-free zones could represent salt or a thick and uniform
sequence of shale. No attribute identifies reflection-free zones unambiguously.
Examples
Figure 18 compares four stratigraphic attributes: amplitude variance, reflection
spacing, parallelism, and divergence. Amplitude variance is computed as a normalized
standard deviation according to equation 25. A comparison with the reflection strength
of Figure 21b in Chapter 4 shows that strong amplitude variance tends to be associated
Figure 18. Stratigraphic attributes for the seismic data of Figure 21a in Chapter 4. (a) Amplitude
variance. (b) Reflection spacing. (c) Parallelism. (d) Divergence; yellow indicates divergent
reflections and light blue indicates convergent reflections. Reflection spacing is an instantaneous
attribute quantified by equation 26; the other attributes are averages in a window of 3 lines by 3
traces by 7 samples. Every 5th trace of the original seismic data is overlain in variable area format.
with weak amplitudes, and weak variance tends to be associated with strong amplitudes.
Reflection spacing is quantified as an instantaneous attribute according to equation 26. It
resembles the complex trace attribute for phase wavelength because the reflection
spacing is driven more by phase changes than by amplitude changes. It also resembles
complex trace attributes for frequency because the reflection dips are moderate. Parallelism
and divergence present new information and do not resemble other attributes.
Summary
Structural and 3D stratigraphic attributes quantify seismic reflection patterns related
to geological structure and stratigraphy.
Structural attributes are among the most important of all seismic attributes, and include
dip, azimuth, and curvature. The plane-wave destructor and gradient squared tensor are
efficient methods for computing dip and azimuth. Complex trace analysis takes twice
the computation time for comparable results. Dip scanning is likewise slow and, being
quantized, it provides poorer resolution. For qualitative interpretation of geological
structure, dip and azimuth are better combined as apparent dip or as seismic shaded
relief. These attributes produce displays that look like natural illuminated topography
and are easier to interpret than a display of dip or azimuth alone, or even dip and
azimuth combined through a 2D color scale. They are well suited for volume blending
with amplitude or stratigraphic attributes (see Chapter 9). Volume curvature attributes
reveal fine structural detail and highlight places where fractures are more likely to
occur, such as reflection bumps, sags, and flexures.
Three-dimensional stratigraphic attributes are inspired by seismic stratigraphy and
include amplitude variance, reflection spacing, parallelism, and divergence, among
others. Their development has lagged that of structural attributes because stratigraphic
properties are more difficult to quantify than structural properties. Further, while structural
attributes are usefully interpreted individually, stratigraphic attributes are better interpreted
together following the methods of seismic stratigraphy, which is more challenging. The
interpretation of stratigraphic attributes would benefit from multiattribute processes that
combine the attributes as seismic facies with geological meaning.
Seismic discontinuity attributes serve both structural and stratigraphic exploration by
highlighting faults and channels. They are the topic of the next chapter.
Chapter 6
Seismic Discontinuity
Introduction
Seismic discontinuity attributes detect breaks in the continuity of seismic reflections.
The breaks are caused by faults, channels, diapirs, pinch-outs, artifacts, and noise. Discon-
tinuity attributes are applied primarily for mapping faults and channels. They are so helpful
in these essential tasks that they have become the most important of all 3D seismic attributes.
Discontinuity attributes employ energy ratios or derivatives. Discontinuity attributes
based on energy ratios include correlation, semblance, covariance, and weighted corre-
lation discontinuity (Bahorich and Farmer, 1995; Marfurt et al., 1998; Gersztenkorn
and Marfurt, 1999; Van Bemmel and Pepper, 2000; Barnes, 2007). They quantify the
degree to which a set of neighboring traces differ from each other. Discontinuity attributes
based on derivatives measure the degree of change between adjacent traces. They include
gradient squared tensor discontinuity and instantaneous wavenumber magnitude disconti-
nuity. Gradient squared tensor discontinuity is becoming increasingly popular because it
naturally accounts for reflection dip.
Most discontinuity measures are comparable, but the details of their implementa-
tion substantially affect attribute quality and computational cost. These details include
window shape, dip corrections, and filters to produce cleaner attributes.
This chapter reviews seismic discontinuity attributes. It introduces discontinuity
measures based on energy ratios, and considers how to modify them to account for reflec-
tion dip. It then introduces discontinuity measures based on derivatives and discusses how
filters improve discontinuity attributes in general.
Discontinuity based on energy ratios

Consider 3D seismic data to be the sum of a continuous component and a discontinuous
component (Figure 1). Describe seismic discontinuity as that proportion of the data that is
discontinuous. It is mathematically convenient to deal with data energy in place of ampli-
tude because energy is always positive. Following these ideas, discontinuity can be quan-
tified as the ratio of the energy in the discontinuous component to the total energy of the
seismic data. Four prominent measures of seismic discontinuity are based on energy
ratios: crosscorrelation, semblance, covariance, and weighted correlation. They differ
only in how they define discontinuous energy. They are sensitive to waveform shapes,
but only semblance is sensitive to waveform magnitude, or overall amplitude level.
Though these methods differ, their results are nearly identical, and there is no theoretical
reason to favor one method over another.
95
Figure 1. Seismic data can be thought of as the sum of a continuous component and a
discontinuous component.
In this development, reflections are assumed to be relatively flat; the complicating

factor of reflection dip is considered in a following section. Seismic traces are represented
as vectors to simplify the mathematics and promote understanding through graphical
examples. Appendix E reviews the concept of vector traces and sample spaces. The key
idea is that, in a sample space, vector direction represents the shape of the trace waveform
and vector length represents its magnitude. For purposes of illustration, the sample spaces
shown are all 2D, implying seismic traces with only 2 samples, but the ideas are general and
hold for sample spaces of any dimension.
Analysis windows for most discontinuity attributes typically encompass 9 or 25 traces
in a square grid pattern, with a vertical length of 9 to 17 samples. Longer windows tend to
be better for detecting faults, and shorter windows tend to be better for detecting channels.
Energy ratios
The total energy E in a set of seismic traces can be represented as the sum of the energy
of that component of the data that is continuous, “the continuous energy” EC, plus the
energy of that component that is discontinuous, “the discontinuous energy” ED:
E = EC + ED . (1)
This resembles the familiar relation that equates the total energy of seismic data to the sum
of its signal and noise energies. However, discontinuous energy is not the same as noise
energy because it includes the effect of faults and other geological discontinuities. In the
context of discontinuity attributes, the discontinuous energy is the signal.
Define continuity C as the fraction of the total energy that is continuous,
EC
C= , (2)
E
Chapter 6: Seismic Discontinuity 97
and define discontinuity D as the fraction of the total energy that is discontinuous,
ED
D= . (3)
E
Continuity and discontinuity sum to 1:
C + D = 1. (4)
Because energy is always positive, continuity and discontinuity both range from 0 to 1.
A discontinuity value of 0 corresponds to perfectly continuous data, and a value of 1 cor-
responds to perfectly discontinuous data. By these definitions, continuity and discontinuity
are opposites: continuity is the degree to which a seismic trace resembles its neighbors, and
discontinuity is the degree to which a trace differs from its neighbors.
It remains only to define the discontinuous energy. The four common ways to do this
are based on measures of correlation, semblance, covariance, and weighted correlation.
This leads to four distinct discontinuity attributes.
Correlation
The correlation discontinuity attribute crosscorrelates adjacent traces to estimate dis-
continuity. The original continuity attribute of Bahorich and Farmer (1995) is of this
type. It employs two orthogonal crosscorre-
lations derived from three traces in an “L”
pattern. Referring to Figure 2, let c12 rep-
resent continuity in the inline direction
measured as the correlation of traces x1
and x2, and let c13 represent continuity in
the crossline direction measured as the cor-
relation of traces x1 and x3. Combining the
inline and crossline continuities, Bahorich
and Farmer (1995) define the correlation
continuity C as
Figure 2. Map view of the 3-trace L-shaped
C= |c12 · c13 |. (5)
spatial window used in the continuity measure of
Because the window is asymmetric and has Bahorich and Farmer (1995). Each dot
represents the location of a seismic trace.
only three traces, its output is misplaced
Correlating traces x1 and x2 provides a measure
slightly in space and is sharp but prone to of continuity in the inline direction, and
noise. In the original implementation, c12 correlating traces x1 and x3 provides a measure of
and c13 are the maximum crosscorrelations continuity in the crossline direction.
found by independently sliding traces x2 Discontinuity derives from these two
and x3 up or down with respect to trace x1. correlations and is assigned to the location of
This procedure is similar to that employed trace x1.
by Waters (1981, p. 260) in a 2D semblance-

based continuity attribute. It removes the
undesired effect of reflection dips, but it also

removes the desired effect of vertical faults.
It is better to handle reflection dip separately.
A more general attribute defines continu-
ity as the average of the squared zero-lag cor-
relations of a set of M traces with the average
trace xa. Assume the seismic traces have zero
mean and normalize them and the average
trace so that they become unit vectors, which
record only directions. Correlation continuity
C is then the average of the dot products
Figure 3. An illustration in 2D sample between the directions of the M traces with
space, S1 and S2, showing that the correlation the direction of the average trace, x̂a :
coefficient of the average trace xa with the
trace xj is the cosine of the angle uj between
1 M
them. Representing traces as unit vectors, the C= (x̂j · x̂a )2 , (6)
discontinuity energy of x̂j is 12j ¼ sin2uj. M j=1
where x̂j is the direction of the jth trace. Correlation discontinuity D follows as D ¼ 1 2 C.
Correlation discontinuity represents an energy ratio. To show this, refer to Figure 3 and
rewrite equation 6 in terms of the angle uj between vector xj and the average vector xa to
obtain
1 M
C= cos2 uj . (7)
M j=1
Discontinuity D becomes
1 M
1 M
D= sin2 uj = 12 , (8)
M j=1 M j=1 j
where 12j represents the discontinuity energy of the jth trace. This is the ratio of the discon-
tinuity energy to the total energy: the sum of the M discontinuity energies equals the total
discontinuity energy, and because the traces have unit length, the total energy equals the
number of traces M. Because it treats traces as unit vectors, correlation discontinuity
is not affected by differences in trace magnitudes.
Semblance
The semblance discontinuity attribute is based on semblance, which is widely applied
in prestack seismic velocity analysis. Semblance C is a normalized continuity measure
defined as the ratio of the energy of the average trace Ea to the average trace energy in a
set of M traces, E/M, where E is total energy:

MEa
C= ; (9)
E
Appendix E defines trace energy. Semblance
discontinuity follows as D ¼ 1 2 C.
Semblance discontinuity is a measure of
the variance of a set of traces around the
average trace, normalized by the average
energy. To show this, consider a set of M
traces. Let 1j be the distance between the
average trace x a and the jth trace x i (Figure 4).
The discontinuity energy of the jth trace is Figure 4. In sample space, 1j is the distance
the square of 1j, between the average trace xa and the jth
trace xj.
12j = |xj − xa |2 , which expands to (10)
12j = Ej + Ea − 2xj · xa . (11)
The total discontinuity energy ED is the sum of the M trace discontinuity energies:

M
ED = 12j . (12)
j=1
Substituting the expression for 12j into this equation yields

M
ED = (Ei + Ea − 2xj · xa ). (13)
j=1
Hence

M
ED = E + MEa − 2xa · xj , or (14)
j=1
ED = E − MEa . (15)
Dividing each side by total energy E yields the equation for semblance discontinuity,
D ¼ 1 2 C. Thus D represents the variance of the traces around the average trace normal-
ized by the average energy E/M.
Unlike correlation discontinuity, semblance discontinuity is sensitive to differences in
trace magnitude. Traces that have the same shape but differ in magnitude can produce a
strong semblance discontinuity.
Covariance
The covariance or “eigenstructure” dis-

continuity attribute employs principal com-
ponent analysis to quantify the continuity
of a set of zero-mean traces as the ratio of
the first principal component to the total
energy of the traces (Gersztenkorn and
Marfurt, 1999). The first principal com-
ponent l1 is the variance of the data along
the best-fitting line in trace space, and
Figure 5. A crossplot between two seismic serves as an estimate of continuous energy
traces, x1 and x2, with 11 samples. Each point (Figure 5). Covariance continuity is thus
represents the trace amplitudes at a different
the energy ratio
sample index. The closer the points lie to a
straight line, the more the two traces resemble l1
each other. The dashed line is the best-fitting line C= . (16)
in a least-squares sense. Projecting the points E
perpendicularly onto the line, the first principal
Covariance discontinuity follows as D ¼
component is the variance of the points about
1 2 C. The first principal component
the origin.
ranges in value from E/N to E, where N
is the number of trace samples in the
window. Thus covariance discontinuity ranges in value from 0 to a maximum of
(N 2 1)/N. More involved discontinuity measures are sometimes formed by combining
other principal components (Cohen and Coifman, 2002), but they do not produce better
results.
To find the first principal component of a set of traces, construct a covariance matrix
whose elements are the trace variances and covariances, according to equation E-16 in
Appendix E. The total energy of the seismic data equals the sum of the variances along
the main diagonal of the matrix. It remains only to find the largest eigenvalue of the
matrix, which is the first principal component l1.
If the number of trace samples is less than the number of traces, then the matrix is
constructed with sample vectors instead of trace vectors. A sample vector is a collection
of all the seismic trace values at a given trace sample index.
Weighted correlation
The weighted correlation discontinuity attribute is similar to correlation discontinuity,
with the difference that the constituent squared correlations are weighted by the trace ener-
gies. Graphically, this method represents discontinuity in sample space as a normalized
variance about a “line of similarity” that passes through the origin and the average trace
(Figure 6). The dissimilarity of any trace is its distance from the line of similarity. This
method approximates covariance discontinuity but is simpler and does not require zero-
mean traces.
Consider two traces to be similar if they

have the same shape, though their magni-
tudes may differ. Thus two traces are

similar if their unit vectors are equal. To
extend this idea to a set of traces, consider
the traces to be similar to the extent that
they have the same direction as the aver-
age trace. Traces that are exactly similar
to the average fall on the line of similarity,
traces that are closely similar lie close to
the line, and traces that are dissimilar lie
far from the line. The continuous energy
is the sum of the individual weighted corre-
lations, and continuity C is
1 M
C= (xj · x̂a )2 . (17) Figure 6. The line of similarity for a set of
E j=1
seismic traces coincides with the direction of
the average trace vector xa. The discontinuity
Discontinuity follows as D ¼ 1 2 C. of any trace xj is the square of its distance 1j
That C represents a weighted average from this line.
correlation becomes clear when equation
17 is recast as
M
Ej (x̂j · x̂a )2
j=1
C= . (18)

M
Ej
j=1
Trace energy Ej acts as a weighting function. This continuity measure reduces to the
standard correlation measure of equation 6 if the trace energies are each set to 1.
Weighted correlation discontinuity represents the normalized variance of the disconti-
nuity energy around the line of similarity. To show this, let 1j be the distance that trace x j
lies from the line of similarity. Define the discontinuity energy of the jth trace as this dis-
tance squared,
12j = xj · xj − (xj · x̂a )2 = Ej − (xj · x̂a )2 . (19)
The total discontinuity energy ED is the sum of the M individual discontinuity energies 12j .
Discontinuity D is therefore
1 M
D= 12 , (20)
E j=1 j
which is the variance around the line of similarity normalized by the total energy.
Comparison of energy-ratio discontinuity attributes

Figure 7 compares the four discontinuity methods presented above computed in a

window of 9 traces in a 3 by 3 pattern with 11 time samples. The results are nearly indis-
tinguishable. The computational cost of covariance discontinuity is several times that of
the three other methods, but otherwise there is scant reason to favor one method over
another. Semblance discontinuity is most common in commercial applications.
Dip corrections
Energy-ratio discontinuity attributes neglect the influence of reflection dip. They
perform well where dips are relatively flat but fail where dips are steep. This is because
a) b)
c) d)
Figure 7. A comparison of discontinuity attributes along a time slice showing there is little
difference between them. (a) Correlation. (b) Semblance. (c) Covariance. (d) Weighted correlation.
Strong discontinuities are red.
they scan horizontally across seismic traces and so confuse reflection dip for valid seismic
discontinuity. This problem is addressed by structurally guided processing. In each analysis
window, the seismic data are flattened before computing discontinuity (Luo et al., 1996;
Marfurt et al., 1999). In effect, the top and bottom of the analysis window are adjusted
to align with the locally dominant dip. The discontinuity computations proceed along
this dip and thus along reflections (Figure 8). To do this, reflection slopes are estimated
and the seismic data are interpolated along the slopes. Slope estimation performs
better when the window for dip analysis is broader but shorter than the window for the dis-
continuity computation. Figures 9 and 10 compare discontinuity computations with and
without dip corrections. Dip corrections improve fault images markedly in regions of
steep dip, but they increase the computational cost by about five times.
Faults confuse estimates of local reflection dip, thereby degrading dip-corrected dis-
continuity estimates. To avoid this, dips could be estimated through a nonlinear or nonpla-
nar process (Marfurt, 2006). This approach has merit, but it is seldom explored because it is
computationally expensive.
Figure 8. Discontinuity computations should

proceed along reflection dips. Here the red window
is suitable because it follows reflection dips,
whereas the yellow window is appropriate only for
flat reflections. Window sizes are exaggerated for
illustration.
Figure 9. Illustration on a vertical section of the effect of dip corrections on weighted-correlation

discontinuity computed in a window of 3 lines by 3 traces by 11 samples. (a) Original seismic
data. (b) Discontinuity without dip correction, and (c) with dip correction. Dip-corrected
discontinuity is cleaner in regions of steep dip.
Discontinuity based on
derivatives
Discontinuity measures based on

derivatives include wavenumber disconti-
nuity and gradient squared tensor disconti-
nuity. Wavenumber discontinuity merits
investigation, but it is not competitive
with standard discontinuity measures and
has found little application. In contrast,
gradient squared tensor discontinuity is
superior to most other discontinuity attri-
butes, and it is becoming increasingly
popular.
Wavenumber discontinuity
Faults are characterized by abrupt
changes in reflection amplitude and phase.
These abrupt changes cause spikes in
instantaneous frequency and wavenum-
ber. Luo et al. (1996) take advantage of
this to develop a high-resolution discon-
tinuity attribute based on instantaneous
horizontal wavenumber magnitude.
Horizontal wavenumber magnitude kh
combines the instantaneous wavenumbers
in the x and y directions, kix and kiy, accord-
ing to

kh = kix2 + kiy2 (21)
(refer to Chapter 4). A short median filter

in time improves this measure as a discon-
tinuity attribute, making it comparable to
a standard discontinuity attribute without
dip corrections (Figure 11). Wavenumber
Figure 10. Illustration on a time slice of the
discontinuity must be scaled so that its
effect of dip corrections on weighted-correlation
discontinuity computed in a window of 3 lines by values fall in the range of 0 to 1. Wave-
3 traces by 11 samples. (a) Original seismic data. number discontinuity is slower and more
(b) Discontinuity without dip correction. involved than standard discontinuity.
(c) Discontinuity with dip correction. The dip- In principle, a measure for wavenum-
corrected discontinuity is cleaner in regions of ber discontinuity should be computed
steep dip, particularly around the diapirs. along reflection dip, as is done for other
discontinuity attributes. This is accom-

plished readily. The average phase wave-
number unit vector k̂a is defined as

kki l
k̂a = , (22)
|kki l|
and is normal to the plane of the domi-
nant reflections. The projection of the
instantaneous wavenumber ki onto this
plane is the dip-corrected wavenumber dis-
continuity kr. It is given by
kr = |ki × k̂a |. (23)
The measure must be normalized to make

its values fall in the range of 0 to 1.
Figure 12 compares wavenumber dis-
continuity without and with dip correction.
Dip correction produces an inferior image.
Gradient squared tensor Figure 11. (a) Weighted-correlation

discontinuity discontinuity computed in a window of 3 lines
by 3 traces by 11 samples. (b) Wavenumber
Gradient squared tensor discontinu- discontinuity smoothed in time with an
ity has the advantage over energy-ratio 11-sample median filter. The two discontinuity
measures in that it implicitly computes attributes are roughly similar.
Figure 12. Wavenumber discontinuity (a) without dip correction, and (b) with dip correction.
continuity along reflection dips. It employs the same gradient squared tensor described in
Chapter 5 for computation of dip and azimuth. For computing discontinuity, the three
eigenvalues of the gradient squared tensor matrix are needed, but not the eigenvectors.
The largest eigenvalue, l1, is the squared gradient in the direction normal to the reflections.
The second largest eigenvalue, l2, is the maximum squared gradient in a direction parallel
to the reflections. If a clear fault is present, this direction is roughly normal to the fault.
The smallest eigenvalue, l3, is the squared gradient in the direction of least change,
which is parallel to the reflections and faults and tends to record noise. For clean and
continuous seismic data, l1 is much larger than the other eigenvalues. For very noisy
data, the three eigenvalues are roughly comparable. For clean seismic data with a fault,
l2 is relatively large but l3 remains small. Randen et al. (2000) empirically define a
continuity measure in terms of these eigenvalues, which they call “chaos.” Expressed as
a discontinuity attribute D, their measure is
2l2
D= . (24)
l1 + l3
A similar and simpler discontinuity measure is
l2
D= , (25)
l1
which derives from the 3D image planarity

(Hale, 2009). A third alternative, which pro-
duces modestly cleaner results, is
l2 − l3
D= . (26)
l1
All three measures range from 0 for per-

fectly continuous data to 1 for perfectly dis-
continuous data.
Figure 13 compares a standard dip-cor-
rected discontinuity with gradient-squared
discontinuity computed through equation
26. Results and computation times are
comparable.
Figure 13. (a) Weighted-correlation Relative amplitude change

discontinuity with dip corrections computed in a
3 line by 3 trace by 11 sample window. Marfurt and Kirlin (2000) introduce
(b) Gradient squared tensor discontinuity the coherent amplitude gradient to record
computed in a comparable Gaussian-tapered changes in the reflection strength along
window. reflections. Based on principal component
analysis and dip corrections, their attribute reveals detail hidden in the amplitudes and acts
like a high-resolution directional discontinuity attribute.
Relative amplitude change is similar to the coherent amplitude gradient but is simpler,
being just a component of the instantaneous group wavenumber introduced in Chapter 4.
It suffers from effects of reflection interference. To serve as a discontinuity attribute,
these effects must be reduced. They are reduced through median filtering or, like the coher-
ent amplitude gradient, by computing amplitude change along reflection dip. Median filter-
ing in a short vertical window removes noise and spikes due to reflection interference,
while retaining faults and channels. Computing relative amplitude change along reflections
is more involved and much slower.
To compute relative amplitude change along reflection dip, quantify the slopes of the
dominant reflections by the average phase wavenumber vector k within a vertical plane in
the x or y direction. Relative amplitude change along reflections is found by projecting the
instantaneous group wavenumber vector onto a line parallel to the dominant reflection. In
terms of the instantaneous group and average phase wavenumber components, the relative
amplitude change along reflection dips in the x direction, gxr, is
gix kz − giz kx
gxr = . (27)
kx2 + kz2
Similarly, relative amplitude change along dips in the y direction, gyr, is
giy kz − giz ky
gyr = . (28)
ky2 + kz2
Figure 14 compares relative amplitude change computed both horizontally and along
reflection dips. Computing along dip reduces the unwanted spikes due to reflection
Figure 14. A comparison on a seismic line between (a) relative amplitude change computed
horizontally, and (b) relative amplitude change computed along reflection dip. Yellow arrows
indicate computation directions.
Figure 15. A comparison of methods for relative amplitude change on a time slice. The yellow
arrow indicates the computation direction. (a) Horizontal relative amplitude change. (b) Relative
amplitude change along dip. (c) Horizontal relative amplitude change followed by a 7-sample
median filter in time. (d) Relative amplitude change along dip followed by a 7-sample median
filter in time.
interference but retains spikes due to discontinuities. A substantial improvement is derived

by applying a median filter to horizontal relative amplitude change (Figure 15). Median fil-
tering is simpler than computing changes along dip, and it suffices for most applications.
Being directional, relative amplitude change naturally looks illuminated when
displayed in monochrome. It should be run in orthogonal directions to capture all
discontinuities.
Improving discontinuity attributes

Discontinuity attributes are improved through tapered windows or Laplacian filtering.
More significantly, “fault filters” transform a discontinuity attribute into a fault attribute.
Tapered windows
Like any seismic attribute, discontinuity attributes should employ tapered windows
to minimize artifacts. However, most implementations of energy-ratio discontinuity
attributes implicitly employ boxcar windows in efficient roll-along computations. Boxcar

windows cause striping that parallels reflections. The stripes are obvious on vertical sections
but not on time slices. Because discontinuity attributes are largely interpreted along time
slices, boxcar windows are acceptable. Sometimes vertical smoothing is incorporated in
the algorithm to reduce striping.
Gradient squared tensor methods are usually based on Gaussian derivatives, in which
case the windows are inherently tapered. This is another advantage that gradient squared
tensor discontinuity enjoys over competing methods.
Laplacian filtering
Laplacian filtering is a method of 2D image processing that sharpens edges and
improves resolution. Laplacian filter operators approximate a second-order 2D derivative.
They take various forms; Figure 16 illustrates a common form. Laplacian filtering is
applied to seismic discontinuity attributes along horizontal slices, or, alternatively, along
reflections. It efficiently sharpens faults and removes the background trend and low wave-
number noise, but it also increases high wavenumber noise (Figure 17).
Figure 16. A 2D Laplacian operator for seismic discontinuity

attributes. It is flipped in sign from a standard Laplacian operator so that
a filtered attribute retains the original polarity. Laplacian filtering is
applied along horizontal slices or along reflections to sharpen faults.
Figure 17. (a) A discontinuity attribute viewed along a time slice. (b) The discontinuity attribute
with Laplacian filtering. Laplacian filtering sharpens discontinuity images on horizontal slices.
By removing low wavenumber signal and noise, it balances the amplitudes and brings out subtle
features.
Fault attribute
The chief application of seismic discontinuity attributes is to aid fault interpretation.

For this purpose, discontinuities not due to faults are noise and should be removed. Non-
fault discontinuities tend to be relatively nonplanar or nonvertical, or small in extent or
amplitude; it is on this basis that they are removed through methods of fault filtering.
The discontinuities that remain after fault filtering are presumed to represent fault
segments. Methods of image processing can then be applied to extend and connect the
fault segments. The final result is a fault attribute, the
key to automatic fault interpretation. The most success-
ful method for transforming a discontinuity attribute
into a fault attribute is based on ant tracking (Pedersen
et al., 2002). Many other methods have been proposed,
and the topic remains an area of active research (Ash-
bridge et al., 2000; Randen et al., 2001; AlBinHassan
and Marfurt, 2003; Tingdahl and Hemstra, 2003;
Jacquemin and Mallet, 2005; Barnes, 2006). Figure 18
illustrates a basic workflow for fault filtering and predic-
tion based on image-processing techniques, and
Figure 19 illustrates fault filtering applied to a standard
discontinuity volume.
Summary
Figure 18. A workflow based on Discontinuity attributes are indispensable aids for
image-processing methods to the interpretation of faults and channels. After ampli-
transform a discontinuity attribute tude attributes, they are the most important seismic
into a fault attribute. attributes.
Figure 19. Fault filtering cleans up discontinuity attributes. (a) Original seismic data along a time
slice. (b) Discontinuity attribute. (c) Discontinuity after fault filtering.
Discontinuity attributes measure variances or derivatives. Those that measure var-

iances can be represented as energy ratios, and include correlation, semblance, covariance,
and weighted correlation. These attributes differ in how they define continuous energy, but
they produce nearly identical results. Most discontinuity attributes are of this type.
Energy-ratio discontinuity attributes do not account for reflection dip inherently. They
require dip corrections to produce correct images where dips are steep. Dip corrections are
usually planar. Nonplanar dip corrections could provide better images near faults, but their
high computational cost inhibits their adoption.
Discontinuity attributes that employ derivatives are slightly more involved, and they
include wavenumber discontinuity and gradient squared tensor discontinuity. Wavenum-
ber discontinuity is an intriguing idea, but it is neither as good nor as fast as competing
measures. It is seldom encountered. In contrast, gradient squared tensor discontinuity is
steadily gaining popularity because it produces superior results that naturally account for
reflection dip.
Discontinuity attributes sometimes perform better when derived from coherency-
filtered seismic data. They are also improved or enhanced by direct filtering. The most
common filter is the 2D Laplacian, which is applied along time slices to improve image
resolution. More significantly, discontinuity attributes can be transformed into fault attri-
butes through sophisticated fault filters. Such filters hold great promise for the future.
This completes the development of 3D seismic attributes. The book now introduces
advanced methods of 1D attribute computation and analysis, beginning with spectral
decomposition and waveform classification, the topics of the next chapter.

