Seismic Attributes for Prospect Identification and Reservoir Characterization Satinder Chopra Kurt J. Marfurt Geophysical Developments LeISBN 978-0-931830-41-9 (Series) ISBN 978-1-S6080-141-2 (Volume) Copyright © 2007 Society of Exploration Geophysicists PO. Box 702740 Tulsa, OK U.S.A. 74170-2740 All rights reserved. No part of this book may be reproduced, stored in a retrieval system, or transcribed in any form or by any means, electronic or mechanical, including photocopying and recording, without prior written permission of the publisher Published 2007 Reprinted 2008 Printed in the United States of America Library of Congress Cataloging-in-Publication Data Chopra, Satinder. Seismic attributes for prospect identification and reservoir characterization / Satinder Chopra, Kurt J. Marfurt p-cm. -- (SEG geophysical developments Includes bibliographical references and index. ISBN 978-1-56080- 141-2 (volume) -- ISBN 978-0-93 1830-4 1-9 (serie 1. Seismic refl series ; no. 11) ) cismic prospecting. I, Marfurt, K. J. II. Title. tion method. 2. QES39.C52 2007 622' .1592--de22 2007024043Table of Coniens. —v Table of Contents About the Authors Acknowledgments, Chapter 1: Overview of Seismie Attributes. 1 Introguetion| 1 Structure of this Book 2 Historical Development of Attributes 2 Digital Recording and Bright-spot Detection 2 Inicoduetion of color in seismie displays. 3 Color plotting of seismic data 4 Seismic-impedance inversion 5 Proliferation of attributes in the 194 5 Response attributes 6 Interpretive workstations 6 Atteihutes fall out of fay a Twoslimensional auibute 4 Horizon and interval attributes 3 Industry Adoption of 3D Seismic Technology Seismic sequence attribute mapping 8 3D seismic exploration comes of age 8 Seismic ohemnee vir wviw MAI | O Speciral decomposition 19 sismic inversion revisited ut Crossploiting of attributes i ‘Automated pattern recognition on attributes 12 Neural-network application for multiatiribute analysis B Enhanced visualization helps atribote interpretation B Trace shape... ee. see ees a4 SHERRILL AE AY MW Texture autributes 15 Curvature Io Examples of Present-day Workslows . re . Attributes used to generate sand probability volumes: 0 Time-lapse analysis 19 Attributes for detecting gas zones below regional velocity inversion 20 Chapter Summary... * sa see 2 References 2Table of Contents vil Mathematical Definition of Curvature ....4¢s¢s0escssevseseeeeeeetereveev eens - B Relationship between Curvature and Practures zs ‘Spectral Estimates of Curvature vcccccccceeececuevssvsvsesseeeeeeeeeeuseeunenes . 7 Volumetric Curvature: Calculations from Volumetric Estimates of Dip and Azimuth 8 Multispectral Estimates of Volumetie Curvatuce 79 ‘The Shape Index 8h Shape Components 87 Lineasnonts oo Examples Central Basin Platform, west Texas, USA Vinton Dome, Louisiana, U.S.A, Delaware Basin, New Mexico, U.S.A Chapter Summacy Rote i ateral Changes in Amplitude and Pattern Recog 9 Induce 9 Seismic Resolution and ‘Thin-hed Tuning =a 99 he She Ler wal Sensitivity of Variance and Semblance Measuces to Amplitude Vaciability Lot Hilbert Transform Estimates of Amplitude Variability 2 Bnergy-weighted Coherentamplitude Gradients 10s Second Derivatives of Amplitude: Amplitude Curvature eeeeerer 10 Pattern Recognition of Seismic Data: Texture im Spectral measures of texture 12 Sicuctural measures of texture 12 sical of toxics im GLCMs tor seismic dat Li Case 16 Case study 2 116 Other examples 19 Seismic facies application of testure attributes Chapter Sun References Chapter 6: Spectral Decompesition and Wavelet Transforms Seismic Resolution and Thin Relevant Co ved Toning ‘opts of Fourier Analysis ‘Spectral Decomposition Using the Short-window Discrete Fourier Transform 128 Examples 129 Visualizing a suite of speetal components 135 The Continuous Wavelet Transform and Instantaneous Spectral Analysis 136 Matched-pursuit technique: 139. Copyrighted materialTable of Contents Channels Paleotopography Structural control of channels Incised valleys and seourit Karst Libus field, Pearl River Basin, China Central Basin Platform, west Texas, U.S.A Kars! and collapse features in the Fort Worth Basin, Texas, U.S.A. Slumps Tor field, Norway (chalk) Vacuurn field, New Mexico, U.S.A, Faults and Fractures Future Trends: Comporer-aided Facies Extraction, Malampaya Field, Philippines Chapter Summary References Chapter 14: Attribute Expression of Deepwater Depositional Environments and Mapping Potential Drilling Hazards Introduction Mass-transport Complexes, Glide Tracks, and Outrunner Blocks Gas-charged Channels Circular Geologie Features Pockmarks Gas chimneys Shale-dewatering features Differential Compaction Iceberg Plow Marks Chapter Summary References Chapter 15: Mapping Reservoir Heterogeneity Introduction Reservair Characterization Workflaws To Estimate Porosity and Thickness Use of Attributes To Understand Reservoir Compartmentalization: Dollarhide Meld, Texas, USA, Geometric Auributes Correlated with Production from Llydrothermally Alt Appalachian Basin, Ohio, U.S.A. Use of Geometric Attrioutes To Map Freeture L Localization: Austin Chalk, Central Texas, US.A Dolomites, ion, Azimuth, and ‘libration of Geometric Attributes To Map Fracture Permeability: Kansas, U.S.A. Calibration of Pault-seal Capability Using Chemical Interwell Tracers, Niberta, Cansda Chemieal-traver-analysis results egtated interpretatio Time-lapse Seismic Survey, East Sehichallion Bield, West of Shetlands Chapter Summacy Referens xi 3x7 340 340. 