Chapter 7
Spectral Decomposition and Waveform

Classification
Introduction
Spectral decomposition and waveform classification are popular and sophisticated
methods for generating and analyzing seismic attribute maps. Though they differ
greatly, they have much in common because both compare waveforms in a narrow interval
of seismic data to reveal channel systems and other stratigraphic features.
Spectral decomposition applies time-frequency analysis to produce a set of maps of the
relative strength of different frequencies. Applying the ideas of thin-bed tuning, the maps
are analyzed to estimate bed thicknesses. Waveform classification maps similar regions by
comparing waveforms observed along a horizon to a set of template waveforms. The tem-
plate waveforms are either supplied by the geophysicist or found through methods of auto-
matic pattern recognition. Spectral decomposition and waveform classification both
require carefully interpreted horizons and carefully selected data intervals.
Spectral decomposition is sometimes applied to entire seismic volumes. The interpret-
ation of volume spectral decomposition is facilitated through red-green-blue (RGB) color
blending. There is no practical way to apply waveform classification to entire volumes.
This chapter presents spectral decomposition and seismic waveform classification. It
begins with a discussion of thin-bed tuning, which underlies the analysis of spectral
decomposition. It describes the concept of filter banks and their application in spectral
decomposition. Finally, it broadly reviews unsupervised waveform classification.
Thin beds
A thin bed is a rock layer, imaged by seismic data, whose thickness is small with
respect to the wavelengths of the seismic wavelet. Reflections from the upper and lower
surfaces of a thin bed interfere so that they cannot be distinguished from each other,
making it difficult to estimate the bed thickness. This is the thin-bed problem. It is important
because many exploration targets are thin beds, including stream channels, point bars,
shoestring barrier sands, and sand-prone delta lobes. Thin-bed thicknesses can be estimated
through tuning analysis based on the thin-bed spectral response.
Thin-bed model
Tuning analysis proceeds from a thin-bed model. The most useful thin-bed model is a
buried isolated sand channel embedded in shale. The sand channel is much broader than it
113
Figure 1. Cross section through a model of a buried sand channel embedded in shale. The sand
channel has thickness D and acoustic impedance Z2 that is less than the acoustic impedance of the
shale, Z1. This model produces reflections with equal magnitude but opposite polarity. The top
reflection coefficient is – r at time t1; the bottom coefficient is +r at time t2. The channel thickness in
time is tb.
Figure 2. (a) A channel imaged on a seismic line as a bright thin bed; negative values are blue and
positive values are orange. (b) The corresponding reflection strength with every other seismic trace
overlain in wiggle trace format; strong reflection strengths are red, weak are blue.
is thick, and its acoustic impedance is lower than that of the shale. Assume that the impe-
dance contrast is large enough to produce a strong composite reflection, yet small enough to
neglect the energy lost in reflection from the top. In this case, the reflections from the
channel top and bottom have the same magnitude. This model produces a negative reflec-
tion over a positive reflection; let the top reflection coefficient be –r and the bottom be +r
(Figure 1). Figure 2 shows a channel thin bed imaged on a seismic line.
The reflectivity series u(t) for the thin-bed model is the sum of two scaled delta
functions,
u(t) = −r d(t − t1 ) + rd(t − t2 ), (1)
where t1 and t2 are the reflection times of the top and bottom of the thin bed. Reflection
magnitude r controls the brightness of the thin-bed response but has no effect on tuning.
Chapter 7: Spectral Decomposition and Waveform Classification 115
The model ignores secondary effects such as multiples, noise, transmission losses, and non-
normal incidence because they are small compared with tuning.
The spectral response of the thin bed is the Fourier transform U( f ) of the thin-bed
reflectivity u(t):
U( f ) = −i2r exp(−i2p f tc ) sin (p f tb ), (2)
where tc is the average of reflection times t1 and t2 and represents the center of the thin bed,
and tb is the time difference t2 – t1 and represents bed thickness. The magnitude of U( f ) is
the amplitude spectrum Au( f ):
Au ( f ) = 2|r sin (p f tb )| (3)
(Figure 3). Discounting the linear component of the phase, which is a function solely of
time tc, the corresponding phase spectrum uu( f ) of U( f ) is
−p
uu ( f ) = sign[r sin (p f tb )]. (4)
2
Depending on the frequency, the phase is either –908 or +908. For frequencies below
1/tb, the phase is a constant – 908. A thin bed ceases to be truly “thin” when its spectrum
extends much beyond 1/tb, so thin beds tend to have constant phase.
Seismic tuning analysis is the study of the spectral response of thin beds to determine
spectral maxima and minima.
Tuning
Tuning is the strong constructive or destructive interference of two or more reflected
seismic wavelets. Tuning depends on the frequency content of the wavelets and the
spacing and reflection coefficients of the reflectors. The frequencies at which
maximum constructive or destructive interference occurs are the tuning frequencies.
Figure 3. Amplitude spectrum of a thin-bed model with a time thickness of 12 ms. Thin-bed
spectra tend not to extend much beyond the second spectral notch, which here occurs at 83 Hz.
The peaks of a thin-bed spectrum corre-

spond to constructive tuning frequencies,
and the notches correspond to destructive

tuning frequencies. The constructive tuning
frequencies fn form the series
n
fn = , n = 1, 3, 5 . . . . (5)
2tb
The lowest constructive tuning fre-
quency is the frequency of the first spectral
peak frequency fp and is informally termed
“the tuning frequency.” The thin-bed thick-
Figure 4. Illustration of tuning for a thin bed ness in time tb is half the period of the
with opposite polarity reflections of first peak frequency,
magnitude r. (a) A sinusoid of frequency 1/2tb
leads to perfect constructive interference.
(b) A sinusoid of frequency 1/tb leads to 1
tb = (6)
perfect destructive interference. 2fp
(Figure 4). The corresponding thickness in depth db, for a thin-bed velocity v, is
v
db = . (7)
4fp
Thus, in principle, thin-bed thickness can be estimated from the spectral response of
the thin bed.
Thin-bed response
The response of the thin bed u(t) to a seismic wavelet w(t) is the composite reflection
x(t), formed as the convolution of the wavelet with the thin bed. The thin-bed response in
the frequency domain is the spectrum of the composite wavelet X( f ), which is the product
of the wavelet spectrum W( f ) with the thin-bed spectrum U( f ). Hence
F
x(t) = w(t) ∗ u(t) ⇐⇒ X( f ) = W( f )U( f ). (8)
The behavior of the composite reflection as a function of thin-bed thickness is

illustrated by the synthetic seismogram of a wedge model (Figure 5). This behavior is
summarized by equations for the total energy and average frequency of the composite
reflection as a function of bed thickness. Consider a zero-phase seismic source wavelet
with boxcar spectrum Aw( f ) ¼ 1 from low frequency fl to high frequency fh. Boxcar
spectra are unrealistic for seismic wavelets, but they serve the purpose of illustration
with simple mathematics.
Figure 5. (a) A model of a sand wedge embedded in a shale that has higher acoustic impedance.
(b) A synthetic seismogram derived from the wedge model given a seismic wavelet with a boxcar
spectrum from 12.5 to 50 Hz. The large trough is taken to be the reflection from the top, and the large
peak is taken to be the reflection from the bottom; both are marked by red lines. The apparent
thickness of the wedge is the distance between the red lines. The tuning thickness, which is that bed
thickness that produces the strongest amplitude, is marked by the dashed blue line.
Referring to the equation for spectral energy, equation 31 in Chapter 3, the energy E(tb)
of the composite reflection from a thin bed as a function of bed thickness tb is

1 fh
E(tb ) = 2 A2w ( f )A2u ( f )df = 2 A2u ( f )df . (9)
0 fl
Recalling equation 3 for Au( f ), the energy is proportional to

fh
E(tb ) / sin(p f tb )2 df , or (10)
fl
fh − fl 1
E(tb ) / + [sin(2pfl tb ) − sin(2pfh tb )]. (11)
2 4ptb
Referring to the equation for average spectral frequency, equation 32 in Chapter 3, the
average frequency fa of the thin-bed composite reflection is

1 fh
1 1
fa (tb ) = fA2w ( f )A2u ( f )df = f sin (p ftb )2 df . (12)
E(tb ) E(tb )
0 fl
Integrating yields the formula for the average frequency of the composite reflection.
Of prime interest is the average frequency of a thin bed as its thickness approaches zero.
This is found by considering that a thin bed acts like a derivative operator when convolved
with a wavelet. The amplitude spectrum of a derivative operator is proportional to fre-
quency f. Hence the power spectrum of the composite wavelet is proportional to f 2 and
its average frequency is
fh
f 3 df
fl 3 fh4 − fl4
fa (0) = = . (13)
fh 4 fh3 − fl3
f 2 df
fl
For sufficiently broad bandwidth, this reduces to fa(0) 3fh/4. This is higher than the
average frequency of the seismic wavelet, which approaches fh/2 for a wavelet with suffi-
ciently broad bandwidth. Thus reflections from thin beds with equal and opposite polarity
exhibit higher average frequency.
Figure 6 shows the average frequency and total energy of a composite reflection from a
thin bed as a function of bed thickness; the seismic wavelet has a boxcar spectrum from
12.5 to 50 Hz. The total energy curve closely matches the amplitude curve of Kallweit
and Wood (1982, Figure 13), which employs the same wavelet. Kallweit and Wood
measure reflection amplitude as the peak absolute amplitude of the composite wavelet,
and they define the tuning thickness to be that thickness at which the reflection amplitude
is maximum. By their measure, the tuning thickness in this example is 14 ms. The tuning
thickness could also be defined as that thickness at which the total energy of the composite
wavelet is maximum, which is 13 ms here. This definition yields closely comparable esti-
mates and mathematically is more tractable.
Figure 6. Total energy and average frequency of a thin bed as a function of bed thickness for a
wavelet with a bandpass spectrum of 12.5 to 50 Hz. The tuning thickness is 13 ms. The Widess limit
of resolution is 8 ms.
Below the tuning thickness, total energy drops off almost linearly while average fre-
quency increases slightly. This implies that thin beds with thicknesses below the tuning
thickness produce composite reflections that look the same except for their magnitudes.
Widess limit of resolution

The Widess limit of resolution is an empirical estimate of the thinnest thin bed whose
thickness can be measured given a particular seismic wavelet. It is defined as one eighth the
dominant wavelength of the wavelet (Widess, 1973). Widess implicitly associates the cor-
responding dominant frequency with the tuning frequency. Assuming that the average fre-
quency fa approximates the dominant frequency, this estimate is expressed more
conveniently as one quarter of the period of the average frequency of the wavelet. For a
wavelet with a boxcar spectrum from low frequency fl to high frequency fh, the Widess
limit of resolution tW becomes
1 1
tW = = . (14)
4fa 2( fl + fh )
The Widess limit of resolution is handy and widely applied. However, it serves only as
a rough estimate of resolution because it neglects the influence of bandwidth; the greater
the bandwidth, the better the resolution. As a rule of thumb, bandwidth should be 2 or
more octaves for the Widess limit to apply.
Uncertainty principle
The resolution of a seismic waveform in time is related to its length; its corresponding
resolution in frequency is related to its bandwidth. Waveform length and bandwidth
cannot be arbitrarily small simultaneously. Beyond a certain point, shortening a waveform
necessarily broadens its bandwidth, and reducing its bandwidth necessarily lengthens the
waveform. This fact is expressed by the uncertainty principle. In the context of waveforms
and signals, the uncertainty principle has the same form as the famous uncertainty prin-
ciple of quantum mechanics, but it is purely a mathematical construct without physical
implications (Cohen, 1995, p. 195). Its name is misleading because it expresses a definite
property of waveforms.
The uncertainty principle states that, for an arbitrary waveform, the product of its
length with its bandwidth can be no smaller than some constant, which depends on how
length and bandwidth are defined. It is convenient and customary to define them as standard
deviations (Berkhout, 1984, pp. 22, 28; Cohen, 1995, p. 46). Define waveform length in
time tl as the standard deviation of the instantaneous power a 2(t) around the center time tc:
1
(t − tc )2 a2 (t)dt
tl2 = −11 , where (15)
−1
a2 (t)dt
1
ta2 (t)dt
tc = −1
1 . (16)
−1
a2 (t)dt
Similarly, define bandwidth fb as the standard deviation of the spectral frequency around
the mean frequency as given by equation 38 in Chapter 3.
With these definitions for waveform length and bandwidth, the uncertainty principle is
1
tl fb ≥ . (17)
4p
The product of waveform length and bandwidth is called the uncertainty; the minimum
uncertainty is 1/4p. Waveforms with minimum uncertainty have the shortest time length
for a given bandwidth, or, conversely, the narrowest bandwidth for a given time length.
Gabor wavelets have minimum uncertainty. For any given amplitude spectrum, the wave-
form with the shortest time length has a constant phase spectrum. Thus minimum uncer-
tainty also implies constant phase.
The uncertainty principle as expressed above is awkward because it quantifies length
and bandwidth as standard deviations instead of as more intuitive measures that quantify
the entire length and bandwidth, which are several times longer. These properties are use-
fully approximated
√ by arbitrarily defining the full time length Tl and full bandwidth Fb as
representing 2 p times tl and fb respectively. With these measures, the uncertainty prin-
ciple becomes
Tl Fb ≥ 1. (18)
This form of the uncertainty principle is more convenient and suffices for seismic attribute
analysis. Claerbout (1992, p. 248) arrives at the same expression through different
definitions.
The uncertainty principle implies that short waveforms naturally have broad band-
width, and narrowband waveforms naturally are long in time. This fundamental fact
underlies all of seismology. Seismic data interpretation demands the best possible resol-
ution in time, so it requires the shortest possible seismic wavelet. If a wavelet has constant
phase, then it can be made shorter only by increasing its bandwidth. For spectral decompo-
sition, narrow bandwidth is as important as a short waveform. These goals compete and
require compromise.
Spectral decomposition
Spectral decomposition applies time-frequency analysis to seismic data to produce fre-
quency maps or volumes. A frequency map derives from the seismic data within a horizon-
guided interval through a volume, and a frequency volume derives from the entire seismic
volume. Both represent the spectral power in a narrow frequency range. Maps are far easier
to deal with. A set of frequency maps typically numbers about 100, and may be organized
as a “tuning volume” to facilitate interpretation. In such a volume, the vertical scale is fre-
quency and horizontal slices are constant frequency maps. Frequency volumes are derived
much like maps. Because it is difficult to interpret multiple volumes simultaneously, their
number is limited to about 5 and the filter bandwidths are increased correspondingly.
Frequency maps distinguish reflections and thin beds on the basis of their tuning
responses, which are a function of bed thicknesses. They sometimes reveal features that
cannot be seen with reflection strength or other attributes, and they have modest inherent
geological meaning in that they suggest bedding thickness. For example, thin beds that
are most prominent on high-frequency maps are relatively thin, and thin beds that are
most prominent on low-frequency maps are relatively thick. The thickness of an isolated
thin bed is estimated from its spectrum. This estimate is reliable to the extent that the
thin bed is free of interfering reflections, the bandwidth is sufficient to resolve the thickness,
and the analysis interval is centered on the thin bed.
There are many and varied methods for spectral decomposition. Modern methods, such
as the constrained least-squares spectral analysis of Puryear et al. (2012), aim to improve
signal localization in both time and frequency simultaneously. Such methods show
promise, but they have yet to become important in commercial applications and are not dis-
cussed further here. More typically, spectral decomposition is accomplished through filter
banks, the most basic form of time-frequency analysis. A filter bank is an ordered set of
narrowband filters. Each filter is characterized by a center frequency and covers a different
subset of the signal bandwidth; taken together, the filters span the entire signal bandwidth
(Figure 7). The number of filters is chosen according to the needs of interpretation. The
filter bank is applied to the seismic data to produce a corresponding set of filtered responses,
each of which leads to a separate attribute.
The discrete Fourier transform and the wavelet transform can be treated as filter banks.
Both are employed in spectral decomposition in a form that restricts the analysis to the fre-
quency band of the seismic signal, with a frequency sample interval that is convenient for
Figure 7. A filter bank is an organized set of narrowband filters. Filter banks are applied in tuning
analysis and for determining the time-variant frequency content of the seismic signal.
interpretation. The frequency sample interval is often set at 1 Hz, which in typical appli-
cations is much smaller than the bandwidth implied by the uncertainty principle (Hall,
2006). This is justified as it helps identify peak spectral frequencies more accurately.
The filter bank is applied to seismic data, and the filtered output is converted to reflec-
tion strength. Each filter produces a single value at each map location. In practice, a differ-
ent approach is taken that yields nearly the same result. The filter-bank wavelets are made
complex and their real and imaginary responses are combined to yield the spectral magni-
tudes for the center frequencies of the wavelet.
This is the essence of spectral decomposition. It remains only to describe the details of
the filter wavelets and aspects of interpretation.
Wavelets
The filters in a filter bank for spectral decomposition are usually Gabor or Morlet wave-
lets, which have the property of minimum uncertainty. Gabor and Morlet wavelets are
essentially the same: both are formed as the product of a Gaussian window with a
complex sinusoid (Figure 8). The distinction between them is how they behave in a filter
bank. Gabor wavelets all have the same length, envelope, bandwidth, and spectral shape
(Figure 9). Morlet wavelets all have the same form derived by stretching or squeezing a
basis wavelet, or “mother wavelet” (Figure 10). Their lengths are inversely proportional
to the center frequencies, so they are long at low frequencies and short at high frequencies.
Their bandwidths are proportional to the center frequencies and encompass a constant
number of octaves. Morlet wavelets are constant-Q filters, where in this context Q refers
to the ratio of the wavelet’s center frequency to its bandwidth.
Consider two filter banks that cover the same frequency range, one built with Gabor
wavelets and the other with Morlet wavelets. Their wavelets differ at the ends of the fre-
quency range but are comparable in the middle. In the analysis of narrowband seismic
data, with bandwidth less than 2 octaves, it
matters little which wavelet is used. In the
analysis of broadband seismic data, with
bandwidth exceeding 2 or 3 octaves, Morlet
wavelets provide sharper images at higher
frequencies because they are shorter, and
they provide more reliable images at low fre-
quencies because they are longer. These
properties offer a distinct advantage in the
analysis of thin beds over a wide range of
thicknesses. Gabor wavelets have trouble
with especially thick or thin features
because they have fixed length. A wavelet
that is longer than a thin bed will likely
overlap other reflections that mask the thin-
Figure 8. The Gabor wavelet is the product bed response. A wavelet that is shorter than
of a sinusoid with a Gaussian window. the thin bed cannot capture it fully to find
its tuning frequency. At low frequencies, Gabor wavelets produce maps that resemble
average amplitude. A disadvantage of Morlet wavelets is that a high-frequency map rep-
resents less data than a low-frequency map, which complicates comparison. In contrast,
Figure 9. Gabor wavelets and their power spectra; dashed lines are envelopes. Gabor wavelets in a
set have the same envelope and equal bandwidth as measured in hertz.
Figure 10. Morlet wavelets and their power spectra; dashed lines are envelopes. Morlet wavelets
in a set are stretched or squeezed versions of the same basis wavelet. They have equal bandwidth as
measured in octaves.
a set of frequency maps created with Gabor wavelets all represent the same data because the
Gabor wavelets have constant length. In general, Morlet wavelets are preferable in spectral
decomposition.
Both Gabor and Morlet wavelets appear in commercial software for spectral decompo-
sition. One popular software tool implements a Fourier transform on the seismic data
within a Gaussian window. In effect, it employs Gabor wavelets in the decomposition.
Other tools implement wavelet transforms with Morlet wavelets.
Complex Gabor wavelets wg(t) have the form

t2
wg (t) = exp − 2 + i2p fc t , (19)
2s
where s controls the length in time and fc is the sinusoidal frequency. For spectral
decomposition, the length must be long enough so that the wavelet has negligible
average value, and thus a negligible dc component. In this case, the frequency spectrum
Wg( f ) of the real part of wg(t) is
√
Wg ( f ) = 2ps exp [−2p2 s2 ( f − fc )2 ]. (20)
Here, s is inversely related to the bandwidth and fc is the center frequency of the spectrum.
It is often desirable that the Gabor wavelets
√ act as unscaled filters so that Wg( fc) ¼ l. This
requires dividing the wavelet by 2ps. For a set of N Gabor wavelets, s is a constant and
the center frequencies are given by
n−1
fc (n) = fl + ( fh − fl ), 1 ≤ n ≤ N, (21)
N−1
where fl is the lowest center frequency in the set and fh is the highest.
Complex Morlet wavelets wm(t) are defined by replacing s with fc/h to obtain
2 2
fc −t fc
wm (t) = exp + i2pfc t , (22)
h 2h2

where h is a constant that defines the basis wavelet. Scaling factor fhc ensures that all
wavelets have the same energy. Again assuming a negligible dc component, the frequency
spectrum Wm( f ) of the real part of wm(t) is

2ph h2
Wm ( f ) = exp −2p2 2 ( f − fc )2 . (23)
fc fc
Because the bandwidth of a Morlet wavelet increases with its center frequency, the Morlet
wavelets in a filter bank are spaced closely at low frequencies and progressively farther
apart at higher frequencies. From low frequency fl to high frequency fh, the center
frequencies of a set of N Morlet wavelets are

n−1/N−1
fh
fc (n) = fl , 1 ≤ n ≤ N. (24)
fl
Estimating thin-bed thickness

Given a set of frequency maps, the thickness of a thin bed is estimated by scanning the
maps to find the frequency of the first spectral peak, and then applying equation 6 to
compute thickness as half the period of the tuning frequency. Scanning is done by eye
or by an automatic process. The underlying assumption is that the observed spectra rep-
resent geology and are free of the overprint of the seismic source wavelet. For this to be
sufficiently true, the source wavelet must have zero phase and a broad flat-amplitude spec-
trum. If the wavelet does not have a white spectrum, then the frequency maps are normal-
ized individually so as to have the same average value (Partyka et al., 1999). Map
normalization suffices for routine studies. Alternatively, spectral whitening could be
applied prior to spectral decomposition.
Another approach for estimating thin-bed thickness matches observed spectra with
various model spectra and takes the thickness to be that of the model that matches best.
This method is more effort but more stable than relying on the first peak frequency
(Figure 11). It could be accomplished in the time domain through a suitably tailored super-
vised waveform classification, which underscores the close relation between spectral
decomposition and waveform classification.
Thin-bed thickness estimates
are constrained by the length of
the analysis interval and the spec-
trum of the signal. Consider a
wavelet spectrum bounded by
low frequency fl and high fre-
quency fh. These frequencies
limit the bed thickness tb that can
be measured through spectral
peak searching to the range
1 1
≤ tb ≤ . (25)
2fh 2fl
Spectral attributes
Figure 11. Thin-bed thickness is estimated from the
Spectral decomposition spectral response by observing the frequency of the first
spreads information across many spectral peak and equating it to the tuning frequency. This
frequency maps. It might seem process fails if the data are noisy or the model is
helpful to summarize this infor- inappropriate. A better approach compares overall
mation as maps of total energy, observed and modeled spectra to find a match.
average frequency, spectral skew, and spectral kurtosis (Marfurt and Kirlin, 2001). Total
spectral energy and average spectral frequency are similar to the standard interval attributes
for total energy and weighted average frequency. Nothing is gained by computing them
through spectral decomposition. There are no corresponding time-domain attributes for
spectral skew and spectral kurtosis, but these quantities are difficult to relate to geology
and have limited value. Spectral attributes that have value estimate thin-bed thickness
and the confidence in the underlying thin-bed model.
A more popular approach for summarizing the information in a set of frequency maps
or volumes employs RGB color blending to combine three spectral components represent-
ing low, intermediate, and high frequencies (Balch, 1971; Stark, 2006; Figure 12). Low fre-
quencies are set red, intermediate frequencies are set green, and high frequencies are set
blue. RGB displays are interpreted qualitatively. Red suggests thick features, green
suggests moderately thick features, and blue suggests thin features. White indicates fea-
tures that have equally strong response at all frequencies, such as solitary bright reflections,
and black indicates features that have weak amplitudes at all frequencies, such as faults and
Figure 12. Red-green-blue blending facilitates interpretation of volume spectral

decomposition by combining low-, intermediate-, and high-frequency components into a
single composite volume.
weak reflections. In practice, RGB blending of frequency volumes is applied more for
revealing detail than for tuning analysis.
Example of spectral decomposition

Figures 13 and 14 illustrate spectral decomposition applied to seismic data that images
a large channel. The channel is characterized by a series of weak reflections with a bright
thin-bed reflection at its base. The channel has a strong positive reflection over a negative
reflection. Figure 13 shows root-mean-square (rms) amplitude extracted in a horizon-
guided 40-ms interval centered on the bright thin bed. Figure 14 compares a set of fre-
quency maps derived along the same horizon, with a corresponding map of reflection
strength to serve as the full spectral response. The frequency maps employ Morlet wavelets.
The channel is imaged best on the 51-Hz frequency map. By equation 6, its time thickness
is estimated as 10 ms.
In thin-bed analysis, the quality of the horizon is as important as the quality of the
seismic data. Here, the seismic data have good quality, but it is challenging to track hor-
izons close to these channels because the reflections are weak and discontinuous.
Instead, the horizon is picked on a clearer, shallower event and shifted down to the level
of the bright channel. The shifted horizon largely passes through the center of the thin
bed, as required for the analysis, but the match is imperfect, which degrades the analysis.
This is a common problem.
Figure 13. A seismic line and an rms amplitude map with a broad channel. The channel is
characterized by a thick sequence of low-amplitude events with a bright thin bed at its base.
The same horizon is employed in Chapter 3 to illustrate amplitude attributes. Data from
the Taranaki Basin, offshore New Zealand.
Figure 14. Frequency maps extracted along the horizon used in Figure 13. The maps are derived
with Morlet wavelets for center frequencies as indicated. The seismic data have not been whitened,
and the frequency maps are scaled individually. The broad channel is imaged best on the 51-Hz map,
as marked by the white circle. Reflection strength is shown for reference and represents the full
spectral response.
Waveform classification
Waveform classification is a method of automatic pattern recognition that identifies
regions of similar waveform along a horizon (Addy, 1997). To the extent that similar wave-
forms represent similar geology, waveform maps reveal stratigraphic features, such as
channels, flood plains, and point bars. Waveform maps often reveal features that cannot
be distinguished on maps of competing attributes.
A seismic waveform is a small segment of a seismic trace that comprises one or more
lobes. A waveform represents a single reflection or a pattern of interfering reflections, and
has characteristic amplitude, frequency, and phase. Individual waveforms lack inherent
geological meaning. Modeling lends them meaning by relating them to known geology
(Poupon et al., 1999). Unfortunately, realistic models entail so many variables that
modeled waveforms are highly nonunique and have little value. As a result, waveform
maps are interpreted qualitatively.
Waveform classification assigns a class number to each waveform in an interval. Inter-
vals follow a horizon and are usually of constant length because a nonconstant length com-
plicates comparison. Each class is characterized by a representative or template waveform.
The number of classes typically ranges from 5 to 25. Waveform maps are thus quantized, so
edges and regions appear sharp and clear, unlike the fuzzy patterns observed on amplitude
or frequency maps. A waveform class is often described as a seismic facies, but its signifi-
cance is more geophysical than geological.
Classification methods
Waveform classification is supervised or unsupervised. Supervised classification
requires that the geophysicist supply the template waveforms. Waveforms are typically
taken at well locations where the geology is known, and in this way they acquire geological
meaning. Supervised classification has great potential, but it is seldom applied because it
demands prior knowledge and considerable effort.
Unsupervised classification automatically determines the template waveforms. Two
methods are common, K-means clustering and the Kohonen self-organizing feature map,
or Kohonen SOFM. Both methods follow a similar workflow: initialization to make a
first guess of the template waveforms, training in a small subset of the data to refine
these waveforms, and classification of the full data set with the final template waveforms.
The essential difference between K-means clustering and the Kohonen SOFM lies in
the ordering of their template waveforms. The Kohonen SOFM naturally orders the wave-
forms according to shape, whereas K-means clustering produces a random order. Taken
individually, class numbers assigned to the template waveforms have no inherent
meaning, but they gain useful relative meaning when the template waveforms are
ordered according to their similarity with each other. This makes the Kohonen SOFM
superior to K-means clustering. Template waveforms found by K-means clustering can
be reordered by waveform amplitude, but reordering by similarity is much more difficult,
being akin to the “traveling salesman problem.” The traveling salesman problem is to find
the shortest path for a salesman to visit
a set of cities, starting and ending at the
same city, but passing through each city
only once. This famous optimization
problem is surprisingly challenging.
Waveform classification is some-
times overly influenced by amplitude.
This is remedied by rescaling all obser-
ved waveforms to have the same rms
amplitude.
Once the template waveforms have
been defined, supervised and unsuper-
vised classification proceed in the
same way. Each waveform observed
Figure 15. Waveform classification compares
in an interval along a horizon is com-
observed waveforms with a set of template
pared to the set of template waveforms waveforms and assigns each observed waveform to
and is assigned the class number of the the class that it most closely matches. In this
template that it most closely matches example, the observed waveform best matches the
(Figure 15). This produces a map of first template waveform, so it is assigned class
class numbers. number 1.
This is the essence of waveform classification. Details of K-means clustering and

the Kohonen SOFM are described in Chapter 9 in the context of pattern recognition for
multiattribute analysis.
Examples of waveform classification

Figure 16 compares two waveform maps and an rms amplitude map derived in
the same 40-ms horizon-guided interval. The waveform maps are generated through unsu-
pervised classification with the Kohonen SOFM. They differ markedly from the rms ampli-
tude map, and show the channel systems to be more complex than would be supposed by
reviewing amplitude alone. In this way, waveform classification complements amplitude
analysis.
The waveform maps illustrate the influence of the number of classes. One map employs
8 classes and the second employs 12. They show the same broad patterns, but details differ
Figure 16. Horizon with amplitude map and Kohonen SOFM waveform maps derived from a
40-ms interval (11 samples) along the horizon. (a) Horizon with shaded relief. (b) Root-mean-square
amplitude. (c) Waveform map with 8 classes. (d) Waveform map with 12 classes.
Figure 17. Template waveforms for the waveform maps of Figure 16. (a) Waveforms for
Figure 16c. (b) Waveforms for Figure 16d.
considerably. The difference is due in part to the difference in number of classes, and in part
to differences in how the template waveforms are ordered, as seen in Figure 17. In some
places, boundaries do not change; in other places, boundaries shift greatly. Increasing
the number of classes increases the resolution of the waveform map but diminishes its
ability to generalize the data. The “right” number of classes depends on the objective
and preference.
Some implementations of waveform classification impose constraints on the template
waveforms to ensure consistent ordering when the number of classes is changed. This
facilitates comparison between maps that otherwise have the same parameterization.
Most implementations ignore this issue, so the order of the template waveforms can
change with the number of classes.
Shortcomings
Unsupervised waveform maps suffer several unique shortcomings. The sharp bound-
aries seen in waveform maps are pleasing to the eye and help visualize patterns.
However, their precise locations cannot always be trusted because they often shift with
changes in parameterization, particularly in the number of classes. Rarely is there a
natural or optimum number of classes in seismic data. Waveforms observed along a
horizon almost always form a continuum of shapes. Where waveforms change gradually,
classification boundaries are artificial and shift depending on the number of classes. Where
waveforms change abruptly, boundaries are meaningful and are less affected by changes in
the number of classes.
Unlike most map attributes, waveform classification is sensitive to small errors in the
guide horizon and to small changes in the interval definition. More critically, waveform
maps are rarely comparable because they use different template waveforms. Supervised
classification does not suffer this shortcoming.
Summary
Spectral decomposition and waveform classification are sophisticated techniques for
producing and analyzing sets of horizon-guided interval maps. Though they differ substan-
tially, they have much in common because both analyze waveforms to reveal stratigraphic
features. Spectral decomposition can even be viewed as a kind of supervised waveform
matching.
Spectral decomposition applies filter banks and tuning analysis to distinguish thin beds
and estimate their thicknesses. Filter banks typically employ a set of Gabor or Morlet wave-
lets characterized by their center frequencies and bandwidths. The bandwidths limit the res-
olution of the analysis in the frequency domain. By the uncertainty principle, bandwidth
can be made smaller only by increasing the length of the wavelet, which degrades resol-
ution in the time domain. Many thin-bed models could be applied, but the most useful
model is the familiar low-impedance channel with two equal magnitude, opposite polarity
reflections. This model retains wide applicability, though it ignores complicating factors.
Assuming this model, and assuming that the analysis interval is centered on the thin
bed, thin-bed thickness is estimated as half the period of the first peak spectral frequency,
where the peak frequency is found qualitatively by inspection of the frequency maps.
Thickness estimates are necessarily rough. The Widess limit of resolution is a useful
estimate of the thinnest thin bed whose thickness can be measured. It is defined as one
eighth the dominant wavelength, or one quarter the period of the dominant frequency.
This constrains thickness estimates derived through spectral decomposition.
Waveform classification applies automatic pattern recognition to identify regions of
similar waveform along a horizon. Because waveforms rarely have inherent geological
meaning, waveform maps are interpreted qualitatively. They provide crisp images that
reveal channels, point bars, flood plains, and other features with greater clarity than
most competing attributes. However, boundaries on waveform maps are somewhat artifi-
cial because they often depend on the number of classes. Further, waveform maps are
overly sensitive to the details of the horizon that defines the interval. Small flaws in the
horizon cause spurious results, and modest changes in interval definition produce strikingly
different maps.
Spectral decomposition is sometimes applied to an entire seismic data volume to
produce a set of frequency volumes. This approach has the advantage that it does not
need interpreted horizons and the disadvantage that it is challenging to analyze multiple
frequency volumes together. For qualitative interpretation, RGB blending of three fre-
quency volumes offers a practical solution. Waveform classification cannot be adapted
to classify entire volumes in some meaningful way. Instead, 3D seismic data are classified
through pattern recognition with multiattribute analysis, as reviewed in Chapter 9.
The next chapter introduces the two poststack geophysical attributes that purport to
record rock properties, relative acoustic impedance and Q.
Chapter 8
Relative Acoustic Impedance and Q
Introduction
Several rock properties are extracted from prestack seismic data with reasonable suc-
cess, principally compressional- and shear-wave velocities and impedances. Density and
quality factor, or Q, are also extracted from prestack data, though less reliably. In contrast,
no rock properties can be extracted with confidence from poststack seismic data. Optimistic
geophysicists nonetheless routinely apply poststack attributes that purport to record two
fundamental rock properties, relative acoustic impedance and Q.
Ideally, relative acoustic impedance records differences in acoustic impedance from a
background trend, and Q attributes record the degree to which the earth attenuates seismic
energy. Relative acoustic impedance is employed primarily as a relative measure of poros-
ity, and Q as a gas indicator.
Offered as poststack seismic attributes, relative acoustic impedance is computed
through recursive inversion, and Q through spectral ratios. Unlike most poststack seismic
attributes, relative acoustic impedance and Q require that the seismic data meet specific
conditions. Acoustic impedance requires that the seismic wavelet be removed from the
data so that the reflection coefficients can be observed. The wavelet is assumed to be inva-
riant. Q requires that the seismic wavelet be preserved so that it can be observed. The
wavelet is assumed to vary in time and space. These requirements conflict and are never
satisfied for either purpose. As a result, poststack relative acoustic impedance and Q attri-
butes are far from ideal. They have some qualitative value, but quantitatively they are
highly unreliable. In spite of their defects, relative acoustic impedance attributes are
widely employed, and Q attributes are widely promoted.
This chapter introduces the ideas that underlie relative acoustic impedance derived
through recursive inversion and Q derived through spectral ratios, and reviews their appli-
cation and their flaws.
Relative acoustic impedance

Acoustic impedance Z is a rock property that quantifies the resistance offered to pro-
pagating seismic compressional waves. It is the product of two other rock properties,
density and compressional-wave velocity. In clastic sediments, acoustic impedance tends
to be related inversely to porosity because increased porosity reduces rock density and
usually reduces compressional-wave velocity (Bacon et al., 2007, p. 147). If the pore
spaces are filled with gas in place of water or oil, then density reduces further and compres-
sional-wave velocity decreases sharply, causing a marked drop in acoustic impedance.
133
The velocity decreases dramatically for low gas saturations up to about 10%, and
then increases slowly for higher gas saturations. As a result, compressional-wave velocity
and acoustic impedance serve as direct hydrocarbon indicators, but they cannot distinguish
economic gas saturations from “fizz gas” or noneconomic saturations (Liner, 2004, p. 511;
Bacon et al., 2007, p. 140). Acoustic impedance is of interest chiefly for inferring porosity.
Impedance logs record acoustic impedance in wells. To determine acoustic impedance
between wells, seismic data are converted to acoustic impedance through inversion.
Seismic data lack both the low and the high frequencies of impedance logs, so a full inver-
sion must supply this missing content by incorporating velocity and density information.
The high frequencies record stratigraphic detail and are not recoverable between wells.
The low frequencies define the background trend in the impedance, and they are essential
for estimating absolute porosities. It is reasonably straightforward to estimate the back-
ground trend given a velocity model and assumptions about density. Without the back-
ground trend, acoustic impedance becomes “relative acoustic impedance.” Relative
acoustic impedance is the only seismic impedance attribute that can be estimated from con-
ventional poststack seismic data alone. All other impedance attributes require velocities,
well logs, multicomponent seismic data, or prestack seismic data.
Most commercial software packages for poststack seismic attributes include relative
acoustic impedance derived through recursive inversion. Recursive inversion acts like an
integration of the seismic trace, which rotates the
phase of the trace by 90 degrees and boosts low frequen-
cies with respect to high frequencies. It is more a trans-
formation than an attribute because it does not subset the
information in the data. Indeed, the original seismic data
can be recovered from a recursive inversion through
differentiation, with a slight error.
Recursive inversion requires that the seismic data
closely approximate a normal-incidence reflection coef-
ficient series. In practice, this means that the data must
have zero-phase and broad bandwidth and be largely
free of noise. To enhance the bandwidth, spectral broad-
ening is often included in a workflow for recursive
inversion (Figure 1).
Recursive inversion
Acoustic impedance inversion transforms seismic
data into estimates of acoustic impedance. There are
many ways to do this; recursive inversion is the oldest,
simplest, and least reliable method (Latimer et al.,
Figure 1. Workflow to 2000). Recursive inversion has long since fallen into
compute relative acoustic disrepute for full acoustic impedance inversion, but it
impedance through recursive remains widely employed for poststack relative acoustic
inversion. impedance attributes. Recursive inversion is founded
Chapter 8: Relative Acoustic Impedance and Q 135
on the assumption that a seismic trace approximates

a reflection coefficient series, from which the associ-
ated impedances derive recursively. The following

development employs an earth model with flat homo-
geneous layers and assumes normal-incidence reflec-
tions (Waters, 1981, p. 271; Russell, 1988; Lines and
Newrick, 2004, p. 151).
Consider an earth model that has two layers with
acoustic impedances Z1 and Z2 (Figure 2). The reflec-
tion coefficient r1 of the interface between the layers
is the relative difference,
Z2 − Z1 Figure 2. Normal-incidence
r1 = . (1) reflection from an interface between
Z1 + Z2 two layers with impedances Z1 and
Turning this around, the impedance of the second Z2. Reflection coefficient r1 is a
layer Z2 can be determined from the impedance of function of the impedances.
the first layer Z1 and the reflection coefficient r1:
1 + r1
Z2 = Z1 . (2)
1 − r1
The impedance of the first layer must be known or
estimated.
An earth model with 3 layers has 2 reflecting
interfaces (Figure 3). The impedance of the second
layer, Z2, is found from equation 2, and the impe-
dance of the third layer, Z3, is found similarly from
Z2 and the second reflection coefficient, r2. Combin- Figure 3. A layered earth model
ing these two steps yields the recursive equation with 3 layers and 2 interfaces. Here,
for Z3, Z1 , Z2 . Z3, so r1 is positive and
r2 is negative.
1 + r2 1 + r1 1 + r2
Z3 = Z2 = Z1 · . (3)
1 − r2 1 − r1 1 − r2
Extending this recursive procedure to an earth model with n layers (Figure 4), the
impedance of the nth layer Zn is determined from the set of n – 1 reflection coefficients,
r1, r2 , . . . , rn – 1, according to
1 + r1 1 + r2 1 + rn−1
n−1
1 + rj
Zn = Z1 · ··· = Z1 . (4)
1 − r1 1 − r2 1 − rn−1 j=1
1 − rj
Taking the logarithm of the multiplicative series produces the additive series,

n−1
1 + rj
Zn = Z1 exp ln . (5)
j=1
1 − rj
For reflection coefficients rj much smaller

than 1,

1 + rj
ln ≈ 2rj . (6)
1 − rj
Employing this approximation in equation 5
yields the inversion formula

n−1
Zn ≈ Z1 exp 2 rj . (7)
j=1
The reflection coefficients rj are unknown;

Figure 4. A layered earth model with only seismic trace samples xj are known. The
n layers and n – 1 interfaces. seismic trace is a band-limited version of the
reflection coefficient series plus noise and mul-
tiples. It lacks both high frequencies and the low frequency background trend (Figure 5).
As a result, only a band-limited version of the impedance can be recovered from seis-
mic data alone. The lack of high frequencies limits the resolution; the lack of low frequen-
cies limits the resolution and restricts the inversion to a relative acoustic impedance.
If noise is negligible, then the seismic trace scaled by a factor k approximates the inter-
mediate-frequency portion of the reflection coefficient series. Ignoring the band-limited
nature of seismic data, this relationship is expressed as kxj rj. The approximation is
flawed because index j on the left-hand side refers to a sample in the seismic trace,
whereas on the right-hand side it refers to a reflection coefficient in the reflectivity
series. Accepting the flaw for the sake of expediency, while acknowledging that the
result must be a relative acoustic impedance Z̃ n , the formula for recursive inversion
becomes

n−1
Z̃ n ≈ Z1 exp 2k xj . (8)
j=1
The exponent tends to be small, roughly less than 0.2 in magnitude. In this case, repre-
senting the exponent as a, ea ≈ 1 + a. Thus

n−1
Z̃ n ≈ Z1 1 + 2k xj . (9)
j=1
This is the practical formula for recursive inversion. The impedance of the first layer, Z1,
must be known. However, Z1 merely scales the relative acoustic impedance and conse-
quently does not affect its appearance. Scale factor k relates the value of the seismic
trace at a reflection to the value of the reflection coefficient. Like Z1, k scales the result
but does not change its appearance. The values assigned to Z1 and k are unimportant in
qualitative studies.
Figure 5. The spectrum of the earth’s reflectivity divided into three frequency ranges:
low, intermediate, and high. Low frequencies are provided by the background trend, intermediate
frequencies are provided by the seismic data, and high frequencies are provided by well logs.
Low-cut filter
Equation 9 approximates an integration of the seismic trace. The spectrum I( f ) of
the integration operator is
i
I( f ) = − (10)
2pf
(Figure 6a). Integration suppresses high frequencies, strongly boosts low frequencies, and
has an infinite dc component. Exploration seismic data are dominated by noise at low fre-
quencies. Integration boosts this low-frequency noise so much that it obscures the signal.
The noise must be removed by a low-cut filter.
“Leaky integration” offers an alternative to low-cut filtering because it limits how
strongly the dc and low-frequency components are boosted (Claerbout, 1992, p. 48). The
integrated trace yn of a discrete seismic trace xn is found through the recursive relation,

n
yn = xj = yn−1 + xn . (11)
j=1
The leaky integration ỹn of trace xn is the similar relation,
ỹn = cỹn−1 + xn , (12)
where factor c is slightly less than 1. The leaky integration operator has the frequency
spectrum Ĩ( f ) given by
1
Ĩ( f ) = . (13)
1 − cei2pfT
Figure 6b illustrates the amplitude spectrum of the leaky integrator. Low frequencies
are boosted strongly with respect to higher frequencies, but the increase is bounded and
controlled by the factor c.
Influence of wavelets and noise

Recursive inversion ignores the

influence of the seismic wavelet by
assuming that the seismic trace
represents a reflectivity series. The
assumption is unjustified and leads
to significant error. This is illustrated
in the simple example of Figure 7,
which compares relative acoustic
impedance derived from a reflection
coefficient series with that derived
from a synthetic seismic trace. The
synthetic trace is created by convol-
ving the reflection coefficients with
azero-meansymmetricsourcewave-
let. The impedances are derived
from equation 9 with Z1 ¼ 5 and
k ¼ 0.1. The impedance derived
directly from the reflection coeffi-
cients is drawn as a blocky function.
It approximates the true impedance
to the extent that the reflection coef-
ficients are small. The impedance
derived from the synthetic trace is
Figure 6. Amplitude spectra scaled by 200p for drawn as a seismic trace. The two
(a) standard integration, and (b) leaky integration impedance functions differ greatly,
with c ¼ 0.95. illustrating the deleterious effect of
the seismic wavelet on inversion.
Restoring the background trend

The missing background trend in relative acoustic impedance can be restored by
incorporating velocity models or well logs in the inversion. A common approach estimates
the background trend by converting interval velocities to acoustic impedances using
Gardner’s empirical relation between density and velocity in clastic sediments, and then
adding this trend to the relative acoustic impedance to produce a full acoustic impedance.
The result is no longer strictly a poststack seismic attribute because it includes information
from well logs or velocities from prestack seismic data.
Examples of relative acoustic impedance

Figure 8 compares a seismic line, processed with and without spectral whitening, to
its relative acoustic impedance estimated through recursive inversion; Figure 9 gives the
Figure 7. Illustration of the key limitation of relative acoustic impedance estimated through
recursive inversion. (a) A zero-mean symmetric source wavelet. (b) Reflection coefficients
(black spikes) and the seismic trace (blue) created by convolving the wavelet with the reflection
coefficients. (c) Relative acoustic impedance derived through recursive inversion of the seismic
trace (green), and acoustic impedance derived through recursive inversion of the reflection
coefficients and represented as a blocky function that approximates the true impedance (red).
Scale factor k ¼ 0.1, and initial impedance Z1 is 5; both curves neglect the initial impedance.
Relative acoustic impedance indicates changes in impedance. However, interpretation is
not straightforward, even in this simple noise-free model.
Figure 8. Examples of relative acoustic impedance estimated through recursive inversion.

(a) Original seismic data. (b) Seismic data with spectral whitening. (c) Relative acoustic
impedance derived from the unwhitened seismic data. (d) Relative acoustic impedance derived
from the whitened seismic data. The relative acoustic impedances are filtered with a 2-Hz
low-cut filter.
amplitude spectra. Both the original seismic data and the spectrally whitened data are fol-
lowed by coherency filtering to remove random noise. The whitened data have higher
frequency content and show better detail. Consequently, relative acoustic impedance
derived from the whitened data has higher frequencies than that derived from the original
data, but the enhanced low frequencies are more striking and just as important. Sometimes
relative acoustic impedance looks cleaner than the original seismic data because it boosts
low frequencies, which are usually more coherent than high frequencies.
Relative acoustic impedance derived through recursive inversion is widely applied
because it is easy to generate and often qualitatively helpful. However, it remains a poor
man’s impedance attribute and must be interpreted cautiously (Russell and Hampson,
2006). Its underlying assumptions are invalid, and it is limited by the lack of both low and
high frequencies. As with any impedance inversion, the validity of relative acoustic
Figure 9. Spectra of the seismic data of Figure 8. (a) Original seismic data. (b) Seismic
data with spectral whitening. (c) Relative acoustic impedance of the unwhitened seismic data.
(d) Relative acoustic impedance of the whitened seismic data.
impedance is checked by comparing with impedance logs, or with synthetic traces derived
from logs (Latimer et al., 2000).
Attenuation
Seismic attenuation is the progressive loss of energy in a seismic wave propagating
through a lossy medium. This loss is proportionally greater for higher frequencies than
for lower frequencies, and is in addition to the frequency-independent losses due to wave-
front spreading. Attenuation is the combination of “intrinsic attenuation,” caused by
absorption, and “apparent attenuation,” caused by scattering. Absorption is a fundamental
rock property that produces friction of the seismic wave as it passes through the rock. Scat-
tering is the cumulative effect of multiple reflections and diffractions of the seismic energy
at rock interfaces. The effects of intrinsic and apparent attenuation are hard to distinguish,
so typically their combined effect is treated as one. Attenuation is quantified by the quality
factor or Q of the rock, which is an inverse measure of a rock’s tendency to absorb seis-
mic energy (Table 1). Rocks with low Q strongly attenuate energy, but rocks with high
Q attenuate weakly or negligibly.
Table 1. Typical quality factors measured in

the laboratory for common rocks and fluids (Hamilton,
1972; Johnston, 1981; Waters, 1981, p. 29; Sheriff and

Geldart, 1995, p. 180). Quality factors estimated from
seismic data are generally lower.
Rock or fluid Quality factor
Shale 20– 80
Marine sediments 30– 100
Sandstone 70– 130
Limestone, dolomite 60– 200
Granite 100– 500
Gas sand 5 – 50
Seawater 63,000
Q is of interest because gas sands tend to have anomalously low Q. Thus Q is a hydro-
carbon indicator. Unfortunately, it is unreliable because a small percentage of gas mixed
with fluid causes Q to decrease sharply (Toksöz and Johnston, 1981). As a result, Q is
like acoustic impedance in that it cannot distinguish between economic and noneconomic
gas saturations.
Valid estimates of Q can be derived from poststack seismic data if the noise is low,
the data have not been processed with time-variant filtering or time-variant deconvolution,
moveout stretch and migration have negligible influence on frequency content, and, most
importantly, changes in the spectrum of the seismic wavelet can be estimated at different
traveltimes on seismic traces. This last condition is especially problematic. Seismic traces
are complicated functions of earth reflectivity, attenuation, source wavelets, noise, and
other factors, making it difficult to observe the seismic wavelet at different times or to dis-
tinguish the small background effects of attenuation from other, more important effects.
Attenuation sometimes manifests itself on seismic data as a low-frequency shadow
directly beneath a highly attenuating zone. Shadows fade away with distance from the
attenuating zone due to wavefront healing.
Standard frequency attributes reveal attenuation, but they are too strongly influenced
by local reflectivity to be useful in estimating Q. The pulse broadening method, which
measures the time from the onset of a wavelet to its first maximum, likewise cannot be
applied because it requires clean seismic wavelets free of interference (Stacey, 1977,
p. 304; Liu, 1988). Other methods are used instead, the most important being spectral ratio-
ing, which is based on differences in the amplitude spectra of consecutive time windows
on a seismic trace (Johnston and Toksöz, 1981, p. 9; Tonn, 1991; Mitchell et al., 1996).
Quality factor
Over the frequency range of seismic exploration data, a sinusoidal plane wave traveling
in a lossy medium suffers amplitude decay proportional to the number of wavelengths
traveled. This gives rise to frequency-dependent exponential decay in wave amplitude with
traveltime. The instantaneous power 1(t, f ) of the wave with frequency f propagating
through an attenuating medium with quality factor Q is
1(t, f ) = 1(0, f )e−2pft/Q , (14)
where 1(0, f ) is the power of the wave at time t ¼ 0 (Stacy, 1977, p. 301). This equation is
the solution of the differential equation that defines frequency-independent Q,
2pf 1(t, f )
Q=− (15)
d
1(t, f )
dt
(Johnston and Toksöz, 1981, p. 2). The expression for 1(t, f ) holds for any frequency or
combination of frequencies, so it also represents the instantaneous power spectrum of a
broadband propagating plane wave, and not just the power of a single frequency component
(Figure 10).
Attenuation of a propagating seismic wave is manifested by decay in its average spec-
tral frequency as well as in its energy. Attenuation also causes dispersion, which distorts
the phase spectrum. Dispersion is usually insignificant on seismic reflection data and is
neglected in most analyses of attenuation.
Attenuation in a homogeneous earth

The decay in energy and average frequency suffered by a seismic plane wave propa-
gating through a homogeneous attenuating earth is derived by putting the expression
Figure 10. Power spectrum after 1-s traveltime in a homogeneous earth with a Q of 100 for a
seismic plane wave with an initial boxcar power spectrum of value 1 from 10 to 80 Hz.
for the instantaneous power spectrum of equation 14 into equations for the energy and
average spectral frequency of a wavelet. For the purpose of illustration, let the initial
power spectrum of the plane wave, 1(0, f ), be a boxcar with bandwidth from low frequency
fl to high frequency fh:

1 fl ≤ f ≤ fh
1(0, f ) = . (16)
0 f , fl or f . fh
The energy of the plane wave as a function of time, E(t), is
1 fh
E(t) = 1(t, f )df = exp (−2pft/Q)df , (17)
0
fl
which yields
Q
E(t) = [exp (−2p fl t/Q) − exp (−2p fh t/Q)]. (18)
2pt
Similarly, for average frequency,
1
f 1(t, f )df fh
1
fa (t) = 1
0
= f exp (−2pft/Q)df , (19)
E(t)
1(t, f ) fl
0
which yields
Q fh exp (−2p fh t/Q) − fl exp (−2p fl t/Q)
fa (t) = + . (20)
2pt exp (−2p fh t/Q) − exp (−2p fl t/Q)
Figure 11 shows energy and frequency decay curves derived from equations 18 and 20 for
a range of typical values of Q. These curves illustrate the overall pattern of spectral decay
that one might observe on seismic reflection data. The energy of the seismic wave decays
more significantly than its average frequency.
Attenuation due to a buried channel

Consider now the spectral change caused by a highly attenuating gas sand. Model the
gas sand as a channel sand with low Q embedded in shale with high Q (Figure 12). Neglect
the energy lost to reflections. Above the channel, the instantaneous power spectrum 1(t, f )
of a propagating seismic plane wave is
1(t, f ) = 1(0, f )e−2pft/Q1 t ≤ t1 . (21)
Within the channel, the power spectrum becomes
1(t, f ) = 1(t1 , f )e−2pf (t−t1 )/Q2 t1 , t ≤ t2 , (22)

a)
100
75
Energy 50
200
25 100
50
20
0
0 500 1000 1500 2000
Time (ms)
b)
50
40
Frequency (Hz)
200
30
100
50
20
20
10
0 500 1000 1500 2000
Time (ms)
Figure 11. (a) Energy and (b) average frequency decay curves for a plane wave with initial boxcar
spectrum from 10 to 80 Hz, for rock quality factors of 20, 50, 100, and 200.
Figure 12. Model of a low-Q sand body embedded in high-Q shale. The sand thickness D
corresponds to the time difference t2 – t1.
and beneath the channel,
1(t, f ) = 1(t2 , f )e−2pf (t−t2 )/Q1

t . t2 , (23)
where 1(t1, f ) and 1(t2, f ) are the power spectra of the plane wave at times t1 and t2. Employ-
ing this definition of the instantaneous power spectrum in the equations for total energy
and average frequency above produces the decay curves shown in Figure 13. Sharp
drops in energy and average frequency occur beneath the channel. The drop in frequency
Figure 13. (a) Energy and (b) average frequency decay curves derived from the channel
model of Figure 12 for a plane wave with an initial boxcar spectrum from 10 to 80 Hz.
The channel top is at 1.0 s and its thickness is 50 ms.
is the low-frequency shadow, which here has a magnitude of about 5 Hz. Such a small
change in overall frequency is difficult to detect on seismic data, given the much larger
and more rapid variations due to reflection interference and other factors. To increase
the chances that shadows can be detected and quantified, long analysis windows are
employed. This analysis ignores wavefront healing, which limits the length of shadows
observed on seismic data.
Figure 14 shows typical observed amplitude and frequency decay curves derived from
fully processed poststack seismic data. The seismic data are of good quality and have
Figure 14. Observed (a) amplitude and (b) frequency decay curves for seismic data with a
bandwidth of 10 to 80 Hz. These curves represent the average of 3111 attribute traces; the
amplitude attribute is reflection strength and the frequency attribute is weighted average
instantaneous frequency derived in a 30-ms (15-sample) analysis window. The two
strong amplitudes between 820 and 900 ms are reflections from gas sands.
an overall bandwidth of 10 to 80 Hz. The curves represent the average of 3111 adjacent
attribute traces through a zone that encompasses two bright reflections from gas reser-
voirs between 820 and 900 ms. The amplitude attribute is reflection strength, and the fre-
quency attribute is weighted average instantaneous frequency derived in a 30-ms analysis
window. In spite of the considerable time and spatial averaging, frequency fluctuations
exceed 10 Hz beneath the gas zone. It is practically impossible to identify frequency
changes that can be attributed to particular reflections with any confidence.
The seismic data have had spherical divergence corrections applied in processing,
removing any overall decay in amplitude, but these do not affect amplitude changes
caused by locally strong attenuation. The frequency curve increases rapidly from time
0 ms to 500 ms, after which it decays more or less as expected. This pattern is typi-
cal. It results from the highly variable offset mix and strong normal moveout (NMO)
stretch that occur in the shallow mute zone. For this reason, attenuation analysis fails in
shallow data.
Spectral ratio method for Q estimation

The spectral ratio method is designed for measuring the quality factor Q of rock
samples in a laboratory. Adapted for seismic reflection data, it assumes that the frequency
spectrum of a window of seismic data represents the spectrum of the seismic source wavelet
and is not unduly influenced by reflectivity or noise. The same assumption underlies the
method of spiking deconvolution. This assumption is often invalid in deconvolution and
is even less valid in spectral ratioing because spectral analysis windows are typically too
short to justify the implicit requirement of white reflectivity. An analysis window that is
long enough to provide valid estimates of Q is almost guaranteed to have unacceptably
poor time resolution. The necessary compromise between resolution and validity is
rarely satisfactory.
The spectral ratio method for Q estimation determines amplitude spectra in successive
windows of a seismic trace and separates frequency-dependent spectral changes from
frequency-independent changes. If the measured spectra record only the seismic wavelet,
then the frequency-dependent changes are related to attenuation and the frequency-
independent changes are related to wavefront spreading. Q estimation by spectral ratioing
relies only on the frequency-dependent changes and therefore is insensitive to automatic
gain control or spherical divergence corrections. This insensitivity is a distinct advantage.
The instantaneous power spectrum 1(t, f ) introduced above is a frequency-dependent
spectrum valid for plane waves. A more general instantaneous power spectrum includes
frequency-independent amplitude changes caused by wavefront spreading or data pro-
cessing. Let g(t) represent the frequency-independent instantaneous power. Multiplying
g(t) with 1(t, f ) gives an instantaneous power spectrum 1g(t, f ) that incorporates both
frequency-dependent and frequency-independent effects:
1g (t, f ) = g(t)1(0, f )e−2pft/Q . (24)
Assuming constant Q, the power spectra at a time t1 and a later time t2 are
1g (t1 , f ) = g(t1 )1(0, f )e−2pf t1 /Q , (25)
and
1g (t2 , f ) = g(t2 )1(0, f )e−2pf t2 /Q . (26)
The logarithm of the ratio of these spectra is the spectral ratio equation,

1(t1 , f ) 2pf (t2 − t1 ) g(t1 )
ln = + ln . (27)
1(t2 , f ) Q g(t2 )
This equation defines a straight line as a function of frequency. Express it concisely as
C( f ) = mf + b, (28)
where function C( f ) is the logarithm of the spectral ratio,

1(t1 , f )
C( f ) = ln , (29)
1(t2 , f )
slope m is
p(t2 − t1 )
m= , (30)
Q
and intercept b is

g(t1 )
b = ln . (31)
g(t2 )
Intercept b is a measure of frequency-independent amplitude decay. It is unrelated to Q
and is ignored. Q is found from slope m as
p(t2 − t1 )
Q= . (32)
m
Figure 15 illustrates the idea behind spectral ratioing.
All that is left to do is to estimate slope m from observed seismic spectral ratios. A plot
of the logarithm of spectral ratios measured at different frequencies approximates a straight
line with positive slope if the spectral changes are due largely to attenuation (Figure 16).
Given a set of N spectral ratios Ci at frequencies fi, where i ¼ 1, 2, . . . N, and fitting a
Figure 15. Illustration of the spectral ratio method. (a) Power spectrum of the seismic data at
time t1. (b) Power spectrum at a later time t2, with relative decay. (c) Logarithm of the ratio
of the two power spectra. The slope of the log-ratio is proportional to inverse quality factor Q 21.
straight line to these points by least squares, slope m is estimated as

N Ni=1 fi Ci − Ni=1 fi · Ni=1 Ci

m= N 2 . (33)
N Ni=1 fi2 − i=1 fi
Slope m is often small, so inverse Q may be preferable to Q because it is more stable.
Implementation
The spectral ratio at a time tm is a function
of the spectra of the seismic data in windows
before and after this time. Figure 17 illustrates
a window strategy. This strategy works well if
the wavelet tends to be fairly constant in each
window, and any anomalous attenuation occurs
in a zone that is narrow with respect to the
window length.
Spectral ratioing is largely unaffected by
processes that do not introduce frequency-
Figure 16. A straight line is fit to a set of dependent spectral change, such as automa-
observed spectral ratios C( f ) in the tic gain control and spherical divergence cor-
bandwidth of the seismic signal. The slope rection. Methods of spectral shaping, such as
m of the line is positive if the spectral spiking deconvolution or spectral whitening,
changes are due to attenuation. Quality also do not affect spectral ratios if they
factor Q is related to the slope. Intercept employ a single operator for the entire trace,
b is a measure of frequency-independent and thus do not introduce time-variant spectral
amplitude change and is not needed in change. The fortunate result is that spectral
attenuation analysis. ratioing is insensitive to the effects of many
Figure 17. Windowing strategy to compute spectral ratios on a seismic trace in time. The
spectral ratio measured at time tm is derived from the spectra in an upper window centered at
time t1 and in a lower window centered at time t2. The time difference between t1 and t2 is Dt, which is
the effective window length and sets the resolution of the method. Window amplitude is W(t).
methods of seismic data processing. Like most attribute techniques, it works best with clean
broad-bandwidth data and fails for noisy narrow-bandwidth data. As a rough rule, spectral
ratioing requires a signal bandwidth of two octaves or more.

Because spectral ratioing employs long analysis windows on individual seismic traces,
Q attributes look streaky and lack lateral continuity. Their appearance is improved mod-
estly by trace mix.
Examples of spectral ratioing

Figure 18 shows an example of the spectral ratio method to estimate inverse Q.
The data seem promising for Q analysis because they exhibit marked frequency changes
laterally and vertically. Anomalies of interest are those with small positive Q and
hence large inverse Q, which indicate places where the frequency content decreases
markedly in time. In this typical example, inverse Q is negative almost as much as
positive. Negative zones indicate an increase in frequency content with time, demon-
strating the reality that within a typical analysis window the earth’s reflectivity has
more influence than changes in the seismic wavelet. As a result, Q has little quantitative
Figure 18. A seismic line and inverse Q derived through spectral ratioing with different window
lengths. (a) Original seismic data. (b) Inverse Q attribute using a 50-ms window, (c) a 100-ms
window, and (d) a 200-ms window. Larger windows improve Q estimation but reduce time
resolution.
significance, though it retains qualitative value for anomaly identification. Here, negative
inverse Q is associated with clean, closely spaced reflections. A strong reflection at
1000 ms on the right side is preceded by a zone of low Q (high inverse Q) and is followed
by a zone of negative Q. This behavior is contrary to that expected for an attenuating
gas sand.
The strong influence of reflectivity makes attenuation analysis prone to misinter-
pretation. For example, changes in the overall strength of the reflectivity alters the
signal-to-noise ratio of the data, which likely alters the observed spectra. As a result, a
zone of weak reflectivity beneath a zone of strong reflectivity could appear as a strong
attenuation anomaly and yet be unrelated to attenuation. More discouragingly, reflection
interference can mask genuine zones of strong signal attenuation. Q attributes derived
through spectral ratioing are unreliable and must be interpreted cautiously.
Summary
Poststack seismic attributes for relative acoustic impedance derived through recursive
inversion and Q derived through spectral ratios purport to reveal rock properties. They
rarely succeed. They retain qualitative value, but quantitatively they are highly unreliable.
At best, relative acoustic impedance suggests relative differences in porosity, and Q
identifies zones of anomalous spectral change. The ideas of impedance and attenuation
are nonetheless of fundamental importance in reflection seismology, so in spite of obvious
shortcomings, relative acoustic impedance is applied widely, and Q attributes remain
popular for research.
Both relative acoustic impedance and Q are limited by the difficulty of separating the
seismic source wavelet from the earth’s reflectivity. For acoustic impedance, the seismic
source wavelet must be removed from the data to observe the reflectivity. The wavelet
is assumed more or less constant throughout the data. In contrast, for Q attributes the reflec-
tivity must be white to observe changes in the wavelet. These requirements conflict and
cannot be fully satisfied. They are often ignored.
Relative acoustic impedance derived through recursive inversion is a poor man’s impe-
dance attribute. It is approximated by an integration of the seismic traces followed by a
low-cut filter. Relative acoustic impedance is more a transform than an attribute because
the original seismic data are recoverable through differentiation. It is popular because
it is easy to generate and sometimes provides useful results. Nonetheless, it must be inter-
preted cautiously. Its implicit assumption that the seismic wavelet has been removed
is never justified, and the missing low-frequency information handicaps quantitative
application.
Improving seismic resolution in time is often cited as an advantage of impedance inver-
sion. However, broad bandwidth is a requirement for successful inversion instead of a
characteristic result. This is certainly true of recursive inversion. By itself, recursive inver-
sion acts like an integration, which boosts low frequencies and thereby often reduces res-
olution. If the goal is solely to improve resolution, then deconvolution or spectral whitening
is the proper solution, not impedance inversion.
Quality factor attributes have been available for many years but have never enjoyed
the respect or popularity of impedance attributes. This is because their defects are clear
and their advantages less so. The chief advantage is that they identify zones of anomalous
spectral change, which are not otherwise notable. The chief defects are unreliable measure-
ments and poor resolution. These result from the difficulty in separating the small, slowly
varying background effects of attenuation from the much larger, rapidly varying effects
of the reflectivity. The spectral ratio method for computing Q produces nearly as many
negative values as positive, proving that it is largely unrelated to signal attenuation. Its
shortcomings are shared by other Q attributes. In spite of this disappointment, new Q attri-
butes appear regularly, accompanied by bold claims. Geophysicists should judge these
skeptically.
Another way to endow seismic attributes with geological meaning is through multiat-
tribute analysis, the topic of the next chapter.

Chapter 9
Multiattribute Analysis
Introduction
Multiattribute analysis is the simultaneous study of two or more attributes to character-
ize seismic data (Figure 1). The motivation is that seismic data are better described by a
set of complementary attributes than by any single attribute. The key methods of multiat-
tribute analysis are volume blending, cross-plotting, principal component analysis, and
automatic pattern recognition.
Volume blending graphically combines two or more seismic attribute volumes into one
display to enable direct comparison. It is popular because it is easy to apply and interpret.
Attribute cross-plotting reveals relationships between attributes. It is used to define ano-
malies or identify duplicate attributes. Principal component analysis finds linear relation-
ships among a set of attributes. It is applied in workflows to reduce the total number of
attributes or to create unique attributes. Automatic pattern recognition finds regions in
seismic data characterized by similar attribute values. It finds application in detailed analy-
sis and geobody extraction.
This chapter reviews the methods of multiattribute analysis.
Volume blending
Volume blending displays two seismic volumes in a composite image to facilitate their
direct comparison (Figure 2). In effect, it overlays a semitransparent foreground attribute
on top of a background attribute so that both are visible. Opacity functions control the trans-
parencies of the attributes. Blended displays are enhanced by varying the opacity of the
foreground attribute to strengthen key values and hide unimportant values (Figure 3).
The design of opacity functions, like that of color bars, remains as much art as science.
The idea of blending evolved from Nigel Anstey’s work in the early 1970s on seismic attri-
butes and over-plotted color displays (see Anstey, 2005). With modern visualization
tools, volume blending has become routine.
An effective strategy for volume blending relates stratigraphic and geophysical infor-
mation to its structural context by displaying a structural attribute in the background in
gray scale and a stratigraphic or geophysical attribute in the foreground in color. Such
blended displays become especially intuitive when the structural attribute looks like illu-
minated apparent topography because apparent topography often suggests true subsurface
geological structure.
155
Figure 1. Multiattribute analysis combines two or more seismic attributes to produce a

composite image, predict geological properties, or create new attributes.
Figure 2. An example of volume blending. (a) Original seismic data. (b) Reflection strength.
(c) Seismic shaded relief. (d) Reflection strength in the foreground blended with shaded relief
in the background. Yellow arrow indicates illumination direction.
Illumination is achieved through bump mapping or directional attributes. It acts as a

directional filter, highlighting features that trend perpendicular to the illumination direction
and hiding features that trend parallel to it.
Bump mapping treats seismic data values like topographic elevations along a surface
and illuminates them to simulate a 3D appearance (Lynch and Lines, 2004; Lynch et al.,
2005; Lynch, 2006). The illumination direction is arbitrary and can be changed in the
display. Bump mapping is appropriate for discontinuity, dip, curvature, and the original
seismic data.
Directional attributes are computed through directional derivatives. They naturally
look illuminated when displayed in monochrome. They achieve much the same effect as
bump mapping except their illumination cannot be changed in the display. To highlight
Chapter 9: Multiattribute Analysis 157
Figure 3. The opacity function of the foreground attribute is varied to enhance the appearance
of a blended display. Here, the seismic data of Figure 2 are blended as the foreground attribute
with seismic shaded relief in the background. (a) Blending with a constant opacity function.
(b) Blending with a variable opacity function. The opacity functions are shown as red lines on
the histograms of the seismic data. Yellow arrow indicates illumination direction.
Table 1. A popular strategy for volume blending, with representative attributes.

Attributes that highlight geological structure are displayed in the background in gray
scale, and attributes that quantify stratigraphic or geophysical properties are displayed
in the foreground in color. Surfaces acquire a 3D texture through standard blending with
directional attributes or through bump mapping with nondirectional attributes.
Background – structural Foreground – stratigraphic or

attributes (gray scale) geophysical attributes (color)
Non-directional Discontinuity Original seismic
Dip magnitude Reflection strength
Bump mapping Curvature Response phase
Fault attribute Average frequency
Directional Azimuth Sweetness
Apparent dip Parallelism
Co-rendering Shaded relief Acoustic impedance
Amplitude change AVO attributes
all trends, directional attributes must be created in pairs with orthogonal illumination.
Directional attributes include apparent dip, azimuth, seismic shaded relief, and amplitude
change.
Table 1 summarizes this blending strategy. It has general applicability and is widely
followed (Chopra, 2002; James and Kostrova, 2005; Pilcher and Blumstein, 2007; Hart,
Table 2. Alternative attribute combinations employed in

volume blending.
Background attribute Foreground attribute

(gray scale) (color)
Discontinuity Curvature
Dip Azimuth
Original seismic data Dip
Original seismic data Sweetness
Seismic shaded relief Discontinuity
2008a, 2008b; Ferguson et al., 2010). A second

strategy aids structural analysis by blending
complementary structural attributes, such as dis-
continuity and curvature (Chopra and Marfurt,
2010; Table 2). A third strategy blends three
related attributes through red-green-blue (RGB)
blending. By this strategy, each attribute controls
a separate color, either red, green, or blue. It is
applied chiefly in spectral decomposition to
combine low-, intermediate-, and high-frequency
tuning volumes, and occasionally in amplitude-
variation-with-offset (AVO) analysis to combine
near-, mid-, and far-angle stacks (Stark, 2006;
see Chapters 7 and 10).
Except in RGB blending, it is difficult to
blend more than two attributes effectively. To
do so, the opacity functions must be designed
to limit overlap of the foreground attributes.
Blended displays are more interpretable when
seismic attributes are smooth and clean. Attri-
butes are made smoother and cleaner either by
coherency filtering the seismic data before attri-
bute computation, or by smoothing the attributes
directly. Figure 4 shows a general workflow for
volume blending.
Figure 4. Workflow for volume

blending. Illumination is introduced Discontinuity, curvature, and
through the structural seismic attribute amplitude change
or through the visualization. It is not
always necessary to filter the seismic Discontinuity is the most widely employed
attributes. structural attribute in volume blending. It makes
Figure 5. Blending helps delineate channels. (a) Reflection strength along a time slice.
(b) Reflection strength blended with discontinuity sharpened through Laplacian filtering.
Data from the Taranaki Basin, offshore New Zealand.
line drawings of faults and channels and other geological features, especially when shar-
pened, so it blends well with amplitude, frequency, sweetness, and other stratigraphic
attributes, which provide color between the lines (Figure 5). Fault attributes serve the
same purpose as discontinuity in volume blending. Curvature attributes also appear in
volume blending, though they serve less well because blending obscures their finer
detail. These attributes are suitable for seismic data in general, whether the structure is
simple or complex.
Relative amplitude change acts as a directional high-resolution discontinuity attribute,
revealing details in faults and channels. In contrast to discontinuity, it is more suitable
where geological structure is uncomplicated.
Dip, azimuth, and shaded relief

Dip and azimuth find occasional application as background attributes in volume
blending. Dip magnitude is applied like discontinuity, and azimuth is applied like seis-
mic shaded relief. Exaggerating the dip enhances subtle features. Azimuth represents a
compass angle and appears directional when displayed with a circular gray scale, which
helps in volume blending. Rotating the gray scale changes the apparent direction of
illumination.
It is natural to combine dip and azimuth to record reflection orientation. They are
sometimes combined through volume blending, with azimuth controlling the color of
the display and dip controlling the shading

(Dalley et al., 1989; Marfurt et al., 1998;
Randen et al., 2000). More usefully, they

are combined as directional dip or as seis-
mic shaded relief, which is superior as a
background attribute in volume blending.
Seismic shaded relief looks like natural
topography. It is an illumination technique
like bump mapping except it illuminates
seismic reflection surfaces instead of ampli-
tudes. Blending shaded relief with seismic
data or attributes makes horizontal slices
look like aerial photographs of topography,
and vertical sections look like rugged can-
yon walls. Shaded relief is effective where
geological structures are complex but offers
little where structures are simple.
Caution must be exercised in interpret-
ing seismic attributes blended with shaded
relief. The apparent topography seen on a
horizontal slice through a shaded relief
volume might accurately represent the sub-
surface structure over a large vertical range
around the slice, but the blended attribute
necessarily derives exactly from the time
or depth of the slice. Thus, whereas shaded
relief might closely resemble an interpre-
ted horizon, a blended attribute will almost
never resemble the attribute extracted along
the horizon.
Figure 6. Structural attributes compared

with reflection strength through blending along
Examples of volume blending
a time slice (2404 ms). (a) Original seismic Figure 6 compares various structural
data. (b) Reflection strength. (c) Discontinuity attributes blended with reflection strength
and (d) Laplacian-filtered discontinuity along a time slice. The structural attribute
blended with strong values of reflection
that is best for blending depends on the
strength. (e) Most negative curvature. (f)
objective. Discontinuity is the most reliable,
Seismic shaded relief blended with reflection
strength. (g) and (h) Relative amplitude change but relative amplitude change shows nearly
in two orthogonal directions blended with the same features and appears more natural.
reflection strength. Yellow arrows indicate the Laplacian-filtered discontinuity and most
direction of illumination. Data from the negative curvature provide more detail than
Taranaki Basin, offshore New Zealand. discontinuity.
Figure 7. Multiple probes allow more than two attributes to be viewed at the same time.
Here, the smaller probe blends seismic shaded relief with strong values of reflection strength,
highlighting a prospective channel. The larger probe blends the original seismic data with shaded
relief. The vertical sections in the background show a discontinuity attribute. Illumination is in
the direction of the yellow arrows. The high resolution of the discontinuity complements the
lower resolution of the shaded relief in showing the geological structure. Reflection strength
reveals bright events, and the seismic data provide a reference for comparison. Data from the
With multiple probes, any number of seismic attributes can be compared readily. In
Figure 7, three probes show seismic data and three attributes: reflection strength, disconti-
nuity, and seismic shaded relief. They combine different kinds of stratigraphic and struc-
tural information in one view. This procedure can be extended to any number of probes.
Crossplots
Crossplots show relationships between two seismic attributes. A crossplot comprises a
scatter of points, where the coordinates of each point are the values of the two attributes
at the same location. Three-dimensional crossplots comprise points defined by three
attributes, but they are awkward to deal with and are seldom encountered.
Crossplot patterns take many forms. Attributes that are unrelated tend to produce
amorphous clouds of points (Figure 8). A cloud pattern indicates either that the attributes
Figure 8. A typical crossplot of reflection strength and average frequency derived along a
horizon. The input attributes have strong signal, so the amorphous cloud of points suggests that
the two attributes are independent. There are no anomalous regions. Most points in the cloud
have moderate amplitude and a frequency within the signal bandwidth, which is about
20– 50 Hz. The strongest amplitudes are confined to this bandwidth and represent seismic signal.
Weak amplitudes that lie outside this bandwidth likely represent noise.
contain unique information or that one or both attributes are dominated by random noise.
Many attributes are closely related and produce simple crossplot patterns, such as lines or
parabolas. Simple crossplot patterns signify duplicate attributes.
Crossplot analysis is usually applied to prestack attributes, but it is also effective with
poststack attributes. A typical workflow takes two attributes with independent information,
crossplots them, selects anomalies in crossplot space, and maps the anomalies back to the
coordinate space of the seismic data to show their spatial distribution. Such attribute
anomalies may indicate prospects, but their definitions rarely have more than local
significance.
The correlation coefficient is the standard similarity measure for linearly related
attributes. For nonlinearly related attributes, the most widely applied similarity measure
is the rank correlation coefficient.
Correlation
The correlation coefficient is familiar as a measure of similarity between two seismic

traces. It also serves as a measure of similarity between two attributes, whether the
attributes take the form of traces, maps, or volumes.
The correlation coefficient r between attributes a1 and a2 that have N samples is
defined as
1 N
(a − m1 )(a2n − m2 )
n=1 1n
r=N , (1)
s1 s2
where m1 and m2 are the mean values of the two attributes, and s1 and s2 are their standard
deviations. In a crossplot of a1 with a2, the correlation coefficient measures how closely
the scatter of points lie along the straight line that fits the points in a least-squares sense.
The correlation coefficient is appropriate for linear relationships but is inappropriate for
strongly nonlinear relationships.
Rank correlation
The rank correlation coefficient is a nonlinear estimate of how closely two attributes are
related (Isaaks and Srivastava, 1989, p. 31). It is more robust than standard correlation
because it is less sensitive to outliers. Rank correlations are appropriate when the two
attributes are related through a monotonic function so that they both increase or decrease
in lockstep.
Rank correlation replaces the values of an attribute with their ranks to produce a
rank attribute R. The ranks are found by ordering the attribute values in a linear array
from smallest to largest value. The rank of an attribute value is its index in the array.
As an example, a trace attribute with values (2.5, 4.0, 7.8, 5.1, 3.6) transforms to a rank
attribute with integral values (1, 3, 5, 4, 2). The rank correlation coefficient rrank
between attributes a1 and a2 is defined in terms of the correlation of their corresponding
rank attributes, R1 and R2, as
1 N
(R1n − mR1 )(R2n − mR2 )
= N
n=1
rrank , (2)
sR1 sR2
where mR1 and mR2 are the mean values of the rank attributes, and sR1 and sR2 are their
standard deviations.
Crossplot example
Figure 9 shows a crossplot between reflection strength and parallelism. The two
attributes appear quite different, and their crossplot suggests only a weak relationship.
Figure 9. A crossplot of reflection strength and parallelism. Strong reflection strengths are red,
weak are blue; parallel reflections are yellow, nonparallel are blue. The two attributes are
related so that strong amplitudes are highly parallel, whereas nonparallel reflections have weak
amplitudes. The rank correlation coefficient suggests a closer relationship than the correlation
coefficient does. Data from offshore Australia.
Their correlation coefficient is 0.39, proving that they are somewhat related. The rank
correction is 0.44, suggesting that the two attributes are related more closely. Inspection
reveals that strong amplitude events are parallel, and highly nonparallel events have
weak amplitude.
This example highlights a pervasive weakness that hinders multiattribute analysis in
general: many attributes are correlated. This is true even for attributes whose definitions
would seem unrelated, such as amplitude and parallelism. The problem is partly inherent
in the seismic method: seismic imaging performs best where structure is simple and
noise is low. In this case, the data tend to be more coherent and have stronger amplitude.
Coherent data tend to be more parallel, so parallelism tends to correlate with amplitude.
Where structures are more complicated and imaging is less successful, attributes are
noisier and correlate less. For this reason, attributes tend to be better correlated along
interpreted horizons, which are usually picked on clear reflections rather than on entire
seismic lines, which typically have mixed reflection quality.
Principal component analysis

Principal component analysis is a general least-squares method for analyzing multi-
dimensional data (Kirlin, 1999; Duda et al., 2001, p. 115). It rearranges the information
in a set of linearly related seismic attributes through translation and linear combination
to produce a new set of attributes called “principal components.” Most seismic attributes
are partly correlated with each other, but principal components are uncorrelated by
design. Principal component analysis separates the dominant information from the less
significant information or noise. It finds modest application in creating unique attributes
and in attribute space reduction.
Principal component analysis finds the set of orthogonal axes, centered on the centroid
of the attribute data, that best fit the data in a least-squares sense. It projects the attributes
along these axes to produce the principal components. The principal components are
unique because the axes are orthogonal. By convention, they are arranged in order of
decreasing energy. The strongest components show the most characteristic features and
represent strong signal or coherent noise. The weaker components represent random
noise or weak signal. Like crossplots, principal component analysis cannot by itself
distinguish noise from signal; this must be done by inspection. Like correlation, it identifies
linear relationships between attributes and is inappropriate for attributes with highly non-
linear relationships.
Illustration of principal component

analysis
To illustrate the idea behind prin-
cipal component analysis, consider a
study of the heights and weights of a
group of people. The goal of the study
is to characterize the people as either
underweight, overweight, or relatively
normal in weight. Principal component
analysis transforms the heights and
weights into two new attributes, princi-
pal components p1 and p2. The dominant
Figure 10. A plot of points representing the
principal component p1 is the position heights and weights of a set of people. Height
along a normal height-weight trend line, and weight attributes h and w are transformed
and the second component p2 measures into principal components p1 and p2. Component
the deviation from this trend (Figure p1 represents the normal height-weight trend and
10). Because the deviations from the component p2 represents the deviation from
normal trend are of interest here, the this trend.
data are projected onto the p2 axis to discard the dominant component and reduce the data
to a single dimension, thereby simplifying subsequent analysis. However, the meaning of
the new attribute p2 is not as simple as it might seem, because a given deviation from the
normal is less important for a large person than for a small person. The original attributes
may remain preferable because they have clearer meaning.
Principal component analysis of seismic attributes exhibits similar strengths and
weaknesses. The key differences in attribute analysis are that the dominant components
are retained, not discarded, and the meanings of the principal components are more
obscure.
Attribute spaces and attribute vectors

Attribute spaces and attribute vectors are central to multiattribute analysis. An attribute
space is a space defined by a set of seismic attributes, each of which constitutes an axis
of the space. The dimension of the space equals the number of attributes, and its boundaries
are limited by the attribute ranges. The most useful attribute spaces comprise independent
attributes. The crossplot of Figure 9 represents an attribute space of two dimensions defined
by amplitude and frequency.
An attribute vector is an ordered set of attribute values measured at a specific loca-
tion in a seismic volume. Attribute vectors define points in attribute spaces (Figure 11).
Each component represents the value of a different attribute. Attribute vectors sim-
plify the mathematics of multiattribute analysis, especially when the number of attri-
butes is large. The mathematics of attribute vectors is the same as that for vector traces
(Appendix E).
Method of principal component analysis

A set of M seismic attributes is transformed
into M principal components as follows. Let
each attribute have N samples. From each attri-
bute, subtract its mean. Rearrange the attributes
as a set of N attribute vectors of dimension M
and form the covariance matrix of the attribute
vectors so that the matrix elements are the
covariances of the attributes in the set (Appendix
E, equation E-17). Find the eigenvalues of the
covariance matrix, arrange them in order of
decreasing magnitude, and derive their associ-
ated unit-length eigenvectors. The eigenvectors
Figure 11. An attribute vector define orthogonal axes about which the data var-
represents a point in an attribute space. iances are minimum, and the eigenvalues rep-
Here, vector A corresponds to the point resent the variances. Transform the attributes
(160, 25) in a 2D space defined by into principal components by projecting the attri-
amplitude and frequency attributes. bute vectors onto the eigenvectors through dot
products. The projection of an attri-

bute vector B onto an eigenvector êi
produces the value bi,
bi = B · êi . (3)
Figure 12 illustrates the process. A

2D space is composed of attributes u
and v. The center of gravity of the
space is defined by the average values
of u and v and is represented by
vector C. Principal component analy-
sis of attributes u and v yields principal
components p1 and p2. An attribute Figure 12. Vector A with coordinates (a1, a2) in
vector A in the space of u and v the attribute space defined by attributes u and v
becomes the vector B ¼ A 2 C in the becomes vector B with coordinates (b1, b2) in the
space of p1 and p2. The principal com- space defined by the principal components p1 and p2.
ponent values of B are its coordinates Vector C represents the center of gravity of the
in the space of p1 and p2. attribute space and defines the origin of the principal
component axes.
Attribute space reduction

An attribute space is reduced by removing redundant information and noise from a set
of seismic attributes to produce a new and smaller set of seismic attributes. Attribute space
reduction is applied primarily to decrease the computational cost in automatic pattern
recognition, which is roughly proportional to the number of attributes.
Attribute space reduction is readily accomplished through principal component analy-
sis (Linari et al., 2003). The idea is to transform a set of seismic attributes into their
principal components and then discard those components that are dominated by random
noise, which are usually the weaker components. This assumes that the original attributes
include redundant information that is consolidated in the stronger components. The draw-
back is that principal components lack simple and familiar meaning, which makes principal
component analysis a poor solution for the problem of too many attributes.
Figure 13 shows an example of principal component analysis applied to a set of five
related seismic attributes: reflection strength, sweetness, amplitude variance, bandwidth,
and average wavelength. The first principal component closely resembles reflection
strength and sweetness, and represents the dominant information in the attributes. It is
roughly 10 times stronger than the second principal component and 100 times stronger
than the fifth principal component. The fifth component is dominated by noise and can
be discarded from subsequent analysis. The remaining components likely represent
signal, but they don’t reveal interesting features and their meanings are obscure. The orig-
inal attributes are highly correlated, but they have the advantage of clearer meaning. This
example is typical. As a result, attribute space reduction through principal component
analysis finds only minor application despite its relative simplicity.
Figure 13. Principal component analysis of five seismic attributes of the seismic data shown in
Figure 21, Chapter 4. The attributes are (a) reflection strength, (b) sweetness, (c) amplitude
variance computed in a window of 5 lines by 5 traces by 7 samples, (d) bandwidth computed in
a 40-ms window, and (e) wavelength computed in a window of 3 lines by 3 traces by 7 samples.
The five principal components are arranged in order of decreasing strength: (f) first, (g) second,
(h) third, (i) fourth, and (j) fifth. The first principal component is much stronger than the other
components and resembles reflection strength. The fifth component is weak and noisy and can
be discarded to reduce the dimension of the attribute space from 5 to 4.
Automatic pattern recognition

Automatic pattern recognition of seismic data is largely a matter of classification.

Classification generalizes seismic data by categorizing the patterns prevalent in the data.
Patterns are characterized by ranges of attribute values. The underlying assumption for
interpretation is that similar patterns imply similar geology. Most methods of automatic
pattern recognition do not indicate directly what the geology might be, so results must
be interpreted qualitatively.
Classification is supervised or unsupervised. Supervised classification is learning by
example; the geophysicist provides the patterns for the algorithm to find in the seismic
data. Unsupervised classification is learning by doing; the algorithm automatically dis-
covers representative patterns in the seismic data. Supervised approaches are attractive
because the geophysicist can choose patterns that have specific geological meaning. In
contrast, patterns found by unsupervised classification lack inherent geological meaning
and may not relate to features of interest. Nonetheless, unsupervised methods enjoy
greater commercial success because they are easier to apply.
An attribute class is a set of attribute values, organized as a vector or as a set of ranges,
that represents a pattern in seismic data. Attribute classes are also called “hybrid attributes”
(Taner, 2001), “meta-attributes” (de Rooij and Tingdahl, 2002), and “seismic facies”
(Linari et al., 2003). Attribute classes can represent almost any seismic pattern; those of
chief interest have geological or geophysical significance. In seismic waveform classifi-
cation, template waveforms take the place of attribute classes.
The number of classes in a particular classification is chosen by the geophysicist or
is determined algorithmically. Attributes rarely exhibit a natural number of classes. The
number of classes is usually chosen to achieve a desired resolution or generalization in
the classified data. Typically, it ranges from 5 to 100.
The value of classification depends on the seismic attributes employed. It is important to
choose good attributes, but it is hard to know how “good” any attribute is in a particular case.
The contributions that individual attributes make to pattern recognition can be estimated
algorithmically, if laboriously (Schuelke and Quirein, 1998). Fortunately, an experienced
geophysicist can usually judge which attributes are most suitable for pattern recognition.
Different attributes occupy different numerical ranges, which complicates the compari-
son of attribute values. The standard solution is to rescale each attribute to have zero
mean and unit standard deviation. Attribute vectors comprised of rescaled attributes can
be compared using simple similarity measures.
Classification assigns each vector to the class that it matches best. The match is deter-
mined through a similarity measure, such as Euclidian distance. As in seismic waveform
classification, classified attribute data are necessarily quantized, meaning their values
are limited to specific values. As a result, classified data exhibit sharp boundaries. Sharp
boundaries appeal to the eye, but their locations are somewhat artificial and cannot be
trusted. This is because classification must impose boundaries even where attributes
change gradationally.
Classified seismic data are often noisy. Standard smoothing filters, whether linear or
nonlinear, are inappropriate for cleaning classified data because typically classes are unre-
lated to each other. Imagine data classified into five rock types: sand, silt, shale, limestone,
and coal. In this context, the average or median makes little sense, but the mode, the value
that occurs the most, does. Mode filtering is appropriate for cleaning and smoothing any
quantized data (Hall, 2007). Applied to classified seismic data, mode filters operate in
a running analysis window and replace the class at the window center with the class that
occurs most in the window. The shape of the analysis window is the same as that of a
typical running average or median filter. Like most 3D filters for seismic attributes,
mode filters perform better when they follow structure.
Distance measures
In most pattern recognition algorithms, the similarity of two attribute vectors is quan-
tified by a distance measure, with smaller distances implying greater similarity. The two
most important measures are Euclidian distance and Manhattan, or city-block, distance.
They do not require zero-mean attributes and are easier to evaluate than correlation
coefficients. The Euclidian distance de between attribute vectors x i and x j with N attributes
is the square root of the sum of the squares of the differences of their components:

N

de = |xi − xj | = (xin − x jn )2 . (4)
n=1
Manhattan distance dm is the sum of the absolute values of the differences of their
components:
N
dm = |xin − x jn |. (5)
n=1
Figure 14 illustrates these two measures. Manhattan distance is simpler and faster
to compute than Euclidian distance, which is attractive when comparing millions of
vectors. However, classification with Euclidian distance tends to converge faster, reducing
overall computation time and suggesting that it is the more natural measure. Most pattern
recognition algorithms employ Euclidian distance.
Figure 14. Illustration of Euclidian distance de and

Manhattan distance dm between attribute vectors A
and B in a two-dimensional space defined by
attributes a1 and a2. The Euclidian distance is the
shortest distance between the ends of the vectors
(long dashes). The Manhattan distance is the shortest
distance along a path that everywhere parallels an
axis (short dashes). In this 2D example, the
Euclidian distance is the length of the hypotenuse of
the right triangle, and the Manhattan distance is the
sum of the lengths of the two other sides. Euclidian
distance is always less than or equal to Manhattan
distance. These ideas carry over to an attribute space
of arbitrary dimension.
Substituting Mahalanobis distance for Euclidian distance avoids the necessity of

rescaling the attributes (Duda et al., 2001, p. 36). Mahalanobis distance is a method for
comparing attributes that are correlated to each other and have Gaussian distributions,
but it is computationally demanding and is rarely implemented in seismic pattern
recognition.
Supervised classification
Supervised classification classifies seismic data with a set of user-defined attribute
classes (e.g., Hampson et al., 2001). Typically, the classes are defined at locations in the
seismic data where geological or geophysical properties are known, such as at wells. In
this way, the classification inherits clear and useful meaning. Once the classes have been
defined, classification is fast and has global applicability. The drawback is that defining
a suitable set of attribute classes requires prior knowledge and considerable effort and
skill. As a result, supervised classification is inconvenient and often unsuitable.
Unsupervised classification
Unsupervised classification classifies a set of seismic attributes with representative
classes that it finds automatically. It discriminates major features well and is easy to
apply. It sometimes reveals channel systems and other patterns whose geological signi-
ficance is readily inferred by inspection. These advantages make unsupervised classi-
fication more popular than supervised classification. Its chief drawback is that the
representative classes lack inherent geological meaning and global applicability. In
some cases, the meaning of a class is suggested by the attributes, or can be inferred by
comparison with other information.
Most methods of unsupervised classification implement a workflow that comprises
subsetting, initialization, training, and final classification (Figure 15). Such methods
resemble nonlinear filtering and entail a similar computational cost.
A small representative subset of the data suffices for training. This is fortunate because
training is iterative, and without subsetting it becomes costly on large data sets. By employ-
ing a subset of 1% or less of the total data size, the time spent training the classifier
reduces to relative insignificance, and the total time for classification becomes roughly
linear with the number of classes and number of attributes.
Initialization makes a first estimate of the classes. Training refines the classes iteratively
and continues until the classes converge to a solution. The final classes are applied to the
entire data set to produce a classified volume.
K-means clustering is employed occasionally for unsupervised classification of
seismic data, but by far the most common algorithm is the Kohonen self-organizing
feature map, or Kohonen SOFM (Strecker and Uden, 2002; Trappe, 2002). Both the
Kohonen SOFM and K-means clustering create a specified number of classes and populate
them roughly equally. K-means clustering is simple and straightforward, but it does not
arrange its classes in any particular order. In contrast, the Kohonen SOFM orders its
classes naturally. As in waveform classification, this is a significant advantage, which is
why the Kohonen SOFM is favored

(Coléou et al., 2003). Classes produced
by K-means clustering can be reordered

in some fashion, such as by vector
length. Reordering by vector shape,
though conceptually appealing, proves
more difficult than clustering itself.
Hierarchical classification is some-
times applied in multiattribute analy-
sis (e.g., Linari et al., 2003). However,
it is suited only for small data sets, and
it must be modified greatly to handle
typical seismic data volumes. Its pur-
ported advantage is that it can discover
the natural number of classes in an attri-
bute space. Because seismic attributes
rarely exhibit natural classes, hierarchical
classification is inappropriate in multi-
attribute analysis.
Image texture analysis has long been
promoted for classifying seismic data
(Love and Simaan, 1984; Gao, 2003;
Chopra and Alexeev, 2006). Texture
analysis replaces seismic attributes with
generic image textures and combines
them in specific ways to identify patterns.
It has not been shown to work better
than more conventional methods of auto-
matic pattern recognition that employ
seismic attributes, and it has the draw-
back that image textures lack direct
Figure 15. Workflow of an unsupervised
relationships to geophysical and geologi-
classifier. Both K-means clustering and the
Kohonen SOFM follow this workflow.
cal properties.
K-means clustering
K-means clustering classifies a set of attribute vectors into K classes, or “clusters,”
where K is chosen by the user. The clusters lack connections between them. There are
many variants of K-means clustering. The basic variant iterates through the training data
and updates all clusters simultaneously after each complete iteration. It employs a
winner-takes-all strategy in assigning attribute vectors to clusters, where the “winner” is
the cluster that is closest to the attribute vector. In this way, each class learns independently
of the other classes. Other variants of K-means clustering try to improve cluster learning
by exaggerating cluster changes, sharing vectors between clusters, updating clusters with
each new member vector, minimizing cluster spreads, and so on. Most refinements have
only a modest effect on results or performance.
K-means clustering is typically initialized with attribute vectors selected randomly

from training data. This approach is unsatisfactory because it produces different clas-
sifications for the same training data and parameters. The different classifications broadly
resemble each other but differ substantially in their details. Recursive initialization is better.
In recursive initialization, the training subset is itself subsetted, and the smaller subset
is used to find representative attribute vectors to initialize the larger subset. This
procedure is applied recursively until the smallest possible subset is reached. Its attribute
vectors are taken to initialize the training of the next largest subset, which is trained
to find representative vectors. These representative vectors initialize the training of the
next largest subset, and so on, until the largest subset is trained. This method produces
repeatable results and converges faster than training with random initialization.
Kohonen SOFM
The Kohonen SOFM is a neural
network that acts like K-means cluster-
ing with continuous updating and with
connections between the attribute
classes, or “neurons” (Duda et al.,
2001, p. 576). The connections enable
winning neurons to share training and
to organize themselves in an order
so that neighboring neurons represent
similar seismic patterns. This is the
strength of the method. The neighbor-
hood of any neuron is set by a window
whose breadth decreases with training.
The window shape weights the relative Figure 16. Examples of (a) linear, (b) grid, and
degree of learning for the neurons in a (c) honeycomb map architectures for a Kohonen
neighborhood. Typical neighborhood SOFM with 12 neurons. These architectures govern
functions take the shape of a boxcar, how attribute classes relate to each other in the
Gaussian, or a “Mexican hat,” which attribute space. A class is similar to its neighbors.
has the shape of a Ricker wavelet. Here, for the linear architecture, class 6 is most
The architecture of the connections similar to classes 5 and 7 and probably least similar
between the various neurons is called to class 12. For the grid architecture, class 6 is
the feature map. Feature maps typically most similar to classes 2, 5, 7, and 10 and probably
least similar to classes 4 and 12. For the
form lines, grids, or honeycombs,
honeycomb architecture, class 6 is most similar to
though they can assume any desired
classes 5, 7, and 12, and probably least similar to
architecture (Figure 16). The Kohonen classes 1, 2, 3, and 9. This analysis is complicated
SOFM updates the neurons with each because these architectures are applied in
training attribute vector according to a multidimensional attribute spaces in which they
learning law that changes each neuron warp to fill the space.
by a fractional amount of its difference between the neuron with the training vector. The
fractional change is the product of the learning rate and the neighborhood function.
The learning law is
ỹ = y + w(n)a(n) · (xn − y), (6)
where ỹ is the updated neuron, y is the original neuron, x n is the nth training vector, w(n)
is the neighborhood function of the neuron, and a(n) is the learning rate. Both the neigh-
borhood function and the learning rate decrease as the number of training vectors n
increases.
The Kohonen SOFM is initialized with small random vectors. During training, the
neurons move about in the attribute space to assume the desired map architecture. Unlike
K-means clustering with random initialization, the final classes are independent of the
initial classes. Classes created by the Kohonen SOFM have slightly more spread in their
constituent attribute vectors than those created by K-means clustering, but the difference
is rarely important.
Examples of attribute classification

Figure 17 compares the classification of seismic data by K-means clustering with that
by the Kohonen SOFM. The two classifications find similar patterns, though they assign
different colors to their classes. They employ five attributes: reflection strength, amplitude
variance, reflection spacing, parallelism, and divergence. Classification is fast, requiring
no more calculation time than a basic 3D seismic attribute. As is often the case, amplitude
drives the classification more than the other attributes.
Figure 17. Classification of seismic attributes into 20 classes by (a) K-means clustering and
(b) the Kohonen SOFM, for the seismic line shown in Chapter 4, Figure 21; the data are from
offshore Australia. Five seismic attributes are employed: reflection strength, amplitude variance,
reflection spacing, parallelism, and divergence (refer to Figures 21 in Chapter 4 and 18 in
Chapter 5).
Figure 18. Mode filtering to clean classified seismic data showing a large channel crossing a
fault. (a) The classified data viewed along a horizon. (b) The classified data after mode filtering
with an operator of 5 lines by 5 traces by 5 samples. The data are classified into 20 classes by
the Kohonen SOFM using four seismic attributes: reflection strength, average frequency,
parallelism, and continuity. Data from the Taranaki Basin, offshore New Zealand.
Figure 18 illustrates the effect of smoothing classified seismic data. Because filters
based on averages or medians are inappropriate for classified data, smoothing is accom-
plished through mode filtering, which selects the prevalent value in the analysis window
as the output. Mode filtering is fast and should be applied when the classification
appears noisy or when there is a need to reduce its resolution.
Summary
Multiattribute analysis is founded on the premise that if a single attribute helps solve a
problem, then multiple attributes should help more. This premise has merit, but because
multiattribute analysis requires greater effort and understanding than analysis of individual
attributes, it attracts only moderate attention.
The main methods of multiattribute analysis are volume blending and cross-plotting.
They are widely available and are popular because they are easy to apply and comprehend.
Volume blending offers an insightful way to combine attributes to aid the visual inspec-
tion of seismic data. Cross-plotting shows relationships between attributes that identify
redundant measures or reveal anomalies.
Principal component analysis appears to offer an attractive solution to the problem
of redundant seismic attributes. Given a set of attributes, principal component analysis
combines duplicate information and isolates unique information and noise to produce a
new set of uncorrelated attributes called principal components. Principal components are
arranged in order of their energy, the strongest first and the weakest last. Dropping
the weakest components achieves attribute space reduction. Unfortunately, principal com-
ponent analysis suffers three drawbacks: the principal components lack clear meaning, the
input attributes are often not linearly related as required, and the method does not
inherently distinguish signal from noise. These drawbacks make principal component
analysis a poor solution for the problem of duplicate or meaningless attributes. Experience
remains the best guide for choosing useful attributes.

Automatic pattern recognition of geophysical data is a large and involved topic and
merits its own book. The breezy description offered here covers only a few key concepts.
Many methods of pattern recognition have been tried in reflection seismology, but few have
met with commercial or scientific success. Automatic pattern recognition stubbornly
remains a minor tool in seismic interpretation, even as it retains alluring promise.
Though there are many seismic attributes, there is a scarcity of independent and useful
attributes. Stratigraphic attributes are especially lacking because stratigraphic properties
are difficult to quantify. Better seismic attributes must be developed for multiattribute
analysis to progress.
Multiattribute analysis is further hindered by the difficulty of automatically assigning
geological meaning to patterns in seismic data. The patterns revealed by unsupervised
classification and principal component analysis rarely have obvious geological meaning
unless they are recognizable as channels or other stratigraphic features. Future methods
of automatic pattern recognition need to incorporate additional geophysical and geological
information to endow results with geological meaning.
This concludes the exposition of poststack seismic attributes and methods. It remains
only to review their application to common tasks of seismic interpretation. This is the topic
of the next, and final, chapter.
Chapter 10
Applying Seismic Attributes
Introduction
Poststack seismic attributes are applied in seismic data interpretation to aid reconnais-
sance, identify bright spots and frequency anomalies, map faults and channels, reveal
details, highlight trends, and enhance technical presentation. Attribute methods also find
modest application in seismic data processing.
Seismic attribute analysis begins by choosing attributes that are appropriate for the
purpose at hand. This is facilitated by discarding duplicate and flawed attributes to obtain
a smaller set of relatively useful attributes. Certain attributes prove to be most successful
for certain objectives. Seismic attributes are clearer when derived from clean seismic data.
This chapter reviews the application of poststack seismic attributes in reflection
seismology.
Choosing suitable attributes

Hundreds of seismic attributes have been invented, and more appear each year. Their
great number and variety is confusing and makes it difficult to choose between them. It
is impractical to test them all to see which attributes help in a particular setting and
which do not. It is also unnecessary. Most seismic attributes can be discarded because
they are duplicates or unstable, or lack useful meaning (Figure 1). Discarding redundant
and flawed attributes leaves a much smaller and more manageable set of attributes that
are relatively unique, stable, and meaningful. From among the useful attributes, certain
attributes are identified as serving best for specific objectives.
Redundant and flawed attributes

Many seismic attributes differ only in detail. This is especially true of attributes for
amplitude, frequency, and discontinuity.
Consider amplitude. There are more than a dozen common amplitude attributes.
All contain nearly the same information. For most purposes, nothing is gained by using
more than one. Reflection strength or root-mean-square (rms) amplitude usually suffices;
discard the rest.
Interval attributes that record maxima or minima, such as smallest value or maximum
peak amplitude, should be applied only when maxima or minima are of specific interest.
They have the advantage that they retain polarity, but they are sensitive to noise and less
likely to represent the interval.
177
Figure 1. There are hundreds of seismic attributes, but many are duplicates, redundant, or flawed
and can be discarded.
Cosine of the phase acts less like a seismic attribute and more like a strong amplitude
gain, making reflection continuity easier to follow. Treat it as such.
Frequency attributes exhibit a great variety of design and purpose, but some are dup-
licates or flawed. Average frequency, rms frequency, and bandwidth can be computed
either in the time domain or in the frequency domain; choose one approach and discard
the other. Dominant frequency attributes measure a variety of quantities because the
term “dominant frequency” lacks definite meaning and could imply a spectral average, or
a spectral peak, or an rms frequency. Dominant frequencies derived from maximum
entropy spectral decomposition are noisy and lack geological or geophysical meaning;
discard them. Response frequency does not record the average frequency of the seismic
source wavelet as advertised but instead resembles a blocky version of the average fre-
quency. Prefer average frequency.
Arc length and sweetness are driven by both amplitude and frequency. Prefer standard
amplitude and frequency attributes because they have clearer meaning. Otherwise, apply
arc length to identify strong amplitude high-frequency events, and apply sweetness to
identify strong amplitude low-frequency events.
Energy half-time closely resembles vertical relative amplitude change. Choose one and
discard the other. The thin-bed indicator does not indicate thin beds and shows the same
spike anomalies as relative amplitude change. Discard it. Discontinuity attributes based
on novel combinations of principal components of the seismic data resemble standard
discontinuity attributes and can be discarded.
Avoid unstable attributes. These include apparent polarity, which is sensitive to reflec-
tion interference, the number of peaks or troughs in an interval, which is sensitive to the
interval definition, and instantaneous quality factor, which is unstable at amplitude
peaks and troughs.
Chapter 10: Applying Seismic Attributes 179
Avoid attributes that lack clear and useful meaning. Average instantaneous phase is
useless because it tends toward zero. Response bandwidth, based on instantaneous band-
width, is worse because it is always zero by definition. Slope of the instantaneous frequency
lacks useful geological or geophysical meaning; avoid it. Both amplitude-weighted instan-
taneous phase and perigram-weighted cosine of the phase are distorted versions of the orig-
inal seismic trace; discard them. Likewise, discard reflection intensity, gradient magnitude,
and Karhunen-Loeve signal complexity.
Useful attributes
Choose an attribute depending on the exploration objective. The two most useful post-
stack seismic attributes are reflection strength and discontinuity. Attributes of secondary
value include maximum amplitude, average frequency, most positive and most nega-
tive curvatures, spectral decomposition, instantaneous phase, waveform, relative acoustic
impedance, seismic shaded relief, and relative amplitude change. Other poststack attributes
find more limited employment.
Table 1 summarizes the attributes that are best at recording important seismic proper-
ties. This table is greatly condensed compared to the attribute categorization presented in
Chapter 1. Table 2 prescribes specific attributes to apply in common applications.
There are no “subsalt attributes,” “carbonate attributes,” “coal attributes,” or other such
geology-specific poststack seismic attributes. Subsalt seismic data are often noisy and
severely band limited. In this case, seismic attributes offer little. Where subsalt data are
of reasonable quality, attribute analysis is applied as it would be elsewhere. Seismic attri-
bute analysis has met with less success in carbonate exploration than in clastics because
amplitude, the most important seismic property, is less diagnostic in carbonates (Abriel,
2008, p. 91). However, structural attributes, phase attributes, and most frequency attributes
are applied in carbonates in the same way that they are applied in clastics. Discontinuity,
curvature, and amplitude change attributes are effective in outlining pinnacle reefs and
buried karst topography (Sullivan et al., 2006; Chopra and Marfurt, 2007, p. 230). Coal
layers tend to be seismically thin and have low density. They are analyzed the same way
as any thin layer that has low impedance.
Most seismic attributes require little or no parameterization beyond selecting a window
size. Choose small windows to enhance resolution and large windows to reduce variance.
Table 1. Seismic attributes suitable for measuring key seismic properties.

Amplitude Phase Frequency Discontinuity Structure Miscellaneous
Reflection Inst. phase Average Discontinuity Dip Spectral
strength frequency decomposition
rms Response Bandwidth Relative Azimuth Waveform
amplitude phase amplitude
change
Maximum Tuning Curvature Parallelism
value frequency
Table 2. Seismic attributes suitable for common applications.

Bright spots Faults Channels Shadows Reconnaissance Stratigraphy
Reflection Discontinuity Reflection Average Reflection Instantaneous

strength strength frequency strength phase
Relative Curvature Discontinuity Spectral Discontinuity Reflection
acoustic decomposition strength
impedance
Maximum Relative Spectral Reflection Shaded relief Average
value amplitude decomposition strength frequency
change
Dip Waveform Inverse Q RGB-blended Relative
spectral acoustic
decomposition impedance
Shaded Relative Parallelism
relief acoustic
impedance
Relative Waveform
amplitude
change
Preparing attributes for interpretation

Seismic attributes are easier to interpret when they are cleaner. Attributes are made
cleaner and more interpretable by coherency filtering the seismic data before computing
attributes (Höcker and Fehmers, 2002; Sheffield and Payne, 2008; Henning et al., 2010),
or by filtering the attributes directly. Coherency filtering should be applied routinely
prior to attribute analysis. With fault preservation, it is especially effective at enhancing
discontinuity and other structural attributes (Figure 2). All data examples in this chapter,
as well as most examples elsewhere in the book, have been coherency filtered. Vertical
median filtering of instantaneous frequency or horizontal relative amplitude change
reduces noise and removes confusing spikes due to reflection interference. Laplacian filter-
ing makes discontinuity attributes look more like line drawings, which often improves
them for interpretation or volume blending.
Seismic attributes reveal certain components of the seismic signal by removing other
components. They are not designed to improve the signal-to-noise ratio of seismic data. Do
not apply attributes to remove noise or enhance the seismic signal. For that purpose, apply
processes such as band-pass filtering, spectral enhancement, or coherency filtering.
Data limitations
Seismic attribute analysis assumes that seismic data represent an image of the earth.
In reality, the image is distorted in various ways that affect attributes. Errors in data
Figure 2. Seismic attributes are cleaner when derived from coherency-filtered seismic data.
(a) Original seismic data along a time slice (820 ms). (b) The seismic data after coherency filtering
with fault preservation. (c) Discontinuity derived from original seismic data. (d) Discontinuity
derived from coherency-filtered data. Data from the Taranaki Basin, offshore New Zealand.
acquisition or processing and inherent limitations of the seismic method cause defects in
seismic data that degrade seismic attributes.
Poor imaging and inconsistent acquisition distort the amplitude and frequency con-
tent of seismic data, making seismic attributes less reliable. These problems are
most pronounced in shallow and deep data, at survey boundaries, and at skips in data
recording.
Seismic attributes are suspect in the mute zone, especially the shallow half, which is
typically the first 500 ms of data. Here, the fold increases with depth and the offset mix
varies from trace to trace. A theoretical gain correction approximately accounts for differ-
ences in fold, but nothing is done to regularize the offset mix. A variable offset mix contrib-
utes to acquisition footprints, which strongly distort structural attributes. Acquisition
footprints are aggravated by errors in data processing, which tend to be worse in shallow
data (Figure 3). Moveout stretch is most pronounced in the shallow data and markedly
reduces the frequency content. For these reasons, shallow amplitude and frequency ano-
malies cannot be compared with deeper anomalies.
Incorrect velocities or statics corrections, as well as locally strong velocity variations,
degrade seismic images by causing mis-stacking in data processing. Mis-stacking acts as a
high-cut frequency filter, reducing both amplitude and average frequency (Ebrom, 2004).
Zones of mis-stacked seismic data masquerade as attenuation anomalies and can occur
anywhere, but they are most prevalent
in the shallow data where moveout cor-
rections are large and prone to error.
Seismic data are routinely proces-
sed to preserve true amplitudes. “True
amplitude” means that the amplitudes
of isolated reflections represent scaled
reflection coefficients so that amplitu-
des at different times can be compared.
However, amplitude preservation is
not always successful, especially in
structurally complex data. In addi-
tion, data sometimes exhibit significant
amplitude decay because reflection
coefficient magnitudes tend to decrease
with depth. Thus, even on true ampli-
tude data, shallow and deep amplitude
anomalies may not have the same
significance.
Amplitude attributes cannot be
trusted if automatic gain or trace balan-
cing have been applied. These processes
have negligible effect on frequency
attributes and most other attributes
(Figure 4). Time-variant filtering and
Figure 3. A shallow imaging problem in a seismic time-variant deconvolution distort fre-
volume likely caused by an error in data processing. quencies and complicate the compari-
(a) Volume view showing the problem at the time son of frequency anomalies at different
indicated by the black arrow. The problem is times.
visible on crosslines but not inlines. (b) Time slice
Most attributes and attribute
at 352 ms through the problem zone of the seismic
methods are insensitive to the phase of
data, blended with relative amplitude change. The
artifact is so severe that it completely obscures the seismic data, but a few require zero
other features in the relative amplitude change. The phase. These include apparent polarity,
yellow arrow indicates the direction of attribute instantaneous phase, response phase,
computation. Data from the Taranaki Basin, offshore relative acoustic impedance, waveform
New Zealand. modeling, and tuning analysis.
Figure 4. A prominent vertical amplitude defect caused by trace balancing through an unusually
strong event. (a) Seismic line. (b) Reflection strength. (c) Average frequency computed in a
window of 11 samples (44 ms). Trace balancing should be avoided because it destroys the
lateral integrity of the amplitudes, though it does not affect frequency.
Reconnaissance and presentation

Seismic data interpretation begins with reconnaissance and ends with presentation.
Though at opposite ends of the interpretation workflow, these two tasks share the same
objective: to simplify seismic data for easier comprehension. For this purpose, seismic
attributes prove ideal.
In data reconnaissance, the goal is to quickly gather an impression of the overall geo-
logical structure and data character. Faults and horizons have not yet been interpreted,
so map-based attributes must be set aside. Details are unimportant, so high-resolution
methods are not needed. Test reflection strength and discontinuity first. If data quality
permits, test average frequency, curvature, volume spectral decomposition, or seismic
shaded relief. Seismic shaded relief is particularly helpful in reconnaissance of structurally
complex seismic data, and it offers a preview of how interpreted horizons will likely look.
It is most effective when blended with the original seismic data or amplitude, or with both
as shown in Figure 5.
Figure 5. Volume reconnaissance is facilitated by volume blending with multiple probes. Here,
the box probe blends seismic shaded relief with reflection strength, and the arbitrary line probe
blends the original seismic data with shaded relief. Illumination is in the direction of the yellow
arrow. Major features can be identified quickly. Data from the Taranaki Basin, offshore
New Zealand.
In data presentation, faults and horizons have been mapped, prospects have been iden-
tified, and the task remaining is to present evidence in support of the interpretation.
Details are most important; the overall structure and character of the data are secondary.
Seismic attributes aid seismic data presentation by reducing data complexity and
highlighting features of interest, making the presentation both more comprehensible and
more attractive. Amplitude maps are all but essential. Maps showing faults or channels
become more interpretable when combined with high-resolution discontinuity. Waveform
classification and spectral decomposition provide complementary information. The role
of other attributes depends on the nature of the interpretation. Attributes often convey
more information when they are combined through volume blending.
Attribute analysis is a science, but its effective presentation involves as much art
as science.
Bright spots and amplitude mapping

Bright spots are seismic reflections with anomalously strong amplitudes. Geophy-
sically they are caused by large reflection coefficients, possibly enhanced by tuning or
reflection focusing. Geologically they are caused by gas, hard streaks, igneous intrusions,
or coal. In shallow young marine clastic sediments, bright spots often indicate gas and thus
serve as direct hydrocarbon indicators. It is for this reason that bright spot exploration has
long been a mainstay of reflection seismology. Unfortunately, bright spots caused by gas
do not necessarily indicate economically important gas accumulations. Sands with low
gas saturations often produce attractive bright spots (O’Brien, 2004). Indeed, beyond a
small threshold, the brightness of a reflection from a gas-charged sand is insensitive to
gas saturation (Liner, 2004, p. 511; Roden et al., 2012). Bright spots must be interpreted
cautiously.
Dim spots also indicate gas under certain conditions (see Brown, 2011, p. 161). They
occur deeper in seismic data than bright spots and are hard to recognize. Flat spots represent
a possible gas-water contact (Forrest et al., 2010).
For general amplitude mapping, apply reflection strength or rms amplitude. Apply
sweetness to identify bright spots associated with thicker reflection spacing. Prefer these
attributes for both volumes and maps. In clean seismic data, maximum amplitude interval
attributes sometimes produce sharper maps than attributes that involve averaging. When
polarity is important, use average peak or average trough amplitude, or the original
seismic data.
Reservoirs often appear as strong amplitude anomalies that conform to structure
and whose bounds indicate the gas-water or gas-oil contact (Forrest et al., 2010). These
anomalies are best imaged by amplitude maps extracted in an interval at the reservoir
level (Abriel, 2008; Ghosh et al., 2010; Figure 6).
Low-frequency shadows
A low-frequency shadow is a zone in seismic data that is characterized by anomalously
low-frequency content, and which results from strong signal attenuation caused by
Figure 6. Reflection strength extracted on a horizon through an anticline. The amplitudes

are conformable with structure, which suggests a gas-water contact. Example from offshore
Malaysia. After Ghosh et al., 2013, p. 86.
absorption or scattering. Seismic reflections within a shadow lack high-frequency content

relative to surrounding reflections. Shadows are expected to have clear beginnings and fade
out gradually over a hundred or several hundred milliseconds. Gas sands are the chief cause
of anomalous attenuation, and it is for this reason that low-frequency shadows are of inter-
est in exploration.
Unambiguous low-frequency shadows due to absorption are rare. Because seismic data
have limited bandwidth, it takes a great deal of absorption to diminish the signal spectrum
measurably. It is difficult to recognize true absorption shadows because the small overall
spectral change due to absorption is masked by the much larger and highly variable effects
of reflection interference. Scattering from shallow gas causes marked low-frequency
shadows in marine seismic data. Such anomalies are widespread in data from young
clastic settings, such as the South China Sea. They degrade the seismic image but otherwise
are of little interest in exploration.
Frequency shadows must be judged skeptically. A zone of thicker reflection spacing
produces low frequencies that sometimes resemble an anomaly. Such anomalies are often
observed at anticlinal crests (Taner et al., 1979; Figure 7). If the thicker reflection spacing
is due to velocity pull-down beneath a gas reservoir, then the frequency anomaly indica-
tes hydrocarbons, though it is unrelated to attenuation. Data problems also introduce fre-
quency anomalies, as do changes in the seismic signal-to-noise ratio caused by changes in
stratigraphy.
Few attributes are effective at identifying attenuation anomalies other than the obvious
anomalies caused by shallow gas. Instantaneous frequency is characterized by strong and
Figure 7. A low-frequency anomaly beneath a strong amplitude anomaly at the crest of a

faulted anticline, suggesting the presence of gas. (a) Seismic line. (b) Reflection strength.
(c) Average frequency derived in a 44 ms window (11 samples). Low frequencies are red,
high frequencies are blue.
rapid variations that obscure the small changes in the background trend that could indicate
a shadow. Average frequency is more stable, but variations due to interfering reflections
still dwarf the small changes due to attenuation.

A thick gas sand usually causes a strong reflection, so true low-frequency shadows
could be expected beneath bright spots. This attribute combination is a more reliable
direct hydrocarbon indicator than either attribute alone. Effects of attenuation on amplitude
and bandwidth are more subtle than on frequency. Applied in conjunction with frequency
attributes, they might modestly bolster a search for attenuation anomalies.
Poststack Q attributes are designed to quantify attenuation effects, such as low-
frequency shadows. Unfortunately, the same difficulties that undermine frequency and
amplitude attributes also undermine Q attributes. It is all but impossible to determine Q
from standard poststack seismic data with sufficient resolution to identify prospective
gas sands.
Volume spectral decomposition can reveal low-frequency anomalies (Castagna et al.,
2003). A simple workflow suffices. Create two frequency volumes, one low frequency
and one high frequency, by band-pass filtering followed by conversion to trace envelope.
On both frequency volumes, compare the response above and below possible attenuating
zones to spot marked differences. True attenuation anomalies exhibit a persistent relative
loss of high-frequency energy beneath an attenuating zone.
To reproduce the workflow with maps, apply spectral decomposition twice to create
two sets of frequency maps with identical scaling. Derive one set above the zone of interest,
and one set below. Interpret the maps in the same way as frequency volumes. This is less
satisfactory than the volume workflow because the persistence of an attenuation anomaly
is difficult to determine from maps.
Faults
Discontinuity is the workhorse seismic attribute for fault interpretation. Discontinuity
helps greatly when interpreting along horizons or horizontal slices, but it helps much less
on vertical sections. This is partly because faults observed on displays of poststack seismic
data are easier to recognize on vertical sections than they are on horizontal slices, whereas
the opposite is true of displays of discontinuity attributes. Converting the discontinuity
attribute to a fault attribute improves fault clarity and continuity on both horizontal and
vertical views.
Discontinuity attributes are designed to detect relatively vertical faults. They fail to
detect low-angle faults or the lower parts of listric faults, which are difficult to detect by
any means. Discontinuity should be computed along seismic reflections. Where enhanced
resolution is necessary, sharpen discontinuity through Laplacian filtering.
Other structural attributes also reveal faults. Most positive and most negative curva-
tures and horizontal relative amplitude change show fault details, whereas dip and seismic
shaded relief show general fault trends. Of these, only curvature has found much appli-
cation. Apply curvature to reveal detail but not for routine fault mapping because it
displaces faults. Most positive curvature tends to identify upthrown sides of faults,
whereas most negative curvature tends to identify downthrown sides (refer to Figure 9
in Chapter 1).
Avoid spectral decomposition and single trace attributes for fault mapping because
they detect only steep faults.
Figures 8 and 9 compare a variety of structural attributes along a time slice through
a faulted zone. Discontinuity suffices for most fault mapping, though Laplacian filtered
discontinuity is better for mapping detail. Relative amplitude change shows the same
faults as discontinuity. Being directional, it resembles apparent topography, which aids
intuition but also renders it less convenient for interpretation. Curvature attributes reveal
small features between faults that are not visible on discontinuity attributes. Here, the
high-resolution curvature attribute shows fine detail that looks like a network of tiny
Figure 8. Faults above igneous intrusions revealed by seismic attributes. (a) Original seismic time
slice (576 ms); the yellow box identifies the area shown in detail in Figure 8. (b) Dip-corrected
discontinuity. (c) Laplacian-filtered discontinuity. (d) Most negative curvature. Discontinuity
suffices for routine fault mapping, but Laplacian filtering and curvature attributes provide better
detail. Data from the Taranaki Basin, offshore New Zealand.
Figure 9. Close-up view of faults on the time slice (576 ms) of Figure 7. (a) Seismic data.
(b) Dip-corrected discontinuity. (c) Laplacian-filtered discontinuity. (d) Relative amplitude change
in the direction of the yellow arrow. (e) Relative amplitude change in a direction orthogonal to (d),
indicated by the yellow arrow. (f) Most positive curvature with moderate resolution. (g) Most
negative curvature with high resolution. (h) Most negative curvature with moderate resolution.
(i) Most negative curvature with low resolution. Data from the Taranaki Basin, offshore
New Zealand.
lineaments. The significance of the tiny lineaments is debatable. It is often suggested that
they are related to fractures, but this interpretation can be made only if supported by cor-
roborating evidence (Narhari et al., 2009). Positive curvature and negative curvature attri-
butes reveal different networks, which complicates their interpretation.
Figure 10. Faults on vertical sections are sometimes easier to follow with seismic shaded relief
than with discontinuity. (a) Seismic line with faults above an intrusion. (b) The seismic data
blended with a discontinuity attribute. (c) The seismic data blended with shaded relief; faults
appear sculpted into the data. Data from the Taranaki Basin, offshore New Zealand.
Figure 10 compares discontinuity and seismic shaded relief in a vertical view, both
blended with the original seismic data. Discontinuity captures fault details, but shaded
relief makes it easier to follow the vertical extent and continuity of the larger faults.
Channels
Channels and channel systems are much easier to track and interpret with seismic attri-
butes than with standard seismic data. A wide variety of attributes and methods reveal
channel geometry or character, including reflection strength, discontinuity, Laplacian
filtering, spectral decomposition, waveform classification, relative amplitude change,
dip, and seismic shaded relief. Spectral decomposition also provides rough estimates of
channel thickness. Because channels are confined to specific stratigraphic levels, they
are often best studied on attribute maps.
In younger basins, gas-prone sand channels are characterized by strong amplitude and
are readily mapped with reflection strength or sweetness (Cross et al., 2009; Ghosh et al.,
2010). Mud-filled channels tend to have weak amplitudes. In older basins, sand channels
are sometimes distinguished by weak amplitudes (Nestvold, 1996). In either case, dis-
continuity or relative amplitude change highlights the channel boundaries. Discontinuity
provides a cleaner image because it uses a longer operator, but relative amplitude
change offers better resolution. Blending an amplitude attribute with a structural attribute
combines channel brightness and geometry in one view (Figure 11).
Curvature attributes do not image channels as well as discontinuity or relative ampli-
tude change. Most positive curvature roughly shows channel sides, and most negative cur-
vature shows channel interiors (see Figure 9 in Chapter 1). Dip and seismic shaded relief
sometimes reveal the geometry of large well-defined channels better than other attributes.
Frequency and bandwidth attributes rarely image channels as well as amplitude attri-
butes. In contrast, spectral decomposition presented through red-green-blue (RGB) color
Figure 11. Channels highlighted by blending seismic attributes on a time slice (1444 ms)
through flat data. (a) Original seismic data. (b) Discontinuity computed in an 11-sample
window. (c) Relative amplitude change in the direction of the yellow arrow, smoothed with a
7-sample vertical median filter. (d) Sweetness; strong values are red, weak values are blue.
(e) Sweetness blended with bump-mapped discontinuity. (f) Strong sweetness values blended
with relative amplitude change. Data from southern Louisiana. After Barnes et al., 2011,
Figure 6, p. 37.
blending provides superior images of channels, as shown in Figure 12. This RGB-blended
display blends low-, intermediate-, and high-frequency tuning volumes and colors them
red, green, and blue, respectively. It reveals more detail in the channel systems than rms
amplitude.
Channels that approximate ideal thin beds with opposite polarity reflections have a
phase limited to –90 or +90 degrees, assuming the seismic data have zero phase.
Diapirs and gas chimneys

Diapirs in sedimentary basins are caused by mobile salt or clay, or by igneous intrusions.
Diapirs are important because they provide traps for hydrocarbon reservoirs on their flanks.
Thus the external shape of a diapir is important, not its internal character.
Diapirs observed on seismic data have steep sides and no internal reflections.
The steeper sides are difficult to image and often lack clear reflections. In the absence of
reflections, diapirs are inferred by changes in signal character. The seismic data from
within a diaper are dominated by noise with low amplitude, low frequency, limited band-
width, low continuity, and low parallelism.
Gas chimneys are caused by gas that leaks from a reservoir and migrates vertically to
shallower traps or to the surface. The gas escapes along leaky faults or through fractured
zones in reservoir seals. Gas chimneys introduce strong velocity contrasts that scatter seismic
energy and disrupt seismic imaging, thereby

reducing the continuity, amplitude, fre-
quency, and bandwidth of the seismic reflec-

tions. Gas clouds are similar to gas chimneys
but are much broader and disrupt large por-
tions of the seismic image. Gas chimneys
and clouds often occur above faulted anticli-
nes. Their significance in exploration is that
they commonly occur above commercial
reservoirs, and therefore serve as hydro-
carbon indicators (Heggland, 2013).
Diapirs and gas chimneys have nothing
in common geologically, but their seismic
images are much the same in that both
have relatively low amplitude, low con-
tinuity, low parallelism, low frequency,
limited bandwidth, and a vertical shape.
It is on this basis that similar workflows of
automatic pattern recognition are applied
to distinguish diapirs and gas chimneys
(Love and Simaan, 1984; Aminzadeh et al.,
2002; Nourollah et al., 2010; Figure 13). If Figure 12. (a) Root-mean-square amplitude
the pattern recognition distinguishes the along an interpreted horizon. (b) Volume
diapirs or gas chimneys distinctly, then spectral decomposition with RGB blending
geobody extraction can be applied to better on the same horizon. Low frequencies (5 to
define their geometry. 25 Hz) are red, mid frequencies (25 to 50 Hz)
are green, and high frequencies (50 to 100 Hz)
are blue. The RGB display complements
Geobody extraction rms amplitude by revealing detail in the
A geobody is a distinct feature in seis- channels. Data from the Taranaki Basin,
mic data that has geological or geophysical offshore New Zealand.
significance. Geobody extraction is an inter-
pretive process that defines a geobody according to a rule. The usual rule is that the
geobody is defined by a set of contiguous 3D sample locations, or “voxels,” whose data
values fall within a specified range (Figure 14). Bright spots, channels, diapirs, and gas
chimneys are commonly extracted as geobodies. Geobodies can be viewed and analyzed
independently of the seismic data like interpreted horizons or faults.
A feature can be extracted as a geobody using any attribute that distinguishes it
clearly. In practice, geobodies are almost always defined by strong amplitudes or by low
impedances. This is partly because bodies with strong amplitude or low impedance are
often prime exploration targets, but it is also because these attributes have proven to be
more successful than other attributes in geobody extraction.
Some features are better defined by two or more attributes than by a single attri-
bute. In this case, geobody extraction is improved by employing multiple attributes.
One way to accomplish this defines the geobody as the intersection of selected
Figure 13. Automatic pattern recognition applied to define an igneous diapir. (a) Seismic line
showing the diapir. (b) Reflection strength; strong values are red, weak are blue. (c) Parallelism;
parallel events are blue, nonparallel are red. (d) Average frequency; low frequencies are red,
high frequencies are blue. (e) Classification of the 3 attributes into 12 classes. Data from the
attribute ranges. Alternatively, the geo-

body values are defined as an ano-
maly on a crossplot of two attributes.
Yet another way is to create a facies
volume from automatic pattern recog-
nition of multiple attributes and ex-
tract geobodies defined by a particular
facies.
Figure 15 shows a well-defined
channel geobody extracted from a
reflection strength volume. In these
data, no other poststack attribute or
combination of attributes performs bet-
ter than reflection strength in geobody
extraction. This is typical. Figure 16
illustrates a diapir geobody extrac-
ted from the classified data volume of
Figure 13. The geobody is highly irre-
gular and so poorly defined that it has
little value. The result is also typical
Figure 14. Illustration of the process of geobody and illustrates the difficulty of extract-
extraction. (a) A vertical section through a seismic
ing geobodies from features that are
attribute represented as a grid of data values.
A user-selected initial starting location or “seed” poorly defined in amplitude.
for geobody extraction is circled in red. (b) The Geobody extraction is sometimes
geobody, starting from the seed, defined as the refined by combining amplitude with
set of all contiguous data locations whose values frequency or other attribute, as shown
equal 9. by Radovich and Oliveros (1998).
Figure 15. A bright channel extracted as a geobody from a reflection strength volume. Data from
the Taranaki Basin, offshore New Zealand.
Seismic attributes in data processing

Seismic attributes find limited but instructive employment in seismic data processing.
Chief applications include automatic gain correction, phase rotation, and coherency
filtering.
Automatic gain
Root-mean-square gain is a common method for automatically gaining seismic data
(Yilmaz, 2001a, p. 85). An rms-gained seismic trace x̃(t) is produced by dividing the
original seismic trace x(t) by its rms amplitude arms(t):
x(t)
x̃(t) = . (1)
arms (t)
The gain window is the window of the rms amplitude attribute. It is usually several
hundred milliseconds long, which is about ten times longer than typical windows in attri-
bute analysis. The strength of the gain is inversely proportional to the length of the window;
short windows produce strong gain, and long windows produce weak gain. Regardless of
window length or the initial trace values, the rms value of the gained trace tends toward
1. In this way, rms gain balances seismic traces.
Figure 16. An igneous diapir extracted as a geobody from the classified data volume of
Figure 13e. The extraction is poor, but it is difficult to do better. Data from the Taranaki Basin,
offshore New Zealand.
A seismic trace is also gained by dividing it by a smoothed version of its envelope. Like
rms gain, “envelope gain” balances the trace and has a gain strength that is inversely
proportional to the window length. The envelope of the gained trace tends toward 1.
A smoothed envelope ã(t) is obtained by filtering the original envelope a(t) with a smooth-
ing operator or low-pass filter w(t):
ã(t) = w(t) ∗ a(t). (2)
The gained trace x̃(t) derives from the seismic trace x(t) according to
x(t)
x̃(t) = . (3)
ã(t)
Figure 17 illustrates envelope gain applied to a seismic trace.
Phase rotation
A seismic trace x(t) can be rotated in phase by a constant uc by deriving its envelope
a(t) and instantaneous phase ui(t), adding uc to ui(t), and recombining the envelope
Figure 17. Envelope gain divides a seismic trace by a smoothed version of its envelope.
(a) Original trace envelope (red) and smoothed envelope (blue); the smoothing window is
200 ms long. (b) Original trace (red) and gained trace (blue). For comparison, the gained seismic
trace is scaled to have the same overall rms value as the original trace.
and modified phase to obtain the phase rotated trace x̃(t):
x̃(t) = a(t) cos[ui (t) + uc ]. (4)
This phase rotation is more efficient when derived directly from the seismic trace x(t) and
its quadrature trace y(t) according to
x̃(t) = cos uc · x(t) − sin uc · y(t). (5)
The Hilbert transform operator must be sufficiently long to rotate the lowest frequencies
in the data without distortion.
Structurally guided processes

“Structurally guided” processes proceed along seismic reflections. Such processes
employ slope attributes to follow the seismic reflections. The chief structurally guided
process is coherency filtering, which smooths seismic data along reflections. Standard
coherency filters smooth across faults. More sophisticated coherency filters incorporate
Figure 18. Coherency filtering of 3D seismic data characterized by faults, intrusions, and steep
dips. (a) Vertical view of the original data. (b) The data after coherency filtering that smooths
across faults. (c) The data after coherency filtering that preserves faults. Data from the Taranaki
Basin, offshore New Zealand.
nonlinear filters to preserve or enhance faults (Höcker and Fehmers, 2002; Figure 18).
Structurally guided processing is also applied in automatic full volume flattening and in
structurally guided interpolation of well log data (Lomask et al., 2006; Hale, 2010). In
seismic attribute analysis, structurally guided 3D filters are important for stratigraphic
analysis, such as in mode filtering of classified facies volumes and median filtering of
noisy attributes.
Conclusion
Seismic attributes are instrumental in many tasks of seismic data interpretation.
Attributes simplify seismic data analysis by subsetting the information in the data,
thereby enabling a “divide and conquer” approach to understanding the data. Attributes
are all but essential in amplitude mapping, anomaly identification, channel exploration,
and fault interpretation, and they accelerate data reconnaissance and clarify technical
presentations. Attributes even find limited but growing application in algorithms for
seismic data processing.
Seismic attribute analysis, however imperfect or informal, is a science, not an art.
Certainly there are elements of art in the design of color bars and opacity functions and
in the effective presentation of attributes. Nonetheless, attribute analysis proceeds from a
well-grounded scientific basis: the seismic signal can be separated into component attri-
butes that have different significance and that can be investigated independently.
Hundreds of seismic attributes have been invented. Making sense of them all is daunt-
ing, but unnecessary. Most seismic attributes are duplicates or flawed and can be discarded.
Discarding redundant and flawed attributes leaves a much smaller, more manageable,
and more comprehensible set of useful attributes. Useful attributes contain an appreciable
measure of unique information that records geological or geophysical properties. The two
most useful poststack seismic attributes are reflection strength and discontinuity. Other
useful attributes and methods include spectral decomposition, relative acoustic impedance,
average frequency, instantaneous phase, waveform maps, most positive and most negative
curvatures, and seismic shaded relief.
Seismic attributes are more interpretable when derived from clean seismic data. This is
true of attributes in general and of structural attributes in particular, such as discontinuity.
Seismic data are readily cleaned by coherency filtering, which should be routine in standard
attribute workflows.
Unambiguous low-frequency shadows due to absorption are rare. Frequency anomalies
are usually due to other factors, such as reflection interference, strong scattering, velocity
pull-down effects, uneven data acquisition, errors in data processing, or marked differences
in subsurface reflectivity that change the signal-to-noise ratio. As a consequence, frequency
anomalies are unreliable hydrocarbon indicators. Suspected attenuation anomalies must be
interpreted cautiously and supported by additional evidence.
Automatic fault interpretation is advancing slowly but steadily. It is likely that some-
day it will rival automatic horizon extraction in importance and utility. Challenging
problems remain, such as reliably detecting low-angle faults or ensuring that extracted
faults are geologically reasonable.
Geobody extraction is a powerful tool for visualizing complex geological features
in 3D. However, it is only modestly successful with current methods. It is most successful
with features characterized by strong amplitude, such as bright channels, but it fares poorly
with features that have weak amplitude, such as diapirs. Geobody extraction awaits new
attributes and pattern recognition methods to distinguish a greater variety of geological
features robustly.
First introduced some 40 years ago as colorful curiosities of unproven value, seismic
attributes have evolved to become indispensible tools for petroleum exploration. Nearly
all aspects of modern seismic data analysis benefit from attributes. The ever-growing
size and complexity of seismic interpretation projects ensure that seismic attribute analysis
will continue to grow in importance, scope, and sophistication in the years to come.

Appendix A
Analysis Windows
Many seismic attributes are derived within short symmetric analysis windows. The size
and shape of the analysis window determines the attribute resolution. Windows act
like smoothing filters and should be designed to avoid Gibbs’ phenomenon and other
artifacts. For 1D time attributes, Gibbs’ phenomenon is manifested by ringing in the
frequency domain, and sometimes also by conspicuous banding in the time domain.
Windows with abrupt edges produce Gibbs’ phenomenon, but tapered windows reduce
it. Prefer tapered windows.
Trace attributes are computed down traces and require 1D windows. Volume attributes
are computed both down and across traces and require 3D windows. The design of 3D
windows is similar to that of 1D windows.
The simplest window is the boxcar or rectangular window, which has no tapering.
The boxcar window is widely applied in attribute computations even though it strongly
introduces Gibbs’ phenomenon.
A 1D boxcar window in time wb(t) of length Tw is defined as
⎧
⎪ Tw
⎪
⎨ 1 |t| ≤ 2
wb (t) = . (A-1)
⎪
⎪
⎩ 0 |t| . Tw
2
The Fourier transform of the boxcar window, Wb( f ), is the sinc function
sin (pf Tw )
Wb ( f ) = . (A-2)
pf
The amplitude spectrum is the absolute value of Wb( f ). It decays slowly and rings with a
period of 1/Tw Hz; this ringing is Gibbs’ phenomenon. Figure A-1 illustrates a boxcar
window and its amplitude spectrum. The 50-ms length causes ringing with a period of
20 Hz.
Gibbs’ phenomenon is reduced by windows that peak at their centers and taper
toward their edges, such as the triangular, Hanning, Hamming, Blackman, and Gaussian
windows (Stanley, 1975). The Hamming or raised cosine window is common in seismic
attribute design. A Hamming window wh(t) of length Tw is defined by
⎧
⎪
⎪ 0.54 + 0.46 cos
2pt
|t| ≤
Tw
⎨
Tw 2
wh (t) = . (A-3)
⎪
⎪ T
⎩0 |t| .
w
2
201
a)
2
Amplitude 1
0
–30 –20 –10 0 10 20 30
Time (ms)
b)
0
–20
Amplitude (dB)
–40
–60
–80
0 20 40 60 80 100 120
Frequency (Hz)
Figure A-1. (a) A 50-ms boxcar window and (b) its amplitude spectrum.
The Fourier transform of the Hamming window, Wh( f ), is

sin (pf Tw ) 0.54 − 0.08 f 2 Tw2
Wh (f ) = . (A-4)
pf 1 − f 2 Tw2
Figure A-2 illustrates the Hamming window with a length of 50 ms. The first notch in the
amplitude spectrum occurs at twice the frequency of the first notch in the spectrum of the
boxcar window, and the spectral ringing is greatly reduced.
The Gaussian window has several theoretical and practical advantages over other
windows and has been gaining favor in recent years. However, in most attribute compu-
tations it makes little difference which window is used as long as it is tapered.
Figure A-3 illustrates the importance of windowing by comparing a boxcar window
and a Hamming window in the computation of a typical seismic attribute, energy half-
time. The attribute is sharper and clearer when derived with a Hamming window than
with a boxcar window. In the frequency domain, the boxcar window produces a ringy
power spectrum, whereas the Hamming window produces a smooth spectrum.
Appendix A: Analysis Windows 203
a)
1
Amplitude
0.5
0
–30 –20 –10 0 10 20 30
Time (ms)
b) 0
–20
Amplitude (dB)
–40
–60
–80
0 20 40 60 80 100 120
Frequency (Hz)
Figure A-2. (a) A 50-ms Hamming window and (b) its Fourier transform.
Three-dimensional attributes require 3D windows developed like 1D windows. The 3D

Hamming window wh(r) is

p r
0.54 + 0.46 cos r≤R
wh (r) = R , (A-5)
0 r.R
where r is the distance from the center of the window to a point in the window, and R is
the distance from the center to a corner of the window. For a window centered at the
origin, the distance r for a point with sample indices x, y, and t is

r= x2 + y2 + t2 . (A-6)
The values in a 3D window along a traverse through the window center match those of a 1D
window.
a)
Ringing spectrum
b)
Smooth spectrum
Figure A-3. Energy half-time computed with (a) a boxcar window and (b) a Hamming window.
Both windows are 60 ms long (15 samples). Hamming windows prevent spectral ringing and
produce clearer images.
The direct cost of applying a tapered window to seismic data is inconsequential,

but indirect costs can be large. Many recursive algorithms rely implicitly on boxcar
windows to achieve efficiency. Employing tapered windows requires a different
approach, which often is much slower. This is particularly true of 3D seismic attributes
such as discontinuity and reflection dip. One remedy is to employ a 3D Gaussian
window, which can be applied more efficiently as three separate 1D Gaussian windows.
More commonly, the tapered window is discarded altogether for the sake of computational
efficiency.
Figure A-4 shows a vertical view of a discontinuity attribute, computed first with
a boxcar window, then with a window tapered in time, and finally with a 3D tapered
window. The small improvement due to windowing may not be worth the great increase
in computation time. This is often true of discontinuity attributes because they are
usually interpreted along time slices where the detrimental effects of Gibbs’ phenomenon
are less noticeable.
Appendix A: Analysis Windows 205
Figure A-4. Comparison of seismic discontinuity computed in three different windows of size
5 lines by 5 traces by 15 samples (60 ms). (a) Boxcar window. (b) 1D Hamming window.
(c) 3D Hamming window.

Appendix B
Hilbert Transform
The Hilbert transform operator or quadrature filter h(t) is an ideal phase rotator that
subtracts 908 of phase from a seismic trace x(t) to produce a quadrature trace y(t).
Hilbert transformation can be represented as the convolution,
y(t) = h(t) ∗ x(t). (B-1)
The Hilbert transform of a cosine wave with amplitude a, nonzero frequency f, and phase u,
is a sine wave of the same amplitude, frequency, and phase. Thus given
x(t) = a cos (2pft + u), then (B-2)
y(t) = h(t) ∗ a cos (2pft + u) = a sin (2pft + u). (B-3)
Because an arbitrary waveform can be decomposed as a sum of sinusoids, equation B-3

uniquely defines the Hilbert transform for any waveform (Vakman, 1998, p. 10). Some
authors define the Hilbert transform so that it adds 908 phase (Claerbout, 1985, p. 20).
This reverses the sign of the quadrature trace, which must be taken into account in appli-
cation. The frequency of the cosine wave cannot be zero because that implies a constant
valued function, for which phase rotation has no meaning. In practice, Hilbert transform-
ation sets the dc component of a seismic trace to zero.
Transforming equation B-3 into the frequency domain, noting that it holds for all
frequencies except zero, the Fourier transform of the Hilbert transform operator H( f ) is
found to be
⎧
⎨ −i f . 0
H( f ) = −isgn( f ) = 0 f =0. (B-4)
⎩
+i f , 0
Inverse transforming H( f ) to the time domain yields the continuous Hilbert transform
operator in time,
1
t=0
h(t) = pt . (B-5)
0 t=0
Figure B-1 illustrates the Hilbert transform operator and its Fourier transform.
207
a) 200
100
Amplitude 0
–100
–200
–50 –25 0 25 50
Time (ms)
b) Im[H(f )]
+1
–f f
–1
Figure B-1. (a) The continuous Hilbert transform operator in time. (b) The imaginary part of the
Fourier transform of the Hilbert transform operator H( f ). The real part of H( f ) is zero.
The discrete Hilbert transform operator as a function of sample index n, h(n) is inde-
pendent of the time sample period; let it have a unit sample period and a Nyquist frequency
of 1/2. To find h(n), zero the spectrum H( f ) beyond the Nyquist frequency on both the
positive and negative axes, inverse Fourier transform, and sample the resultant operator
with unit sample period to obtain
2
n odd
h(n) = np (B-6)
0 n even or 0
(Figure B-2). This operator drops off slowly with n and so requires a long length, though
half its samples are 0 and can be ignored. Other discrete Hilbert transform operators are
designed to have optimal properties in some theoretical sense (Parks and Burrus, 1987).
Hilbert transformation in the frequency domain derives the quadrature trace as
the inverse Fourier transform of the product of H( f ) with the Fourier transform of
the seismic trace. This is seen readily by taking the Fourier transform of equation B-1,
Appendix B: Hilbert Transform 209
h(n)
2
π
–10 –8 –6 –4 –2 0
–n n
2 4 6 8 10
2
–
π
Figure B-2. The central 21 points of the discrete Hilbert transform operator h(n).
letting X( f ) and Y( f ) be the Fourier transforms of the seismic and quadrature traces x(t)
and y(t):
F
y(t) = h(t) ∗ x(t) ⇔ Y( f ) = H( f )X( f ). (B-7)
Alternatively, the complex trace is derived from the real trace in the frequency domain.
The complex trace z(t) can be expressed as the convolution of the complex delta function
dc(t) with the real trace x(t),
z(t) = dc (t) ∗ x(t), (B-8)
where dc(t) is defined in terms of the delta function d(t) and the Hilbert transform operator
h(t) as
dc (t) = d(t) + ih(t). (B-9)
The Fourier transform of the complex delta function is twice the unit step function in fre-
quency U( f ):
F
dc (t) ⇔ 2U( f ). (B-10)
Hence the Fourier transform Z( f ) of the complex trace z(t) is

⎧
⎨ 2X( f ) f . 0
Z( f ) = 2U( f )X( f ) = X( f ) f =0 (B-11)
⎩
0 f ,0
(Poularikas and Seely, 1985, p. 217). The complex trace z(t) follows as the inverse Fourier
transform of Z( f ). In words, equation B-11 implicitly performs Hilbert transformation in
the frequency domain by transforming the seismic trace to the frequency domain,
zeroing the negative frequency half of the spectrum, leaving the dc component unchanged,
doubling the positive frequency half, and inverse transforming back to the time domain to
obtain the complex seismic trace.

Appendix C
Derivative Filter
Discrete differentiation is accomplished through derivative filters or difference equa-

tions. Both are common in seismic attribute computations. Difference equations are
more efficient and easier to implement, but derivative filters provide better results and
are sometimes preferable.
The derivative filter is derived much like the discrete Hilbert transform operator.
The Fourier transform D( f ) of an ideal differentiator is
D( f ) = i2pf (C-1)
(Figure C-1). To derive the discrete derivative filter, zero the spectrum D( f ) beyond
a frequency fN on both the positive and negative axes and inverse Fourier transform to
ˆ
obtain a band-limited continuous derivative filter, d(t):

fN
ˆ = i2p
d(t) f exp (i2pft)df . (C-2)
−fN
ˆ ¼ 0. Equation C-2 integrates to

Note d(0)
ˆ = 2fN cos(2pfN t) − sin (2pfN t) .

d(t) (C-3)
t pt2
ˆ with unit sample period, set fN to
To obtain the discrete derivative filter d(n), sample d(t)
the Nyquist frequency, which is 1/2, and generalize to arbitrary time sample period by
dividing by period T:
⎧ n
⎨ (−1) n = 0 Im[D(f )]
d(n) = nT . (C-4)
⎩ π
0 n=0 pe
=2
Slo
Like the discrete Hilbert transform –f f
operator, the derivative operator is
infinitely long and its magnitude
decays slowly with index n
(Figure C-2).
In seismic attribute computations, Figure C-1. The imaginary part of the Fourier
two-point difference equations are transform D( f ) of an ideal differentiator. The real
often substituted in place of derivative part of D( f ) is zero.
211
d(n)
–n n
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6 7 8 9 10
–1
Figure C-2. The central 21 points of the discrete derivative operator d(n) for unit sample period.
filters. The derivative of a seismic trace x(t) is approximated by a two-point forward differ-
ence over a sample period T according to
dx(t) x(t + T) − x(t)

≈ . (C-5)
dt T
This operation shifts the result backward a half sample. The backward difference operation,
dx(t) x(t) − x(t − T)

≈ , (C-6)
dt T
shifts the result forward a half sample.

The amplitude spectrum of a forward difference operator is the same as that of a
backward difference operator. It is found as follows. Represent the backward difference
operator as the z transform D(z), given by
1−z
D(z) = (C-7)
T
Claerbout (1992, p. 9). Substitute exp(i2pf T) for z to obtain the corresponding frequency
spectrum D( f ),
1 − ei2pf T
D( f ) = . (C-8)
T
Appendix C: Derivative Filter 213
The magnitude of D( f ) is the 125

amplitude spectrum of the differ-
ence operator, AD( f ), 100
Amplitude (x 2π)
|sin(pf T)| 75
AD ( f ) = 2 . (C-9)
T
50
Figure C-3 compares the
25
amplitude spectra of a difference
operator and an ideal differen-
0
tiator. They match closely at low 0 25 50 75 100 125
frequencies but diverge at high Frequency (Hz)
frequencies.
Figure C-3. Amplitude spectra of the ideal
differentiator (red) and the forward difference operator
(gray), for a sample period of 4 ms.

Appendix D
Discrete Formulas for Approximating

Instantaneous Frequency and Relative
Amplitude Change
Taner and Sheriff’s (1977) formula for instantaneous frequency, equation 10 in

Chapter 4, requires two derivatives, which are accomplished through filters as described
in Appendix C. Scheuer and Oldenburg’s (1988) formula replaces the derivatives with
an efficient difference approximation. It is derived as follows.
Instantaneous frequency fi(t) is the time derivative of the instantaneous phase ui(t)
scaled to units of hertz. Replacing the derivative with a two-point forward difference
over time sample period T yields the approximate formula for instantaneous fre-
quency,
1 dui (t) ui (t + T) − ui (t)

fi (t) = ≈ . (D-1)
2p dt 2pT
Expressing instantaneous phase as the argument of the complex seismic trace z(t),

arg z(t + T) − arg z(t) 1 z(t + T)
fi (t) ≈ = arg . (D-2)
2p T 2pT z(t)
In terms of the seismic and quadrature traces x(t) and y(t), this becomes

1 x(t)y(t + T) − x(t + T)y(t)
fi (t) ≈ arctan . (D-3)
2pT x(t)x(t + T) + y(t)y(t + T)
This is Scheuer and Oldenburg’s approximation for instantaneous frequency. It is con-

strained to yield a maximum value of Nyquist frequency, in contrast to Taner and
Sheriff’s formula, which is unconstrained.
A discrete approximation for relative amplitude change s(t) is derived as follows.
By equation 11 in Chapter 4, relative amplitude change is the derivative of the logarithm
of the instantaneous amplitude a(t). It can be expressed in terms of the instantaneous power
p(t), or amplitude squared, as
d ln a(t) a′ (t) p′ (t)

s(t) = = = . (D-4)
dt a(t) 2p(t)
215
This is approximated by
1 p(t + T) − p(t)
s(t) ≈ . (D-5)
T p(t + T) + p(t)
Claerbout (1985, p. 20; see also Yilmaz, 2001b, p. 1907) offers an alternative discrete
approximation for instantaneous frequency,
′
1 z (t) 1 z(t + T) − z(t)
fi (t) = Im ≈ Im . (D-6)
2p z(t) pT z(t + T) + z(t)
Figure D-1. A comparison of instantaneous frequency computed through the standard formula,
given by Taner and Sheriff (1977), and the approximations given by Scheuer and Oldenburg
(1988) and by Claerbout (1985). Time shifts have been removed from the approximations
through 2-point averaging. The sample period is 2 ms, so the Nyquist frequency is 250 Hz,
which is well above the frequency content of these data. The three curves are nearly identical
except at spikes. Other differences are evident only at large display scales.
Appendix D: Discrete Formulas for Approximating Instantaneous Frequency 217
Figure D-2. Instantaneous frequency of a 5-second linear sweep from 10 to 110 Hz computed
with Scheuer and Oldenburg’s (1988) formula (red line), and with Claerbout’s (1985) formula
(blue line). The sweep (not shown) has a 4-ms sample period, for which Nyquist frequency is
125 Hz. Scheuer and Oldenburg’s formula accurately tracks the instantaneous frequency, but
Claerbout’s formula consistently overestimates it, more so at higher frequencies.
In terms of the seismic and quadrature traces, this becomes

2 x(t)y(t + T) − x(t + T)y(t)
fi (t) ≈ . (D-7)
pT (x(t) + x(t + T))2 + (y(t) + y(t + T))2
These difference formulas introduce a half-sample shift upward in time. The shift is
sometimes ignored, but it can be removed by a running 2-sample average.
Both Scheuer and Oldenburg’s and Claerbout’s frequency formula closely approxi-
mate Taner and Sheriff’s formula for frequencies below one-quarter Nyquist frequency
(Figure D-1). For frequencies above one-half Nyquist frequency, Claerbout’s approxi-
mation overestimates instantaneous frequency unacceptably (Figure D-2).
Formulas for frequency and relative amplitude change yield wavenumber by replacing
the time variable with a spatial variable.

Appendix E
Vector Traces
The mathematical development of discontinuity and waveform attributes is simplified

by treating seismic traces as vectors. This approach aids comprehension because vector
operations can be illustrated with intuitive graphs.
Let xi represent the ith seismic trace in a set of traces, each of which have N samples.
Trace xi is written as a column vector or as the transpose of a row vector according to
⎡ ⎤
x1i
⎢ x2i ⎥
⎢ ⎥
xi = ⎢ . ⎥ = [x1i x2i . . . xNi ]T , (E-1)
⎣ .. ⎦
xNi
where the superscript T denotes transpose. A trace is represented as a position vector in

“sample space.” A sample space of dimension N has N orthogonal axes that correspond
to trace sample indices. Figures E-1 and E-2 illustrate the idea of a 2D sample space.
Figure E-1 shows the traces within an analysis window that encompasses the first two
samples. Cross-plotting the first sample against the second sample of each trace yields
the graph of position vectors in 2D sample space shown in Figure E-2. The vector directions
quantify the shapes of the trace segments. The vector magnitudes, or vector lengths, quan-
tify their overall amplitude levels. This concept extends to traces with any number of
samples.
The dot product of two traces xi and xj is

N
xi · xj = xTi xj = xki x kj , (E-2)
k=1
where k is the trace sample index. The dot product of a trace xi with itself is the trace
energy Ei, and equals the square of the vector magnitude |xi|:

N
Ei = xi · xi = xTi xi = |xi |2 = x2ki . (E-3)
k=1
A unit vector, or direction vector, has a length of 1 and is denoted by the “hat” symbol.
Unit vector x̂i is defined by
xi
x̂i = . (E-4)
|xi |
219
Figure E-1. Five seismic traces. The dashed rectangle defines an analysis window of 5 traces
by 2 samples.
Unit vectors describe trace shapes. The dot

product of two unit vectors x̂i and x̂j is the
cosine of the angle u between them,
cos u = x̂i · x̂j . (E-5)
It is a measure of how much traces xi and xj

look alike.
Let mi be the expected or average value
of trace xi. The variance s 2i of a seismic
trace xi is the square of its standard devi-
ation si, given by
Figure E-2. The seismic data from the analysis 1 N

s 2i = (xki − mi )2 . (E-6)
window of Figure E-1 displayed as vectors in 2D N k=1
sample space, where S1 refers to the first sample
of each trace and S2 refers to the second sample. The covariance sij of traces xi and xj is
1 N
sij = (xki − mi )(xkj − mj ). (E-7)
N k=1
The covariance normalized by the standard deviations of the two traces is the correlation
coefficient cij,
sij
cij = . (E-8)
si sj
The correlation coefficient is independent of differences in amplitude scale and responds
only to differences in trace shape. If the trace means are zero, the covariance equals the
dot product of the traces and the correlation coefficient equals the dot product of their
unit vectors.
Appendix E: Vector Traces 221
Let X be a set of M discrete seismic traces: x1, x2, . . . , xM. Represented as a column
vector,
⎡ ⎤T
x1
⎢ x2 ⎥
⎢ ⎥
X = ⎢ . ⎥ = [x1 x2 . . . xM ]. (E-9)
⎣ .. ⎦
xM
If each trace in X has N samples, then X forms an N by M matrix of samples:

⎡ ⎤
x11 x12 . . . x1M
⎢ x21 x22 . . . x2M ⎥
⎢ ⎥
X=⎢ . .. .. .. ⎥. (E-10)
⎣ .. . . . ⎦
xN1 xN2 . . . xNM
The columns of X are traces and the rows are samples, or horizontal slices. The
average trace xa of the set X is the vector average
1 M
xa = xi . (E-11)
M i=1
The kth sample of xa, xka, is the average value of the kth horizontal slice in X:
1 M
xka = xki . (E-12)
M i=1
The total energy E of the traces X is the sum of the individual trace energies Ei,

M
E= Ei . (E-13)
i=1
By the Schwartz inequality, the total energy E is always greater than or equal to the number
of traces M times the energy of the average trace Ea:
E ≥ MEa . (E-14)
The vector cross product of traces xi and xj is the N by N matrix,

⎡ ⎤
x1i x1j x1i xj2 . . . x1i xNj
⎢ x2i x1j x2i x2j . . . x2i xNj ⎥
⎢ ⎥
xi × xj = xi xj = ⎢
T
⎢ .. .. .. .. ⎥ ⎥. (E-15)
⎣ . . . . ⎦
xNi x1j xNi x2j . . . xNi x Nj
The covariance matrix C of the traces in the set X is the M by M matrix

⎡ ⎤
s 21 s12 ··· s1M

⎢s s 22 ··· s2M ⎥
⎢ 21 ⎥
C=⎢
⎢ .. .. .. .. ⎥.
⎥ (E-16)
⎣ . . . . ⎦
sM1 sM2 ··· s 2M
Covariance discontinuity employs this matrix.

Principal component analysis of seismic attributes is accomplished with a covariance
matrix that replaces traces with attribute vectors. If the M attribute vectors xj of length N
have zero mean, as is common, then the covariance matrix is an N by N matrix that is
the sum of the cross products of all the vectors:

M
C = XXT = xj xTj . (E-17)
j=1
Glossary
acoustic impedance: A rock property equal to the product of the rock density with its
compressional-wave velocity. Acoustic impedance is the resistance offered to propagating
compressional waves. Acoustic impedance serves as a rough inverse indicator of porosity
in young clastic sequences.
amplitude: The magnitude of the seismic trace values. Amplitude is the most important
seismic property and is quantified by many attributes including reflection strength, rms
amplitude, and average absolute amplitude.
amplitude acceleration: The second derivative of the logarithm of the instantaneous
amplitude. It reveals discontinuities and fine details but appears noisy.
apparent polarity: The sign of the trace at envelope maxima, held constant in each
interval bounded by envelope minima, and scaled by the reflection strength. Apparent
polarity is a response attribute.
arc length: Arc length L is the total length of the wiggles of a waveform, and is approxi-
mated by
1 N−1
L= |xn+1 − xn |, (G-1)
N − 1 n=1
where xn is the nth data sample, and N is the number of samples. This idea only makes
sense on a wiggle-trace display. Arc length is driven by amplitude and frequency.
attenuation: The progressive loss of high frequencies with time in propagating seismic
waves. Attenuation is caused by absorption or scattering of seismic energy as it passes
through the earth. The degree to which rock absorbs seismic energy is inversely quantified
by the quality factor. Attenuation is difficult to quantify with poststack seismic data.
attribute: See seismic attribute.
attribute space: In multiattribute analysis, a coordinate space defined by a set of seismic
attributes, each of which defines an axis in the space.
average frequency: The average Fourier spectral frequency weighted by the amplitude or
power spectrum; the average instantaneous frequency weighted by the instantaneous
amplitude or power. Average frequency attributes are derived in a short running interval
and are applied to detect spectral changes or to characterize the reflection spacing.
azimuth: The downdip direction of a 3D reflection surface.
223
bandwidth: The breadth of an amplitude spectrum. Often quantified as the standard

deviation of the frequency spectrum about the mean frequency.
broadband: A bandwidth of more than about 2 octaves, typically 3 octaves or more.

classification: A method of automatic pattern recognition that assigns each data value or
data vector to one of a small set of template classes. Classification is applied to identify
regions in seismic data that are characterized by similar attribute values.
coefficient of variation: A statistical measure C of strictly positive data, defined as the
standard deviation s divided by the mean m:
s
C= . (G-2)
m
The coefficient of variation acts as a relative standard deviation. It is applied in amplitude
variance attributes that are comparable between different seismic data sets.
coherence: The degree of consistency in the amplitude and phase along seismic reflections;
the degree to which neighboring traces are similar. In attribute analysis, coherence is
treated as synonymous with continuity and similarity. See discontinuity.
complex seismic trace: A seismic trace z(t) whose values are complex and formed by the
combination of a seismic trace x(t) and its quadrature trace y(t) according to
z(t) = x(t) + iy(t). (G-3)
In polar form, the complex trace is expressed in terms of reflection strength a(t) and instan-
taneous phase ui(t):
z(t) = a(t) exp[iui (t)]. (G-4)
The complex trace is the basis of complex seismic trace analysis.

complex seismic trace analysis: An important method for generating seismic attributes.
Complex seismic trace analysis extracts the amplitude and phase information of a
seismic trace as separate attributes from which other attributes are derived through differ-
entiation or averaging.
continuity: A measure of the similarity of the amplitude and phase along a seismic reflec-
tion; the degree of similarity between neighboring traces. Continuity is usually treated as
the opposite of discontinuity, and as synonymous with coherence. Continuity attributes
reveal faults and channels.
correlation: A method that quantifies the degree of similarity between two traces or
signals. Normalized correlations produce correlation coefficients. Correlation is employed
in several discontinuity attributes.
correlation coefficient: A normalized measure of the similarity of two traces defined as the
value of the zero-lag correlation between two zero-mean traces normalized by the square
Glossary 225
root of the product of their energies. It is equal to the covariance of the traces normalized
by their standard deviations. Correlation coefficient C between traces x1 and x2 with N
samples is
N
x1 · x2 n=1 x1n x2n
C= =
N 2
N 2 . (G-5)
|x1 | · |x2 |
n=1 x1n · n=1 x2n
The correlation coefficient is sensitive only to the shape of the trace waveforms and is
independent of overall trace magnitudes. Its value varies from 21 to +1. A value of 1 indi-
cates the two traces have identical shape, 0 indicates they are completely unrelated, and 21
indicates their shapes differ only in polarity.
cosine of the phase: A complex seismic trace attribute equal to the cosine of the instan-
taneous phase. Cosine of the phase acts as the perfect automatic gain control because
it removes all amplitude information from a trace, making reflection continuity easier to
follow.
covariance: A nonnormalized measure of the similarity of two traces. The covariance sij
of traces xi and xj, with N samples, and with means mi and mj, is defined as
1 N
sij = (xik − mi )(x jk − mj ). (G-6)
N k=1
Covariance is used in principal component analysis and in some attribute computations.
curvature: The rate of change of dip and azimuth on a reflection surface. Curvature attri-
butes reveal finer detail than dip or azimuth but appear noisier. There are many curvature
attributes, but only most positive and most negative curvatures are applied widely. Strong
curvature suggests where fractures are most likely to occur.
dip: The angle that a planar reflection makes with the horizontal. It is sometimes called
dip magnitude. Slope is the tangent of the dip, but the distinction is often ignored and
slope is referred to as dip.
dip-azimuth: Reflection dip and reflection azimuth combined so that azimuth controls
color and dip controls shading. Slope is often employed in place of dip. Dip-azimuth
presents the same information as seismic shaded relief.
dip variance: The degree to which the dips along reflections vary from the average dip.
Used as a measure of reflection parallelism.
direct hydrocarbon indicator: A seismic attribute that is sensitive to effects in the seismic
data caused by hydrocarbons.
directional attribute: An attribute derived from a directional operator, usually a spatial
derivative. Directional attributes include apparent dip, seismic shaded relief, and
relative amplitude change. They resemble illuminated topography when displayed in
monochrome.
discontinuity: A measure of how much the amplitude and phase vary along seismic reflec-
tions; the degree of dissimilarity between neighboring traces. Discontinuity attributes
reveal faults and channels.

divergence: The degree to which succeeding reflections in a sequence diverge downdip.
Divergent reflections are characterized by constant azimuth and increasing dip with
depth. Divergence identifies sequence margins and the sides of large channels.
dominant frequency: An imprecise term, variously used to refer to the average frequency
of a signal, or to its largest or most significant frequency component.
edge detection: A method that highlights the outlines of features in an image. Edge detec-
tion is applied to interpreted seismic horizons to identify faults and places where dips
change abruptly, but it is generally unsuitable for application to standard seismic data.
effective bandwidth: An empirical measure of bandwidth derived from the autocorrelation
of the seismic data in an interval. Effective bandwidth is applied chiefly as a map attribute.
energy: The integral of the trace instantaneous power; the integral of the power spec-
trum of the trace. Sometimes referred to as total energy. As a map attribute, energy E is
defined by
N
E= x2n , (G-7)
n=1
where xn is the nth data sample, and N is the number of samples. In the context of a complex
trace with instantaneous amplitude a(t), energy is defined as

1
E= a2 (t)dt. (G-8)
−1
energy half-time: An interval attribute that records where in the interval the seismic
energy is concentrated. Computed as a trace attribute, energy half-time measures relative
amplitude changes.
envelope: An amplitude measure that envelops the seismic trace, often referred to as trace
envelope or as signal envelope. In common usage, envelope is synonymous with reflection
strength and instantaneous amplitude. See reflection strength.
filter bank: An ordered set of narrowband filters with different passbands. Spectral
decomposition employs filter banks.
first moment formula: A formula that equates the average instantaneous frequency of a
seismic trace in time to its average Fourier spectral frequency.
frequency: The number of cycles on a seismic waveform that occur within a period of
time. Frequency attributes include instantaneous frequency, average frequency, and
tuning frequency. They identify zones of anomalous seismic attenuation and serve as
rough measures of reflection spacing.
Glossary 227
Gabor wavelet: A wavelet formed as the product of a Gaussian window with a sinusoid.
Commonly used in filter banks for spectral decomposition. Gabor wavelets in a filter bank
have the same envelope, length, and bandwidth measured in hertz.

Gaussian window: An analysis window whose shape is a Gaussian function.
geobody: A distinct 3D feature, extracted from seismic data, that has geological or
geophysical significance. Geobodies are defined in terms of attribute values. Bright
spots, channels, diapirs, and gas chimneys are extracted as geobodies to aid their
interpretation.
geobody extraction: An interpretive process that sculpts geobodies from seismic data
given a definition of the geobodies in terms of their attribute values and spatial distribution.
group vector: The 3D gradient of the logarithm of the instantaneous amplitude. The
horizontal components of the group vector highlight faults and other discontinuities.
Hamming window: A type of tapered window of finite length with the shape of a cosine
function from 2908 to +908, raised by a small value. Sometimes referred to as a raised
cosine window.
Hilbert transform: The process of applying a quadrature filter that subtracts 908 of phase
from a seismic trace without changing its amplitude spectrum. Some authors define
the Hilbert transform so that it adds 908 of phase, which reverses the polarity of the
filter operator. See quadrature filter.
horizon: A surface that represents a stratigraphic level in a seismic line or volume.
Horizons tend to follow continuous reflections characterized by relatively consistent
phase and amplitude.
horizon attribute: An attribute of a seismic horizon. Horizon attributes are necessarily
structural and include dip, azimuth, curvature, and discontinuity.
imaginary trace: The imaginary part of a complex trace; synonymous with quadrature
trace. The imaginary trace is derived as the Hilbert transform of a seismic trace, or real
trace.
instantaneous: With regard to attribute computations, computed at a point instead of in a
window.
instantaneous amplitude: Synonymous with reflection strength and trace envelope. See
reflection strength.
instantaneous bandwidth: A complex seismic trace attribute b(t), defined as the absolute
value of the time derivative of the logarithm of the instantaneous amplitude a(t), scaled to
have units of hertz:

1 d
b(t) = ln a(t). (G-9)
2p dt
Instantaneous bandwidth is closely related to relative amplitude change in time.

instantaneous frequency: A complex seismic trace attribute fi(t), defined as the time
derivative of the instantaneous phase ui(t) scaled to have units of hertz:
1 d
fi (t) = ui (t). (G-10)
2p dt
Instantaneous frequency is highly variable and prone to spikes.

instantaneous phase: A complex seismic trace attribute defined as the argument of the
complex trace. Instantaneous phase has a sawtooth appearance because its values are
constrained to the range of +1808 to 21808.
instantaneous power: Envelope squared. See power.
instantaneous quality factor: A complex seismic trace attribute defined as instantaneous
frequency divided by twice the instantaneous bandwidth. Instantaneous quality factor is
dimensionless and unrelated to rock quality factor.
instantaneous rms frequency: A complex seismic trace attribute defined as the
square root of the sum of the squares of instantaneous frequency and instantaneous
bandwidth.
instantaneous wavenumber: A spatial derivative of the instantaneous phase or of the
logarithm of the instantaneous amplitude.
interval: A window with constant length in time or depth that follows a horizon in a
seismic volume; the region between two selected horizons. Intervals define the seismic
data to use for deriving map attributes.
interval attributes: Map attributes computed in a narrow horizon-guided interval through
a seismic volume; trace attributes computed in a window that runs down the trace. Common
interval attributes include rms amplitude, largest value, average frequency, and number
of peaks.
K-means clustering: A method of unsupervised pattern recognition that is sometimes
applied in waveform mapping and attribute classification. K-means clustering orders its
classes randomly.
Kohonen self-organizing feature map (Kohonen SOFM): A method of unsupervised
pattern recognition that is commonly applied in waveform mapping and attribute classifi-
cation. The Kohonen SOFM orders its classes by similarity.
kurtosis: The relative sharpness or flatness of a distribution relative to a Gaussian distri-
bution. Kurtosis is occasionally employed as an interval attribute.
Laplacian filter: A linear image processing filter that is applied along horizontal slices or
along reflections to sharpen lineations. Laplacian filters are chiefly applied to discontinuity
attributes to enhance faults and channels.
local: With regard to seismic attribute computations, computed in a small window instead
of at a point.
Glossary 229
low-frequency shadow: An anomalous drop in average frequency, observed on seismic

data, that is produced by locally strong signal attenuation. Gas is a primary cause of
seismic attenuation, so low-frequency shadows serve as possible direct hydrocarbon indi-

cators. Prospective low-frequency shadows are rare.
mean: The average of a set of data values, often called sample mean; the expected value of
a random variable. The distinction between sample mean and expected value is important
mathematically but often is ignored in attribute analysis.
median filter: A nonlinear filter that replaces the data value in the center of an
analysis window with the median of the data values in the window. Median filters are
applied to remove outliers or spikes from seismic attributes, such as instantaneous
frequency.
mode filter: A nonlinear filter designed for quantized data that replaces the data value in
the center of an analysis window with the mode of the data values in the window, which is
the value that occurs most often. Mode filters are applied to smooth classified waveform
maps or classified attribute volumes.
Morlet wavelet: A Gabor wavelet whose length is inversely proportional to its center
frequency, commonly used in filter banks for spectral decomposition. The Morlet wavelets
in a filter bank look the same except for scale, and they have the same length measured in
cycles and the same bandwidth measured in octaves.
narrowband: A bandwidth of less than about 2 octaves, typically 1 octave or less.
one-dimensional (1D) attribute: Trace attribute. One-dimensional attributes are com-
puted down individual seismic traces. Examples include reflection strength, rms amplitude,
instantaneous phase, average frequency, and relative acoustic impedance computed
through recursive inversion.
parallelism: A measure of how parallel reflections in a sequence are to each other.
Parseval’s theorem: A theorem that equates the energy of a seismic trace in the time
domain to its energy in the frequency domain.
peak frequency: The frequency of the spectral component that has the largest magnitude.
perigram: Reflection strength minus its low-frequency content.
phase: The relative position along a sinusoid; the average value of the phase spectrum of a
signal.
phase vector: The 3D gradient of the instantaneous phase. The phase vector is the basis of
much of 3D complex seismic trace analysis and provides measures of wavelength, dip, and
azimuth.
polarity: The sign of the seismic data with respect to a standard. Polarity is described as
being either normal or reverse. In seismic attribute analysis, it is convenient to define
normal polarity so that a positive reflection corresponds to a positive reflection coefficient,
which implies an increase in acoustic impedance.
power: The rate of change of energy. For a seismic trace, power refers to the square of
the original trace values, or to the square of the envelope (see instantaneous power). In
the frequency domain, power refers to the square of the amplitude spectrum. In either
domain, the integral of the power is the energy.
power spectrum: The square of the amplitude spectrum.
principal component analysis: A general least-squares method to find linear relationships
between the various components of multidimensional data. Principal component analysis is
applied in discontinuity attributes, in coherency filters, and in multiattribute analysis for
attribute space reduction.
principal components: Measures of data variance derived through principal component
analysis, arranged in order of largest to smallest. The first principal component of a set
of seismic traces measures the continuity of the reflections.
Q: See quality factor.
quadrature filter: A filter that subtracts 908 of phase from a seismic trace without altering
the amplitude spectrum. In common usage, synonymous with Hilbert transform.
quadrature trace: The Hilbert transform of a seismic trace; a seismic trace rotated in
phase 2908; synonymous with imaginary trace.
quality factor: A measure of the degree to which rocks pass acoustic energy without
attenuation. Often called Q, it is sensitive to the presence of gas. Quality factor attributes
quantify overall spectral change, but their relationship to rock quality factor is tenuous.
rank correlation: A nonlinear method for estimating how closely two signals or attributes
are related. The rank correlation coefficient is a more robust measure of similarity than the
correlation coefficient.
real trace: The real part of a complex trace; the original seismic trace.
recursive inversion: A recursive method for inverting a seismic trace to produce an esti-
mate of the relative acoustic impedance. Recursive inversion assumes that the trace
approximates a reflection coefficient series. It closely resembles an integration of the trace.
reflection orientation: The dip and azimuth of a reflection; a unit vector normal to a reflec-
tion surface.
reflection spacing: The distance between two successive reflections measured perpen-
dicularly to the reflections. Reflection spacing is difficult to quantify as an attribute. It is
estimated roughly by frequency or wavelength attributes.
reflection strength: A complex seismic trace attribute that measures seismic amplitude
independently of its phase or polarity. Reflection strength is defined as the magnitude of
the complex seismic trace and is the most important amplitude attribute. Synonymous
with instantaneous amplitude and with trace envelope.
reflection surface: A surface in a seismic volume that follows a seismic reflection and has
relatively consistent amplitude and phase. The concept of reflection surfaces underlies the
computation of many 3D seismic attributes.
Glossary 231
relative acoustic impedance: Acoustic impedance minus the background trend; acoustic
impedance derived only from poststack seismic data. Relative acoustic impedance is
restricted to the bandwidth of the seismic signal and lacks the low-frequency and high-
frequency components of a full acoustic impedance.
relative amplitude change: The rate of change of the logarithm of the instantaneous
amplitude in a given direction. Relative amplitude change in time s(t) is defined as
d a′ (t)
s(t) = ln a(t) = , (G-11)
dt a(t)
where a(t) is instantaneous amplitude.

response amplitude: A response attribute that records the values of the envelope peaks.
response attributes: A set of attributes derived from instantaneous complex seismic trace
attributes through selection at envelope peaks. Response attributes have a blocky appear-
ance and are free of spikes. They include amplitude, phase, frequency, apparent polarity,
and sweetness.
response frequency: A response attribute that records instantaneous frequencies at
envelope peaks. Response frequency equals an average spectral frequency of the seismic
wavelet if the reflections are free of noise and interference.
response phase: A response attribute that records instantaneous phases at envelope peaks.
Response phase equals the phase of the seismic wavelet if the reflections are free of noise
and interference.
RGB blending: Red-green-blue color blending. A method of volume blending whereby
three seismic data volumes are combined graphically so that one volume controls the inten-
sity of red, the second controls the intensity of green, and the third controls the intensity of
blue. Commonly employed in spectral decomposition with low-, mid-, and high-frequency
spectral volumes.
rms: Root-mean-square; a procedure for finding a representative value of a set of data
values. The rms value xrms of a seismic trace xn with N samples is

N
1
xrms =
x2 . (G-12)
N n=1 n
Like the average absolute value, the rms value is independent of the sign of the data values,
but it is more influenced by larger values.
rms amplitude: An amplitude attribute equal to the square root of the average trace energy.
Root-mean-square (rms) amplitude roughly resembles the trace envelope but is always
smaller.
rms spectral frequency: The square root of the average frequency squared of the power
spectrum.
seismic attribute: A measure of a seismic property. Seismic attributes represent subsets of

the information in seismic data.
seismic facies: A recognizable pattern in seismic data that has geological or geophysical
significance. Seismic facies are often characterized by seismic attributes.
seismic property: A characteristic of seismic reflection data. Key seismic properties
include amplitude, amplitude change, phase, frequency, dip, curvature, continuity, and
parallelism.
seismic shaded relief: A directional seismic attribute that combines reflection dip and
azimuth to make horizontal slices look like illuminated topography.
semblance: A measure of the degree of similarity between the traces in a set. Sem-
blance is defined as the energy of the average of the traces divided by the average of
the trace energies. Given M traces with total energy E, and average trace energy Ea, sem-
blance S is
MEa
S= . (G-13)
E
Semblance is employed in the computation of dip, azimuth, and discontinuity.
short time-window Fourier transform: See STFT.
signal length: The effective length of a signal or waveform in time, sometimes measured
as the standard deviation of the instantaneous power around the center of the signal.
similarity: The degree to which the traces in a set resemble each other. Similarity lacks a
precise definition and is sometimes used as a synonym for continuity.
skew: A nondimensional measure of the degree of asymmetry of a set of data values. Skew
is occasionally employed as an interval attribute.
slope: The tangent of the dip; the ratio of the change in depth or time of a reflection over a
horizontal distance. For depth data, slope is dimensionless. For time data, slope has units of
milliseconds per trace or meter. In common usage, slope is referred to as dip.
spectral decomposition: A method that decomposes a waveform into a set of sinu-
soidal components or wavelets with different center frequencies and bandwidths.
Spectral decomposition is accomplished through filter banks and is applied in tuning
analysis.
spectral ratio method: A method for estimating the quality factor from seismic data based
on changes in the amplitude spectra between consecutive windows on a trace.
STFT: Short time-window Fourier transform; a basic method of time-frequency analysis,
sometimes applied in spectral decomposition.
structurally guided: In reference to a seismic process, guided by the reflection slopes to
proceed along reflections. Also called dip guided. Coherency filters are structurally guided,
as are some discontinuity attributes, stratigraphic attributes, and 3D filters.
Glossary 233
sweetness: An empirical measure designed to highlight “sweet spots,” places that are oil
and gas prone. Sweetness is defined either as response amplitude divided by the square
root of response frequency, or more commonly as reflection strength divided by the

square root of instantaneous or average frequency.
thin bed: A rock layer, imaged by seismic data, whose thickness is small with respect to the
wavelengths of the seismic wavelet. Reflections from the upper and lower surfaces of a thin
bed interfere and are difficult to distinguish so that the thickness cannot be estimated
reliably.
thin-bed indicator: A complex seismic trace attribute defined as instantaneous frequency
minus average frequency. Despite its name, the thin-bed indicator does not indicate thin
beds.
three-dimensional (3D) attribute: Volume attribute. Three-dimensional attributes are
computed down traces as well as across traces in both the x and y directions. Examples
include discontinuity, dip, curvature, and parallelism.
time-frequency analysis: The study of local spectral properties as a function of time
for a signal or seismic trace. Key methods of time-frequency analysis include the
short time-window Fourier transform and the wavelet transform. Time-frequency analysis
is applied in spectral decomposition, Q estimation, and the computation of spectral
attributes.
tuning: The strong constructive or destructive interference of two or more reflected seismic
wavelets. Tuning depends on the frequency content of the wavelets and the spacing of the
reflectors.
tuning analysis: The study of the frequency content of interfering reflections. Tuning
analysis is applied to estimate thin-bed thicknesses.
tuning thickness: For a particular seismic wavelet, the thickness of a thin bed that pro-
duces the maximum peak amplitude or maximum total energy in the composite reflection.
two-dimensional (2D) attribute: Line attribute. Two-dimensional attributes are computed
both down and across traces along a seismic line. Examples include slope components,
directional dip, and horizontal relative amplitude changes. Horizon attributes are also
2D attributes, but are computed on a surface across inlines and crosslines.
uncertainty principle: In the context of signals, a fundamental property that states that
the product of signal length with signal bandwidth is greater than or equal to a constant.
The constant depends on the exact definitions of length tl and bandwidth fb; setting tl
and fb equal to standard deviations, the uncertainty principle becomes
tl · fb ≥ 1/4p. (G-14)
This implies that the length and bandwidth of a signal or waveform cannot both be made
arbitrarily small at the same time. The name uncertainty principle is misleading because it
quantifies a definite property of signals.
variance: A measure of the degree to which the values in a set vary from their mean.
The variance V of seismic trace xn with mean m is defined as
1 N
V= (xn − m)2 , (G-15)
N n=1
where xn is the nth data sample, and N is the number of samples. Variance is employed
in many attributes, including bandwidth, amplitude variance, dip variance, and
discontinuity.
volume blending: A method of computer graphics to display simultaneously two or more
seismic volumes by adjusting their opacity and overlaying them. Volume blending enables
different seismic volumes to be compared readily.
waveform: A segment of a seismic trace that encompasses one or several lobes. In seismic
attribute analysis, a waveform typically represents a seismic wavelet or a composite of
interfering reflections.
waveform classification: A method for creating attribute maps that identifies regions of
similar waveform along horizons. Waveform classification reveals details in channel
systems and other stratigraphic features.
wavelength: The length of one cycle of a sinusoidal waveform.
wavenumber: Inverse of wavelength. In attribute analysis, wavenumber is derived as the
rate of change of instantaneous phase along a spatial axis.
weighted average: The average of a set of values scaled by a set of weights. Weighted
averaging is employed in spectral attributes, discontinuity attributes, and average
complex trace attributes. In the time domain, the weighting function is usually instan-
taneous power; in the frequency domain, it is usually spectral power.
Widess limit of resolution: An empirical limit on how thin a rock layer can be
before seismic reflections from its top and bottom are no longer distinguishable.
For depth data, the Widess limit of resolution is one eighth of the dominant wavelength
of the seismic wavelet. For time data, it is more conveniently expressed as a two-way
time thickness tw equal to one quarter of the period of the average frequency fa of the
seismic data:
1
tw = . (G-16)
4 fa
The Widess limit of resolution is important in spectral decomposition and waveform

classification.
window: A small 1D or 3D operator for selecting data for attribute computations. Tapered
windows, such as Hamming and Gaussian windows, reduce Gibbs’ effects. The size and
shape of the window set the resolution of the attribute.
Glossary 235
zero-crossing frequency: A crude measure of average frequency fzc defined as half

the number of zero-crossings Nzc in an interval divided by the interval length in time tl:
Nzc
fzc = . (G-17)
2tl
Zero-crossing frequency is usually offered as a map attribute.

References
Abriel, W. L., 2008, Reservoir geophysics: Applications: SEG Distinguished Instructor

Series No. 11, http://dx.doi.org/10.1190/1.9781560801634.
Addy, S. K., 1997, Attribute analysis in Edwards limestone in Lavaca County, Texas: 67th
Annual International Meeting, SEG, Expanded Abstracts, 737 –740.
AlBinHassan, N. M., and K. J. Marfurt, 2003, Fault detection using Hough transforms: 73rd
Al-Dossary, S., and K. J. Marfurt, 2006, 3D volumetric multispectral estimates of reflector
curvature and rotation: Geophysics, 71, no. 5, P41 –P51, http://dx.doi.org/10.1190/
1.2242449.
Aminzadeh, F., D. Connolly, R. Heggland, P. Meldahl, and P. de Groot, 2002, Geohazard
detection and other applications of chimney cubes: The Leading Edge, 21, no. 7,
681– 685, http://dx.doi.org/10.1190/1.1497324.
Anstey, N., 1972, Seiscom ’72 (Seiscom Limited internal report).
Anstey, N., 1973a, Seiscom ’73 (Seiscom Limited internal report).
Anstey, N. A., 1973b, The significance of color displays in the direct detection of hydro-
carbons: 43rd Annual International Meeting, SEG, Expanded Abstracts.
Anstey, N. A., 1977, Seismic interpretation: The physical aspects: International Human
Resources Development Corp., http://dx.doi.org/10.1007/978-94-015-3924-1.
Anstey, N., 2001, Snapshots from a lifetime in geophysics, in A. McBarnet, ed., EAGE,
1951– 2001: Reflections on the first 50 years: Blackwell Science, 3– 10.
Anstey, N., 2005, Attributes in color: the early years: CSEG Recorder, 30, no. 3, 12– 15.
Ashbridge, J., C. Pryce, F. Coutel, and M. Welch, 2000, Fault and fracture prediction from
coherence data analysis. A case study: The Magnus Field, UKCS: 70th Annual Inter-
national Meeting, SEG, Expanded Abstracts, 1564– 1567.
Bacon, M., R. Simm, and T. Renshaw, 2007, 3-D seismic interpretation: Cambridge Uni-
versity Press.
Bahorich, M. S., and S. R. Bridges, 1992, Seismic sequence attribute map (SSAM): 62nd
Bahorich, M., and S. Farmer, 1995, 3-D seismic discontinuity for faults and stratigraphic
features: The coherence cube: The Leading Edge, 14, no. 10, 1053– 1058, http://
dx.doi.org/10.1190/1.1437077.
Balch, A. H., 1971, Color sonograms: a new dimension in seismic data interpretation:
Geophysics, 36, no. 6, 1074– 1098, http://dx.doi.org/10.1190/1.1440233.
Barnes, A. E., 2001, Seismic attributes in your facies: CSEG Recorder, 26, no. 7, 41 –42,
44– 47.
237
Barnes, A. E., 2006, A filter to improve seismic discontinuity data for fault interpretation:
Geophysics, 71, no. 3, P1– P4, http://dx.doi.org/10.1190/1.2195988.
Barnes, A. E., 2007, Redundant and useless seismic attributes: Geophysics, 72, no. 3,
P33 –P38, http://dx.doi.org/10.1190/1.2716717.
Barnes, A. E., M. Cole, T. Michaels, P. Norlund, and C. Sembroski, 2011, Making seismic
data come alive: First Break, 29, no. 6, 33–38.
Bergbauer, S., T. Mukerji, and P. Hennings, 2003, Improving curvature analyses of
deformed horizons using scale-dependent filtering techniques: AAPG Bulletin, 87,
no. 8, 1255– 1272, http://dx.doi.org/10.1306/0319032001101.
Berkhout, A. J., 1970, Minimum phase in sampled-signal theory: Ph.D. thesis, Delft Uni-
versity of Technology.
Berkhout, A. J., 1984, Seismic resolution: Resolving power of acoustical echo techniques:
Geophysical Press.
Bodine, J. H., 1984, Waveform analysis with seismic attributes: 54th Annual International
Meeting, SEG, Expanded Abstracts, 505 –509.
Bracewell, R. N., 1965, The Fourier transform and its applications: McGraw-Hill.
Brown, A. R., 2011, Interpretation of three-dimensional seismic data, 7th ed.: AAPG
Memoir 42 and SEG Investigations in Geophysics No. 9, http://dx.doi.org/
10.1190/1.9781560802884.
Castagna, J. P., S. Sun, and R. W. Siegfried, 2003, Instantaneous spectral analysis: Detec-
tion of low-frequency shadows associated with hydrocarbons: The Leading Edge, 22,
no. 2, 120, 122, 124 –127, http://dx.doi.org/10.1190/1.1559038.
Castagna, J. P., H. W. Swan, and D. J. Foster, 1998, Framework for AVO gradient
and intercept interpretation: Geophysics, 63, no. 3, 948– 956, http://dx.doi.org/
10.1190/1.1444406.
Chen, Q., and S. Sidney, 1997, Seismic attribute technology for reservoir forecasting
and monitoring: The Leading Edge, 16, no. 5, 445 –456, http://dx.doi.org/10.1190/
1.1437657.
Chopra, S., 2002, Coherence cube and beyond: First Break, 20, no. 1, 27– 33, http://
dx.doi.org/10.1046/j.1365-2397.2002.00225.x.
Chopra, S., and V. Alexeev, 2006, Applications of texture attribute analysis to 3D seismic
data: The Leading Edge, 25, no. 8, 934 –940, http://dx.doi.org/10.1190/1.2335155.
Chopra, S., and K. J. Marfurt, 2007, Seismic attributes for prospect identification and reser-
voir characterization: SEG.
Chopra, S., and K. J. Marfurt, 2010, Integration of coherence and volumetric curvature
images: The Leading Edge, 29, no. 9, 1092 –1107, http://dx.doi.org/10.1190/
1.3485770.
Chopra, S., and K. J. Marfurt, 2011, Structural curvature versus amplitude curvature: 81st
Chopra, S., and K. J. Marfurt, 2012, Seismic attribute expression of differential compac-
tion: The Leading Edge, 31, no. 12, 1418 –1422, http://dx.doi.org/10.1190/
tle31121418.1.
Claerbout, J. F., 1985, Fundamentals of geophysical data processing: Blackwell Scientific
Publications.
References 239
Claerbout, J. F., 1992, Earth soundings analysis: Processing versus inversion: Blackwell
Scientific Publications.
Close, C. M., 1966, The analysis of linear circuits: Harcourt, Brace & World, Inc.
Cohen, I., and L. C. Coifman, 2002, Local discontinuity measure for 3-D seismic data:
Cohen, L., 1995, Time-frequency analysis: Prentice-Hall PTR.
Coléou, T., M. Poupon, and K. Azbel, 2003, Unsupervised seismic facies classification:
A review and comparison of techniques and implementation: The Leading Edge, 22,
no. 10, 942– 953, http://dx.doi.org/10.1190/1.1623635.
Connolly, P., 1999, Elastic impedance: The Leading Edge, 18, no. 4, 438 –452, http://
dx.doi.org/10.1190/1.1438307.
Conticini, F., 1984, Seismic facies quantitative analysis: New tool in stratigraphic inter-
pretation: 54th Annual International Meeting, SEG, Expanded Abstracts, 680 –682.
Cross, N. E., A. Cunningham, R. J. Cook, A. Taha, E. Esmaie, and N. El Swidan, 2009,
Three-dimensional seismic geomorphology of a deep-water slope-channel system.
The Sequoia field, offshore west Nile Delta, Egypt: AAPG Bulletin, 93, no. 8,
1063– 1086, http://dx.doi.org/10.1306/05040908101.
Dalley, R. M., 2008, Value of visual attributes: revisiting dip and azimuth displays for 3D
seismic interpretation: First Break, 26, no. 4, 87–91.
Dalley, R. M., E. C. A. Gevers, G. M. Stampfli, D. J. Davies, C. N. Gastaldi, P. A. Ruijten-
berg, and G. J. O. Vermeer, 1989, Dip and azimuth displays for 3D seismic interpret-
ation: First Break, 7, no. 3, 86 –95.
de Figueiredo, R. J. P., 1982, Pattern recognition approach to exploration, in K. C. Jain and
R. J. P. de Figueiredo, eds., Concepts and techniques in oil and gas exploration: SEG,
267– 286.
De Groot, P. F. M., 1999, Volume transformation by way of neural network mapping: 61st
Conference and Exhibition, EAGE, Extended Abstracts, 3– 37.
Denham, J. I., and H. R. Nelson Jr., 1986, Map displays from an interactive interpretation:
de Rooij, M., and K. Tingdahl, 2002, Meta-attributes – the key to multivolume, multiattri-
bute interpretation: The Leading Edge, 21, no. 10, 1050– 1053, http://dx.doi.org/
10.1190/1.1518445.
Dobrin, M. B., 1976, Introduction to geophysical prospecting, 3rd ed.: McGraw-Hill, Inc.
Dorn, G. A., and D. A. Fisher, 1989, Detailed fault interpretation of 3-D seismic data for
EOR planning: 59th Annual International Meeting, SEG, Expanded Abstracts,
747– 750.
Duda, R. O., P. E. Hart, and D. G. Stork, 2001, Pattern classification, 2nd ed.: Wiley-
Interscience.
Ebrom, D., 2004, The low-frequency gas shadow on seismic sections: The Leading Edge,
23, no. 8, 772, http://dx.doi.org/10.1190/1.1786898.
Farnbach, J. S., 1975, The complex envelope in seismic signal analysis: Bulletin of the
Seismological Society of America, 65, 951– 962.
Ferguson, C. J., A. Avu, N. Schofield, and G. S. Patton, 2010, Seismic analysis workflow
for reservoir characterization in the vicinity of salt: First Break, 28, no. 10, 107 –113.
Forrest, M., R. Roden, and R. Holeywell, 2010, Risking seismic amplitude anomaly pro-
spects based on database trends: The Leading Edge, 29, no. 5, 570–574, http://
dx.doi.org/10.1190/1.3422455.
Gabor, D., 1946, Theory of communication: Journal of the Institution of Electrical Engin-
eers, 93, no. 26, 429–457.
Gao, D., 2003, Volume texture extraction for 3D seismic visualization and interpretation:
Gelchinsky, B. Y., N. A. Karev, and L. D. Kogan, 1969, Application of Hilbert trans-
form for group correlation of seismic waves, in G. I. Petrashen, ed., Problems in the
dynamic theory of the propagation of seismic waves: Collection X, Nauka Press,
121– 123.
Gelchinsky, B., E. Landa, and V. Shtivelman, 1985, Algorithms of phase and group corre-
lation: Geophysics, 50, no. 4, 596 –608, http://dx.doi.org/10.1190/1.1441935.
Geophysical Service Inc., 1964, advertisement: Geophysics, 29, no. 4, back cover.
Geophysical Service Inc., 1965, advertisement: Geophysics, 30, no. 5, back cover.
Gersztenkorn, A., 2012, A new approach for detecting topographic and geologic
information in seismic data: Geophysics, 77, no. 2, V81 –V90, http://dx.doi.org/
10.1190/geo2011-0216.1.
Gersztenkorn, A., and K. J. Marfurt, 1999, Eigenstructure-based coherence computations
as an aid to 3-D structural and stratigraphic mapping: Geophysics, 64, no. 5,
1468– 1479, http://dx.doi.org/10.1190/1.1444651.
Ghosh, D., A. E. Barnes, and N. H. Ngoc, 2013, Seismic attributes for prospect evaluation
in SE Asia basins: Petroleum Geoscience Conference & Exhibition 2013, EAGE Short
Course 2.
Ghosh, D., M. F. A. Halim, M. Brew, B. Viratno, and N. Darman, 2010, Geophysical issues
and challenges in Malay and adjacent basins from an E & P perspective: The Leading
Edge, 29, no. 4, 436–449, http://dx.doi.org/10.1190/1.3378307.
Goodway, W., T. Chen, and J. Downton, 1997, Improved AVO fluid detection and lithol-
ogy discrimination using Lamé petrophysical parameters: 67th Annual International
Gridley, J., and G. Partyka, 1997, Processing and interpretational aspects of spectral
decomposition: 67th Annual International Meeting, SEG, Expanded Abstracts,
1055– 1058.
Hakami, A. M., K. J. Marfurt, and S. Al-Dossary, 2004, Curvature attribute and seismic
interpretation: Case study from Fort Worth Basin, Texas, USA: 74th Annual Inter-
Hale, D., 2009, Structure-oriented smoothing and semblance: CWP Report 635, Colorado
School of Mines, 261– 270.
Hale, D., 2010, Image-guided 3D interpolation of borehole data: 80th Annual International
Hall, M., 2006, Resolution and uncertainty in spectral decomposition: First Break, 24,
no. 12, 43– 47.
Hall, M., 2007, Smooth operator: Smoothing seismic interpretations and attributes: The
Leading Edge, 26, no. 1, 16–20, http://dx.doi.org/10.1190/1.2431821.
References 241
Hamilton, E. L., 1972, Compressional-wave attenuation in marine sediments: Geophysics,

37, no. 4, 620–646, http://dx.doi.org/10.1190/1.1440287.
Hampson, D. P., J. S. Schuelke, and J. A. Quirein, 2001, Use of multiattribute transforms

to predict log properties from seismic data: Geophysics, 66, no. 1, 220– 236, http://
dx.doi.org/10.1190/1.1444899.
Hardage, B. A., 1987, Seismic stratigraphy: Handbook of geophysical exploration, v. 9:
Geophysical Press.
Hardage, B. A., V. M. Pendleton, J. L. Simmons Jr., B. A. Stubbs, and B. J. Uszynski, 1998,
3-D instantaneous frequency used as a coherency/continuity parameter to interpret
reservoir compartment boundaries across an area of complex turbidite deposition:
Hart, B. S., 2008a, Channel detection in 3-D seismic data using sweetness: AAPG Bulletin,
92, no. 6, 733–742, http://dx.doi.org/10.1306/02050807127.
Hart, B. S., 2008b, Stratigraphically significant attributes: The Leading Edge, 27, no. 3,
320– 324, http://dx.doi.org/10.1190/1.2896621.
Hatton, L., M. H. Worthington, and J. Makin, 1986, Seismic data processing: Theory and
practice: Blackwell Scientific Publications.
Heggland, R., 2013, Hydrocarbon trap classification based on associated gas chimneys: in
F. Aminzadeh, T. B. Berge, and D. L. Connolly, eds., Hydrocarbon seepage: from
source to surface: SEG Geophysical Developments Series No. 16, 221 –230.
Henning, A. T., R. Martin, and G. Paton, 2010, Data conditioning and seismic attribute
analysis in the Eagle Ford Shale Play: Examples from Sinor Ranch, Live Oak County,
Texas: 80th Annual International Meeting, SEG, Expanded Abstracts, 1298– 1301.
Höcker, C., and G. Fehmers, 2002, Fast structural interpretation with structure-oriented fil-
tering: The Leading Edge, 21, no. 3, 238– 243, http://dx.doi.org/10.1190/1.1463775.
Hoetz, H. L. J. G., and D. G. Watters, 1992, Seismic horizon attribute mapping for the
Annerveen Gasfield, The Netherlands: First Break, 10, no. 2, 41 –51.
Horn, B. K. P., 1981, Hill shading and the reflectance map: Proceedings of the IEEE, 69,
no. 1, 14– 47, http://dx.doi.org/10.1109/PROC.1981.11918.
Isaaks, E. H., and R. M. Srivastava, 1989, An introduction to applied geostatistics: Oxford
University Press.
Jacquemin, P., and J. L. Mallet, 2005, Automatic faults extraction using double
Hough transform: 75th Annual International Meeting, SEG, Expanded Abstracts,
755– 758.
James, H., and T. Kostrova, 2005, Use of lighting in the 3D display of seismic data: First
Break, 23, no. 3, 53 –57.
Johnston, D. H., 1981, Attenuation: A state-of-the-art summary, in M. N. Toksöz and D. H.
Johnston, eds., Seismic wave attenuation: SEG Geophysics Reprint Series No. 2,
123– 135.
Johnston, D. H., 1993, Seismic attribute calibration using neural nets: 63rd Annual Inter-
Johnston, D. H., and M. N. Toksöz, 1981, Definitions and terminology, in M. N. Toksöz
and D. H. Johnston, eds., Seismic wave attenuation: SEG Geophysics Reprint Series
No. 2, 1 –5.
Kalkomey, C. T., 1997, Potential risks when using seismic attributes as predictors of
reservoir properties: The Leading Edge, 16, no. 3, 247– 251, http://dx.doi.org/
10.1190/1.1437610.
Kallweit, R. S., and L. C. Wood, 1982, The limits of resolution of zero-phase wavelets:
Kirlin, R. L., 1999, Data vectors and covariances matrices, in R. L. Kirlin and W. J. Done,
eds., Covariance analysis for seismic signal processing: SEG, 5– 18.
Klein, P., L. Richard, and H. James, 2009, 3D curvature attributes: a new approach for
seismic interpretation: First Break, 26, no. 4, 105–111.
Koefoed, O., 1955, On the effect of Poisson’s ratios of rock strata on the reflection coeffi-
cients of plane waves: Geophysical Prospecting, 3, no. 4, 381– 387, http://dx.doi.org/
10.1111/j.1365-2478.1955.tb01383.x.
Koefoed, O., 1960, Measurements of amplitudes of reflected seismic waves: Geophysical
Prospecting, 8, no. 1, 25– 46, http://dx.doi.org/10.1111/j.1365-2478.1960.tb01485.x.
Latimer, R. B., R. Davison, and P. van Riel, 2000, An interpreter’s guide to understanding
and working with seismic-derived acoustic impedance data: The Leading Edge, 19,
no. 3, 242– 256, http://dx.doi.org/10.1190/1.1438580.
Lavergne, M., and C. Willm, 1977, Inversion of seismograms and pseudo velocity logs:
Geophysical Prospecting, 25, no. 2, 231 –250, http://dx.doi.org/10.1111/j.1365-
2478.1977.tb01165.x.
Linari, V., M. Santiago, C. Pastore, K. Azbel, and M. Poupon, 2003, Seismic facies analysis
based on 3D multiattribute volume classification, La Palma Field, Maracaibo, Vene-
zuela: The Leading Edge, 22, no. 1, 32 –36, http://dx.doi.org/10.1190/1.1542752.
Lindseth, R. O., 1979, Synthetic sonic logs – a process for stratigraphic interpretation:
Lindseth, R. O., 1982, Digital processing of geophysical data: A review: SEG.
Liner, C. L., 2004, Elements of 3D seismology, 2nd ed.: PennWell.
Lines, L. R., and R. T. Newrick, 2004, Fundamentals of geophysical interpretation:
SEG Geophysical Monograph Series No. 13, http://dx.doi.org/10.1190/1.97815
60801726.
Lisle, R. J., 1994, Detection of zones of abnormal strains in structures using Gaussian
curvature analysis: AAPG Bulletin, 78, 1811 –1819.
Liu, H. P., 1988, Effect of source spectrum on seismic attenuation measurements using the
pulse-broadening method: Geophysics, 53, no. 12, 1520– 1526, http://dx.doi.org/
10.1190/1.1442433.
Lomask, J., A. Guitton, S. Fomel, J. Claerbout, and A. A. Valenciano, 2006, Flattening
without picking: Geophysics, 71, no. 4, P13 –P20, http://dx.doi.org/10.1190/
1.2210848.
Love, P. L., and M. Simaan, 1984, Segmentation of stacked seismic data by the classifi-
cation of image texture: 54th Annual International Meeting, SEG, Expanded Abstracts,
480– 482.
Luo, Y., W. G. Higgs, and W. S. Kowalik, 1996, Edge detection and stratigraphic analysis
using 3D seismic data: 66th Annual International Meeting, SEG, Expanded Abstracts,
324– 327.
References 243
Lynch, S., 2006, Improving the interpretability of seismic data using achromatic seismic
information: 76th Annual International Meeting, SEG, Expanded Abstracts,
1073– 1077.
Lynch, S., and L. Lines, 2004, Combined attribute displays: 74th Annual International
Lynch, S., J. Townsley, M. Dennis, and C. Gibson, 2005, Enhancing fault visibility using
bump mapped seismic attributes: CSEG National Convention, Expanded Abstracts,
243– 246.
Marfurt, J., 2006, Robust estimates of reflector dip and azimuth: Geophysics, 71, no. 4,
P29 –P40, http://dx.doi.org/10.1190/1.2213049.
Marfurt, K. J., and R. L. Kirlin, 2000, 3-D broad-band estimates of reflector dip and ampli-
tude: Geophysics, 65, no. 1, 304 –320, http://dx.doi.org/10.1190/1.1444721.
Marfurt, K. J., and R. L. Kirlin, 2001, Narrow-band spectral analysis and thin-bed tuning:
Marfurt, K. J., R. L. Kirlin, S. L. Farmer, and M. S. Bahorich, 1998, 3-D seismic attributes
using a semblance-based coherency algorithm: Geophysics, 63, no. 4, 1150– 1165,
http://dx.doi.org/10.1190/1.1444415.
Marfurt, K. J., V. Sudhaker, A. Gersztenkorn, K. D. Crawford, and S. E. Nissen, 1999,
Coherency calculations in the presence of structural dip: Geophysics, 64, no. 1,
104– 111, http://dx.doi.org/10.1190/1.1444508.
Masson, P. A. A. H., and F. J. Agnich, 1958, Seismic survey of Sinai and the Gulf of Suez:
Mathieu, P. G., and G. W. Rice, 1969, Multivariate analysis used in the detection of
stratigraphic anomalies from seismic data: Geophysics, 34, no. 4, 507 –515, http://
dx.doi.org/10.1190/1.1440027.
Merlini, E., 1960, A new device for seismic survey equipment: Geophysical Prospecting,
8, no. 1, 4–11, http://dx.doi.org/10.1111/j.1365-2478.1960.tb01483.x.
Milkereit, B., and C. Spencer, 1990, Multiattribute processing of seismic data: Application
to dip displays: Canadian Journal of Exploration Geophysics, 26, 47–56.
Mitchell, J. T., N. Derzhi, E. Lichman, and E. N. Lanning, 1996, Energy absorption
analysis: A case study: 66th Annual International Meeting, SEG, Expanded Abstracts,
1785– 1788.
Mitchum, R. M., P. R. Vail, and J. B. Sangree, 1977, Seismic stratigraphy and global
changes of sea level: Part 6 — Stratigraphic interpretation of seismic reflection patterns
in depositional sequences, in C. E. Payton, ed., Seismic stratigraphy: Applications to
hydrocarbon exploration: AAPG Memoir 26, 117 –133.
Mlsna, P. A., and J. J. Rodriguez, 2009, Gradient and Laplacian edge detection, in
A. Bovik, ed., The essential guide to image processing: Academic Press.
Narhari, S. R., A. L. Al-Kandari, V. K. Kidambi, S. Al-Ashwak, B. Al-Qadeeri, and
C. Pattnaik, 2009, Understanding fractures through seismic data: North Kuwait case
study: 79th Annual International Meeting, SEG, Expanded Abstracts, 547 –551.
Nestvold, E. O., 1996, The impact of 3-D seismic data on exploration, field development,
and production, in P. Weimar and T. L. Davis, eds., Applications of 3-D seismic data to
exploration and production: AAPG Studies in Geology 42, pp. 1 –10.
Nikolaidis, N., and I. Pitas, 2001, 3-D image processing algorithms: Wiley-Interscience.
Nourollah, H., J. Keetley, and G. O’Brien, 2010, Gas chimney identification through
seismic attribute analysis in the Gippsland Basin, Australia: The Leading Edge, 29,
no. 8, 896– 901, http://dx.doi.org/10.1190/1.3479999.
O’Brien, J., 2004, Seismic amplitudes from low gas saturation sands: The Leading Edge,
23, no. 12, 1236–1243, http://dx.doi.org/10.1190/leedff.23.1236_1.
Oldenburg, D. W., T. Scheuer, and S. Levy, 1983, Recovery of the acoustic impedance
from reflection seismograms: Geophysics, 48, no. 10, 1318–1337, http://dx.doi.org/
10.1190/1.1441413.
Oliveros, R. B., and B. J. Radovich, 1997, Image-processing display techniques applied to
seismic instantaneous attributes on the Gorgon gas field, North West Shelf, Australia:
67th Annual International Meeting, SEG, Expanded Abstracts, 2064– 2067.
Ostrander, W. J., 1982, Plane wave reflection coefficients for gas sands at nonnormal angles
of incidence: 52nd Annual International Meeting, SEG, Expanded Abstracts,
216– 218.
Parks, T. W., and C. S. Burrus, 1987, Digital filter design: John Wiley & Sons, Inc.
Partyka, G., J. Gridley, and J. Lopez, 1999, Interpretational applications of spectral
decomposition in reservoir characterization: The Leading Edge, 18, no. 3, 353– 360,
http://dx.doi.org/10.1190/1.1438295.
Pedersen, S. I., T. Randen, L. Sonneland, and O. Steen, 2002, Automatic fault extraction
using artificial ants, 72nd Annual International Meeting, SEG, Expanded Abstracts,
512– 515.
Picou, C., and R. Utzmann, 1962, La “Coupe sismique vectorielle”: Un pointé semi-
automatique: Geophysical Prospecting, 10, no. 4, 497 –516, http://dx.doi.org/
10.1111/j.1365-2478.1962.tb00004.x.
Pilcher, R. S., and R. D. Blumstein, 2007, Brine volume and salt dissolution rates
in Orca basin, northeast Gulf of Mexico: AAPG Bulletin, 91, no. 6, 823– 833,
http://dx.doi.org/10.1306/12180606049.
Poularikas, A. D., and S. Seely, 1985, Signals and systems: PWS Engineering.
Poupon, M., K. Azbel, and G. Palmer, 1999, A new methodology based on seismic facies
analysis and litho-seismic modeling: The Elkhorn Slough field pilot project, Solano
County, California: 69th Annual International Meeting, SEG, Expanded Abstracts,
927– 930.
Press, W. H., S. A. Teukolsky, W. T. Vetterling, and B. F. Flannery, 1992 (reprinted 1995),
Numerical recipes in C, 2nd ed.: Cambridge University Press.
Proubasta, D., 2000, Interview with Mike Forrest, “father of bright spots” and “oil finder” at
Shell Oil: The Leading Edge, 19, no. 11, 1184 –1186, http://dx.doi.org/10.1190/
1.1438500.
Puryear, C. I., O. N. Portniaguine, C. M. Cobos, and J. P. Castagna, 2012, Constrained
least-squares spectral analysis: Application to seismic data: Geophysics, 77, no. 5,
V143 –V167, http://dx.doi.org/10.1190/geo2011-0210.1.
Radovich, B. J., and R. B. Oliveros, 1998, 3-D sequence interpretation of seismic instan-
taneous attributes from the Gorgon field: The Leading Edge, 17, no. 9, 1286– 1293,
http://dx.doi.org/10.1190/1.1438125.
References 245
Randen, T., E. Monsen, C. Signer, A. Abrahamsen, J. O. Hansen, T. Saeter, J. Schlaf, and

L. Sonneland, 2000, Three-dimensional texture attributes for seismic data analysis:
70th Annual International Meeting, SEG, Expanded Abstracts, 668– 671.

Randen, T., S. Pedersen, and L. Sonneland, 2001, Automatic extraction of fault surfaces
from three-dimensional seismic data, 71st Annual International Meeting, SEG,
Expanded Abstracts, 551– 554.
Rich, J., 2008, Expanding the applicability of curvature attributes through clarification of
ambiguities in derivation and terminology: 78th Annual International Meeting, SEG,
Rijks, E. J. H., and J. C. E. M. Jauffred, 1991, Attribute extraction: An important appli-
cation in any detailed 3-D interpretation study: The Leading Edge, 10, no. 9, 11– 19,
http://dx.doi.org/10.1190/1.1436837.
Roberts, A., 2001, Curvature attributes and their application to 3D interpreted hori-
zons: First Break, 19, no. 2, 85 –100, http://dx.doi.org/10.1046/j.0263-5046.2001.
00142.x.
Robertson, J. D., and D. A. Fisher, 1988, Complex seismic trace attributes: The Leading
Edge, 7, no. 6, 22–26, http://dx.doi.org/10.1190/1.1439517.
Robertson, J. D., and H. H. Nogami, 1984, Complex seismic trace analysis of thin beds:
Roden, R., M. Forrest, and R. Holeywell, 2012, Relating seismic interpretation to reserve/
resource calculations: Insights from a DHI consortium: The Leading Edge, 31, no. 9,
1066– 1074, http://dx.doi.org/10.1190/tle31091066.1.
Rummerfield, B. F., 1954, Reflection quality, a fourth dimension: Geophysics, 19, no. 4,
684– 694, http://dx.doi.org/10.1190/1.1438038.
Russell, B. H., 1988, Introduction to seismic inversion methods: SEG.
Russell, B. H., and D. Hampson, 2006, The old and the new in seismic inversion: CSEG
Recorder, 31, no. 10, 5– 11.
Rutherford, S. R., and R. H. Williams, 1989, Amplitude-version-offset variations in gas
sands: Geophysics, 54, no. 6, 680 –688, http://dx.doi.org/10.1190/1.1442696.
Saha, J. G., 1987, Relationship between Fourier and instantaneous frequency: 57th Annual
International Meeting, SEG, Expanded Abstracts, 591– 594.
Sangree, J. B., and J. M. Widmier, 1977, Seismic stratigraphy and global changes of sea
level: Part 9 — Seismic interpretation of clastic depositional facies, in C. E. Payton, ed.,
Seismic stratigraphy: Applications to hydrocarbon exploration: AAPG Memoir 26,
165– 184.
Savit, C. H., 1960, Preliminary report: A stratigraphic seismogram: Geophysics, 25, no. 1,
312– 321, http://dx.doi.org/10.1190/1.1438697.
Scheuer, T. E., and D. W. Oldenburg, 1988, Local phase velocity from complex
seismic data: Geophysics, 53, no. 12, 1503– 1511, http://dx.doi.org/10.1190/
1.1442431.
Schmidt, I., S. Lacaze, and G. Paton, 2013, Spectral decomposition combined with geo-
model interpretation: Creating new workflows by integrating advanced technologies
for imaging and interpretation: 75th Conference and Exhibition, EAGE, Extended
Abstracts, Tu1709.
Schuelke, J. S., and J. A. Quirein, 1998, Validation: A technique for selecting seismic attri-
butes and verifying results: 68th Annual International Meeting, SEG, Expanded
Abstracts, 936 –939.

SDS Data Systems, 1966, advertisement: Geophysics, 31, no. 2, A-18.
Seiscom Delta Inc., 1977, advertisement: Geophysics, 42, no. 4, A-21.
Sharma, R. K., and S. Chopra, 2012, New attribute for determination of lithology
and brittleness: 82nd Annual International Meeting, SEG, Expanded Abstracts,
1– 5.
Sheffield, T. M., and B. A. Payne, 2008, Geovolume visualization and interpretation: What
makes a useful visualization seismic attribute? 70th Annual International Meeting,
SEG, Expanded Abstracts, 849–853.
Sheriff, R. E., 1975, Factors affecting seismic amplitudes: Geophysical Prospecting, 23,
no. 1, 125– 138, http://dx.doi.org/10.1111/j.1365-2478.1975.tb00685.x.
Sheriff, R. E., 1980, Seismic stratigraphy: International Human Resources Development
Corp., http://dx.doi.org/10.1007/978-94-011-6395-8.
Sheriff, R. E., and L. P. Geldart, 1995, Exploration seismology, 2nd ed.: Cambridge Uni-
versity Press. http://dx.doi.org/10.1017/CBO9781139168359.
Smith, G. C., and P. M. Gidlow, 1987, Weighted stacking for rock property estimation and
detection of gas: Geophysical Prospecting, 35, no. 9, 993–1014, http://dx.doi.org/
10.1111/j.1365-2478.1987.tb00856.x.
Sonneland, L., 1983, Computer aided interpretation of seismic data: 53rd Annual Inter-
Sonneland, L., O. Barkved, M. Olsen, and G. Snyder, 1989, Application of seismic wave-
field attributes in reservoir characterization: 59th Annual International Meeting, SEG,
Stacey, F. D., 1977, Physics of the earth, 2nd ed.: John Wiley & Sons, Inc.
Stanley, W. D., 1975, Digital signal processing: Reston Publishing Co., Inc.
Stark, T. J., 2004, Relative geologic time (age) volumes — Relating every seismic sample
to a geologically reasonable horizon: The Leading Edge, 23, no. 9, 928– 932, http://
dx.doi.org/10.1190/1.1803505.
Stark, T. J., 2006, Visualization techniques for enhancing stratigraphic inferences from
3D seismic data volumes: First Break, 24, no. 4, 75– 85.
Strecker, U., and R. Uden, 2002, Data mining of 3D poststack seismic attribute volumes
using Kohonen self-organizing maps: The Leading Edge, 21, no. 10, 1032– 1037,
http://dx.doi.org/10.1190/1.1518442.
Sullivan, E. C., K. J. Marfurt, A. Lacazette, and M. Ammerman, 2006, Application of new
seismic attributes to collapse chimneys in the Fort Worth Basin: Geophysics, 71, no. 4,
B111 –B119, http://dx.doi.org/10.1190/1.2216189.
Taner, M. T., 2000, Attributes revisited, www.rocksolidimages.com/pdf/attrib_revisited.
htm, accessed October 17, 2015.
Taner, M. T., 2001, Seismic attributes: CSEG Recorder, 26, no. 7, 48–50, 52 –56.
Taner, M. T., F. Koehler, and R. E. Sheriff, 1979, Complex seismic trace analysis: Geo-
physics, 44, no. 6, 1041 –1063, http://dx.doi.org/10.1190/1.1440994.
References 247
Taner, M. T., and R. E. Sheriff, 1977, Application of amplitude, frequency, and other attri-
butes to stratigraphic and hydrocarbon exploration, in C. E. Payton, ed., Seismic stra-
tigraphy: Applications to hydrocarbon exploration: AAPG Memoir 26, 301 –327.

Tingdahl, K., and N. Hemstra, 2003, Estimating fault-attribute orientation with gradient
analysis, principal component analysis and the localized Hough-transform: 73rd
Toksöz, M. N., and D. H. Johnston, 1981, Seismic wave attenuation: SEG Geophysics
Reprint Series No. 2.
Tonn, R., 1991, The determination of the seismic quality factor Q from VSP data: A com-
parison of different computational methods: Geophysical Prospecting, 39, no. 1, 1 –27,
http://dx.doi.org/10.1111/j.1365-2478.1991.tb00298.x.
Trappe, H., J. Schubarth-Engelschall, and E. Laggiard, 2002, Seismic facies classification
of deep Rotliegend sandstones by neural network techniques: 64th Conference and
Exhibition, EAGE, Extended Abstracts, H39.
Treadgold, G., B. Campbell, and B. McLain, 2011, Eagle Ford shale prospecting with 3D
seismic data within a tectonic and depositional system framework: The Leading Edge,
30, no. 1, 48–53, http://dx.doi.org/10.1190/1.3535432.
Vakman, D., 1998, Signals, oscillations, and waves: Artech House, Inc.
Van Bemmel, P. P., and R. E. F. Pepper, 2000, Seismic signal processing method and
apparatus for generating a cube of variance values: U. S. Patent 6 151 555.
van Hoek, T., S. Gesbert, and J. Pickens, 2010, Geometric attributes for seismic strati-
graphic interpretation: The Leading Edge, 29, no. 9, 1056– 1065, http://dx.doi.org/
10.1190/1.3485766.
Van Melle, F. A., D. L. Faass, S. Kaufman, G. W. Postma, and A. J. Seriff, 1963, Geophy-
sical research and progress in exploration: Geophysics, 28, no. 3, 466 –478, http://
dx.doi.org/10.1190/1.1439200.
Walls, J. D., M. T. Taner, G. Taylor, M. Smith, M. Carr, N. Derzhi, J. Drummond,
D. McGuire, S. Morris, J. Bregar, and J. Lakings, 2002, Seismic reservoir characteriz-
ation of a U.S. Midcontinent fluvial system using rock physics, poststack seismic attri-
butes, and neural networks: The Leading Edge, 21, no. 5, 428– 436, http://dx.doi.org/
10.1190/1.1481248.
Waters, K. H., 1981, Reflection seismology: A tool for energy resource exploration, 2nd
ed.: John Wiley and Sons.
White, R. E., 1991, Properties of instantaneous seismic attributes: The Leading Edge, 10,
no. 7, 26– 32, http://dx.doi.org/10.1190/1.1436827.
Widess, M. B., 1973, How thin is a thin bed?: Geophysics, 38, no. 6, 1176 –1180, http://
dx.doi.org/10.1190/1.1440403.
Wu, X. M., and D. Hale, 2014, Horizon volumes with constraints: 84th Annual Inter-
Yilmaz, Ö., 1987, Seismic data processing: SEG.
Yilmaz, Ö., 2001a, Seismic data analysis: Processing, inversion, and interpretation of
seismic data, in S. M. Doherty, ed., SEG Investigations in Geophysics No. 10, v. 1,
http://dx.doi.org/10.1190/1.9781560801580.
Yilmaz, Ö., 2001b, Seismic data analysis: Processing, inversion, and interpretation of
seismic data, in S. M. Doherty, ed., SEG Investigations in Geophysics No. 10, v. 2,
http://dx.doi.org/10.1190/1.9781560801580.
Zeng, H. L., 2010, Geologic significance of anomalous instantaneous frequency: Geophys-
ics, 75, no. 3, P23– P30, http://dx.doi.org/10.1190/1.3427638.
Index
A examples of attribute classification, 174–175

K-means clustering, 172–173
absorption, 141 Kohonen SOFM, 173–174
acoustic impedance (Z), 133 supervised classification, 169, 171
amplitude, 7 –8, 34 –35 unsupervised classification, 169, 171–172
acceleration, 53 average attributes, 57
attributes, 7–8 average frequency, 60
change, 9, 159 average quality factor, 62
curvature, 86 bandwidth, 61
mapping, 185 phase attributes, 66 –67
preservation, 182 thin-bed indicator, 61 –62
variance, 93– 94 AVO analysis. see amplitude-variation-with-offset
amplitude-variation-with-offset analysis analysis (AVO analysis)
(AVO analysis), 14 –15, 22, 158 azimuth, 10, 64 –65, 75, 159–160
analog-to-digital converters, 13– 14 apparent slope and dip, 79 –80
analysis windows, 201 –205 complex seismic trace analysis, 76
Anstey, Nigel, 19–20 dip scanning, 76–77
apparent polarity, 59, 178 exaggerated slope and dip, 78 –79
arc length, 38, 178 examples, 80
attenuation, 141 gradient squared tensor, 78
due to buried channel, 144 –148 plane-wave destructor, 77
examples of spectral ratioing, 151 –152
in homogeneous earth, 143– 144 B
quality factors, 142 –143
spectral ratio method for Q estimation, bandwidth, 9, 37, 61
148 –150 blended displays, 158
attribute boxcar window, 201, 204, 205
analysis, 1 bright spots, 185
chaos, 92 bump mapping, 156
class, 169
revival, 24 C
space reduction, 167
spaces, 166 channels, 191– 192
spikes, 54 chaos attribute, 92
vectors, 166 Claerbout’s formula for instantaneous frequency.
attribute maps, 27 see Scheuer and Oldenburg’s
horizon attributes, 27 –31 approximation
interval attributes, 31–43 coherence. see seismic discontinuity
automatic gain, 195 –196 complex notation, 53
automatic pattern recognition, 169, 194 complex seismic trace analysis, 20– 21, 45, 76.
classified seismic data, 169– 170 see also multiattribute analysis
distance measures, 170 –171 average attributes, 57 –62
249
complex seismic trace analysis (Continued) discrete approximation for relative amplitude
gallery of complex trace attributes, 67–72 change, 215
1D complex seismic trace analysis, discrete derivative filter, 211

45– 54 discrete differentiation, 211
response attributes, 57 –62 discrete Fourier transform, 121–122
spikes, 54 –57 discrete Hilbert transform operator, 208, 209
3D complex seismic trace analysis, 62–67 distance measures, 170–171
complex trace, 53 –54 divergence, 92
complex trace attributes examples, 67 dominant frequency, 178
1D attributes, 68–70
3D attributes, 70–71 E
continuous Hilbert transform operator, 208
correlation, 97 –98, 163 edge detection, 28
correlation coefficient, 162 –163, 220 energy, 34–35
covariance, 100 energy half-time, 38– 39, 178
discontinuity, 222 energy ratios, discontinuity as, 95
matrix, 222 comparison of attributes, 102
crossplots, 161 correlation, 97–98
analysis, 162 covariance, 100
correlation, 163 dip corrections, 102– 103
example, 164 –165 semblance, 98 –99
rank correlation, 163 weighted correlation, 100–101
curvature, 10– 12, 29 –31, 75, 159 envelope gain, 196
exaggerated slope, 78–79
D F
data processing, 13 –14 fault attributes, 110, 159
derivative filter, 211–213 faults, 188–191
discrete, 211 feature maps, 173
destructive interference, 54–56 filter bank, 121, 122
diapirs, 192 –193 Fourier transform, 31, 45 –46, 115, 124
digital recording discrete, 121
and bright spots, 13, 18 of boxcar window, 201
data processing, 13 –14 of complex delta function, 209
digital revolution, 17 of complex trace, 209
reflection seismology, 16 of Hamming window, 202, 203
dim spots, 185 of Hilbert transform operator, 207–208
dip, 10, 65–66, 75, 78–79, 159 of ideal differentiator, 211
apparent slope and dip, 79– 80 inverse, 31, 208, 209
complex seismic trace analysis, 76 frequency, 8–9, 35–37, 48 –50
corrections, 102 –103 attributes, 178
examples, 80 change, 53
gradient squared tensor, 78 shadows, 186
plane-wave destructor, 77 frequency maps, 121. see also
scanning, 76–77 attribute maps
directional attributes, 156, 157
discontinuity, 12, 24, 104, 158, 159, 179, 188 –189. G
see also energy ratios, discontinuity on
gradient squared tensor, 105 –106 Gabor wavelets, 122– 123, 124
relative amplitude change, 106– 108 gain window, 195
wavenumber, 104 –105 gained trace, 196
Index 251
gas chimneys, 192 –193 interval attributes, 31, 177. see also horizon
gas sands, 186 attributes
Gaussian curvatures, 11, 22 amplitude attributes, 34 –35, 40, 41

Gaussian derivatives, 109 arc length, 38
Gaussian window, 202 bandwidth attributes, 37, 42
geobody extraction, 193– 195 energy, 34–35
geological attributes, 2 energy half-time, 38 –39
geophysical attributes, 2– 3 examples, 39 –43
Gibbs’ phenomenon, 201 frequency, 35– 37
gradient squared tensor, 78 frequency attributes, 42
discontinuity, 105 –106 spectral bandwidth, 42–43
group vector, 67 statistical measures, 33– 34
inversion methods, 25
H
K
“half bandwidth”, 9
Hamming window, 201, 202, 204, 205 K-means clustering, 172–173
Hilbert transform, 46, 207, 209 Kohonen SOFM, 129, 130, 173–174
Hilbert transform operator, 197, 207
continuous, 208
L
discrete, 208, 209
horizon attributes, 27. see also interval Laplacian filtering, 109
attributes lithological attributes, 2
curvature, 29 –31 low-cut filter, 137
edge detection, 28 low-frequency
examples, 31 anomaly, 187
slope, dip, and azimuth, 28 shadows, 185–188
horizon azimuth, 28
horizon dip, 28 M
horizon slope, 28
hybrid attributes, 169 mathematical attributes, 3
mean curvatures, 11
I median filtering, 107, 108
meta-attributes, 169
interference mode filtering, 170, 175
constructive, 115 –116 Morlet wavelets, 122–123, 124–125
destructive, 54–56, 115 –116 multiattribute analysis, 24 –25, 155, 156.
igneous diapir, 196 see also seismic attributes
illumination, 156 automatic pattern recognition, 169–175
image texture analysis, 172 crossplots, 161–165
instantaneous amplitude acceleration, 53 principal component analysis, 165–168
instantaneous attributes, 4, 57 volume blending, 155– 161
instantaneous bandwidth, 51 multiple probes, 161
instantaneous dip, 65
instantaneous frequency, 48– 49, 50, 186,
N
188, 215
change, 53 Nyquist frequency, 211
comparison, 216
discrete approximation for, 216
O
Scheuer and Oldenburg’s approximation, 215
instantaneous phase, 47–48 1D attributes, 68–70
instantaneous wavelength, 65 1D boxcar window, 201
1D complex seismic trace analysis, 45. recursive initialization, 173

see also 3D complex seismic trace recursive inversion, 134, 138
analysis earth model, 135

amplitude, 47 inversion formula, 136
amplitude acceleration, 53 spectrum of earth’s reflectivity, 137
complex trace, 53 –54 red-green-blue color (RGB color), 18, 113,
foundation, 46–47 191, 192
Fourier frequency transforms, 45 RGB blending, 158
frequency, 48–50 redundant attribute, 177–179
frequency change, 53 reflection
phase, 47–48 amplitude variance, 90
quality factor, 52 divergence, 91–92
relative amplitude change and bandwidth, parallelism, 91
50– 51 patterns, 92–93
rms frequency, 52 reflection-free zones, 93
opacity functions, 155, 157 spacing, 90–91
reflection strength, 7–8, 19, 47, 179, 186
P crossplots of, 162, 164
relative acoustic impedance, 133. see also
parallelism, 164 attenuation
peak frequency, 37 and Q, 133
phase, 8 examples, 138–141
curvature, 86 low-cut filter, 137
rotation, 196 –197 recursive inversion, 134–137
vector, 63– 64 restoring background trend, 138
plane-wave destructor, 77 seismic data, 134
polarity, 8 wavelets and noise, 138
poststack attributes, 188 relative amplitude change, 106– 108
poststack seismic attributes, 1, 2, 177 reservoirs, 185
seismic attributes, 1–7 response attributes, 4, 57. see also seismic
seismic properties, 7–12 attributes
presentation, 184 –185 apparent polarity, 59
prestack attributes, 1 response frequency, 57–59
principal component analysis, 155, 165– 168 response phase, 57 –59
illustration, 165 –166 sweetness, 59 –60
of seismic attributes, 166, 168, 222 “wavelet” attributes, 57
proliferation, 22– 23 response bandwidth, 57, 59, 179
pulse broadening method, 142 response frequency, 57 –59
P-wave, 1 response phase, 57 –59
RGB color. see red-green-blue color (RGB color)
Q root-mean-square (rms), 4, 34 –35, 127
amplitude, 35, 193
quality factor (Q), 52, 70, 133, 142 –143. frequency, 52
see also attenuation gain, 195–196
spectral ratio method for estimation, 148 –150 spectral frequency, 36
R S
raised cosine window. see Hamming window Scheuer and Oldenburg’s approximation of
rank correlation, 163 instantaneous frequency, 215, 217
real notation, 53 Schwartz inequality, 221
reconnaissance, 184 –185 seismic attenuation, 141
Index 253
seismic attributes, 1, 2, 13, 27, 177 –179. see also discontinuity, 12

multiattribute analysis; response attributes frequency, 8–9
analysis, 177 phase, 8

and methods, 2 polarity, 8
attribute selection, 177 slope, dip, and azimuth, 10
bright spots and amplitude mapping, 185 seismic reflections, 186
categorization, 4, 5 seismic shaded relief, 80
channels, 191–192 examples, 84 –85
characteristics, 4–7 illumination formulas, 83 –84
complex seismic trace analysis, 20– 21 illumination models, 81 –83
data limitations, 180– 183 seismic stratigraphy, 89
for common applications, 180 seismic traces, 96 –97, 142, 196, 220
for measuring key seismic properties, 179 seismic tuning analysis, 115
developments, 25–26 seismic waveform, 128
diapirs, 192– 193 semblance, 98 –99
digital recording and bright spots, 13 –18 shaded relief, 160
discontinuity and attribute revival, 24 signal attenuation, 18
disillusionment, 22 –23 similarity. see seismic discontinuity
faults, 188 –191 slope, 10, 64– 65
gas chimneys, 192 –193 sparse spike inversion, 22
geobody extraction, 193 –195 spectral decomposition, 4, 24, 113
in data processing, 195– 198 discrete Fourier transform, 121–122
low-frequency shadows, 185 –188 example, 127–128
methods of computation, 3–4 filter bank, 121, 122
multiattribute analysis, 24 –25 frequency maps, 121
Nigel Anstey’s attributes, 19–20 RGB blending facilitates, 126
preparation for interpretation, 180 spectral attributes, 125–127
presentation, 184 –185 thin-bed thickness estimation, 125
proliferation, 22 –23 thin beds, 113–120
reconnaissance, 184 –185 time-frequency analysis, 120
redundant and flawed attributes, 177 –179 wavelets, 122– 125
seismic stratigraphy and inversion, 21–22 spectral ratioing, 150– 152
shallow imaging problem, 182 spectral ratio method for Q estimation, 148–150
signal attenuation, 18 spikes, 54
types, 2–3 destructive interference, 54– 56
vertical amplitude defect, 183 faults, 57
seismic continuity, 12 synthetic seismic data, 55, 56
seismic data interpretation, 184 statistical attributes, 4, 33–34
seismic discontinuity, 12, 95 stratigraphic attributes, 89
as derivatives, 104 –108 examples, 93 –94
as energy ratios, 95 –103 reflection amplitude variance, 90
fault attribute, 110 reflection divergence, 91–92
improving discontinuity attributes, 108 reflection parallelism, 91
Laplacian filtering, 109 reflection patterns, 92–93
tapered windows, 108 –109 reflection spacing, 90 –91
seismic facies, 89, 169 structural attributes, 2, 75
seismic horizon, 27, 32 azimuth, 75–80
seismic properties, 7 dip, 75 –80
amplitude, 7 –8 seismic shaded relief, 80 –85
amplitude change, 9 volume curvature, 85 –89
bandwidth, 9 structurally guided processes, 197–198
curvature, 10 –12 supervised classification, 25, 169, 171
supervised waveform classification, 129 alternative attribute combinations, 158

S-wave, 1 amplitude change, 159
sweetness, 59 –60, 70, 178, 185 azimuth, 159–160

blended displays, 158
T bump mapping, 156
curvature, 159
tapered windows, 108 –109 delineate channels, 159
thin beds, 113 dip, 159
indicator, 61 –62 directional attributes, 156, 157
response, 116 –119 discontinuity, 158, 159
thickness estimation, 125 examples, 160–161
thin-bed model, 113– 115 illumination, 156
tuning, 115 –116 opacity functions, 155, 157
Widess limit of resolution, 119 shaded relief, 160
3D attributes, 70–71 strategy for, 157
3D complex seismic trace analysis, 62. see also 1D workflow for, 158
complex seismic trace analysis volume curvature, 85
average phase attributes, 66 –67 amplitude curvature, 86
azimuth, 64–65 examples, 86 –89
dip, 65– 66 phase curvature, 86
group vector, 67 reflection slopes, 85 –86
phase vector, 63 –64 volume reconnaissance, 184
slope, 64 –65 volume spectral decomposition, 188, 193
wavelength, 65 –66
time-frequency methods, 4
time-variant deconvolution 182 W
time-variant filtering, 182
trace attributes, 201 waveform classification, 113, 128
trace envelope, 47 classification methods, 129–130
traveling salesman problem, 129 errors in, 131–132
true amplitude, 182 examples, 130–131
tuning, 115 –116 seismic waveform, 128
unsupervised waveform maps, 131
U waveform maps, 130–131
wavelength, 65 –66
uncertainty principle, 119 –120 wavelet(s), 122, 133
unit vector, 219 –220 attributes, 57
unsupervised classification, 169, 171 –172 Gabor wavelets, 122– 123, 124
unsupervised methods, 25 Morlet wavelets, 122–123, 124–125
unsupervised waveform transform, 121–122
classification, 129 wavenumber discontinuity, 104–105
maps, 131 weighted correlation, 100–101
weighting function, 60
V Widess limit of resolution, 119
variance of seismic traces, 220
vector traces, 219 –222 Z
volume attributes, 201
volume blending, 155, 156 zero-crossing frequency, 35 –36

A Handbook For Seismic Attributes

Uploaded by

Copyright:

Available Formats

You might also like

A Handbook For Seismic Attributes

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Handbook For Seismic Attributes

Uploaded by

Copyright:

Available Formats

Downloaded 12/13/16 to 155.247.166.234. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.

Geophysical References Series No. 21

# 2016 Society of Exploration Geophysicists

Library of Congress Control Number: 2016945977

ISBN 978-0-931830-47-1 (Series)

This page has been intentionally left blank

About the Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

Chapter 1: Overview of Poststack Seismic Attributes .............................. 1

Chapter 2: History of Seismic Attributes ......................................... 13

Chapter 3: Attribute Maps and Interval Attributes ............................... 27

Chapter 4: Complex Seismic Trace Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Gallery of complex trace attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

Chapter 5: Structural and Stratigraphic Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

Chapter 6: Seismic Discontinuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

Chapter 7: Spectral Decomposition and Waveform Classification ................ 113

Chapter 8: Relative Acoustic Impedance and Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

Chapter 9: Multiattribute Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

Chapter 10: Applying Seismic Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Choosing suitable attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Appendix A: Analysis Windows ................................................. 201

This page has been intentionally left blank

Arthur E. Barnes received a BS (1974) in physics from Denison University, an MS (1980)

This page has been intentionally left blank

The Handbook of Poststack Seismic Attributes is a general reference for poststack

This page has been intentionally left blank

This page has been intentionally left blank

Overview of Poststack Seismic Attributes

Attributes and methods

geological information. Geophysical attributes are relatively straightforward to measure,

Statistics Average value, rms amplitude, largest value, maximum

Characteristics of seismic attributes

Mathematical Geophysical Geological

Statistics Miscellaneous Wave Wavelet Reflection patterns

Attributes quantify properties

Attributes are filters

Attributes should have useful meaning

Attributes should be unique

Attributes should be easy to use

the display color scale.

attribute computed by smoothing a rectified

puted through complex seismic trace analysis

Differentiation +908 sin(u) cos (u) Derivative filter

include instantaneous frequency, zero-

quency, rms frequency, and tuning

Slope, dip, and azimuth

History of Seismic Attributes

Digital recording and bright spots

Figure 2. A conceptual illustration of a dendritic channel system mapped by hand according to

Reflection seismology changed dramatically in 1963 with the introduction of digital

Chapter 2: History of Seismic Attributes

Nigel Anstey’s attributes

Nigel Anstey’s landmark exposition of seismic attributes followed on the heels of

Complex seismic trace analysis

Seismic stratigraphy and inversion

include amplitude, spacing, continuity, parallelism, and other reflection configurations.

Proliferation and disillusionment

attributes.” Robertson and Fisher (1988) championed response attributes as time-variant

Discontinuity and attribute revival