32 M2 M2 33 346 48 38 349 a9 351 354 381 381 382 38h 392 303 309 401 402 403 404 408Digital recording of seismic data —— 1970 ‘Saisie eats in colar (2) Fest seismic atibutesinresuced sy ‘Concept of selsric tbls presented sts Dose 1 Seismic version inteduced LorgnLnsth) — —» Bright-spot technology | Energy crisis Over cstributes 3 Figure 1 of seismic artributes, and the attributres! time line of the developmen: relation t key advanees in seismic explo Madified from Barnes ration technolo, (20015 ‘Cemex trace analysis Complex vace analysis poset S60 Taw) ——af Campi ace ana 2d satigtapry — Stismeinvesion (Taper a Ste) 1 CGomple trace ana it Gt (lane eta Lt Response atibutes Response ates intatvcee asno) af! 1990 "| Crossloting of atibtes (as ‘tributes cles (Torr ata Paster recopritonlnewabnenork sna an atte Texture atibtesimrocuced Imuinroue analysis peks up bidding in the offshore Louisiana shelf and the Gulf of Mexico, Time/structure, velocit main the mnost fundamental attributes in use today, and seismic amplitude re Introduction of colar in seismic displays In 1971, A. H. Balch developed a computer graphics: photographic system, called the color sonogram, to display Ihe frequency spectra of seismic events simultancously with their time-varying waveforms (Balch, 1971) In sueh a display, the waveforms are presented using a conventional variable-area Scheme, but wilh the positive lobe colored 10 represent the changes in rock attenuation, or the loss of high Frequencies oquency components of the data, The latersl Proliferation of attributes interval atures Seismic stratigraphy and atvoutes lose oredity 20 evo edlecion and iieppresion emerge CCemeback ofattroute analysis Caherence atrioue introduced > ana Farrar Se thas nverson insocuced éecempeston, because of slight lateral changes in move-out yelocity and the like, could show up as color shifls on such displays. Baleh’s 1971 paper is credited with being the first publica tion in Groriysics to display seismic data in color. His, work heralded the beginning of an era in whieh color, with the enhaneed dynamic ra hat ivoffers. is used for mean ingful analysis of seismic data AL around the same time (1968-69), Nigel Arstey at Selscom Limited U.K. was working on innovative seismic displays and played 4 key role in introducing color on seis nic sections (N. Anstey, 2005, and personal communica- ton, 2005). Anstey and his team (Ron O'Donerty, Peser Judy Farrell, and later, Loyd Chapman) installed the first gray-scale laser plotter in their office in Londonstations, Interpretive workstations had (0 mitin advantages that beneficially influonced attributes. First, the use of color became pervasive and (unless you wanted a hardcopy!) economical. Second, calculation of a great many attributes hecame interactive. The benefit of that was more one of personal risk reduction rather than of speed. A daring inter- preter could simply try out an idea in the dead of night and show favorable results to his orher boss the aext d idea enhanced the map. Thus, the 1980s brought a rapid expansion in the seis- mic processing and display capabilities necessary for the plosion of attribute techniques that would occur in the middle 1990s, However, use of attributes actually decreased in the 1980s relative to that in the late 1970s, i the Attributes fall out of favor Complex-trace auribues suffer from waveform inter arising from nearby interfaces that can obscure subtle trends in the data. In particular, instantaneous fre- quency estimates can fall outside the seismic bandwidth and even can 9 Although a few workers understood that phenomenon and could use it as an Indicator of an unconformity oF a thin bed, the deleterious 1ce were not published erate negative values, consequences of wavelorm interfer Thus, interpreters who attempted fo associate physical mean- ing with such attributes were frustrated by artifacts, Inter- proters also found it difficult to relate those attributes di rectly to logged reservoir properties such as porosity, so the ‘tributes could not be used to quantify reservoir properties, As the 1080s passed. seismie attributes lost erodibility with interpreters. That loss probably was coupled with a loss of faith in seismic statigraphy as well, because numerous dry holes were drilled on the basis of seismic stra sraphie pre dictions. Following are three contributing factors suggested hy J.D. Robertson (J.D. Robertson, personal communica- sion, 2005) for the industry's disillusionment with attributes and seismic siratigraphy Limits fo resolution Given the limited resolution of seismic date available in the 19805, coupled with the lack oF geologic input to in- terpretation, interpreters lost sight of what seismic data re ally could resolve, compared with the stratigraphic resolu- tion they were seeking. Numerous interpretations of geologic detail simply were unjustified by the resolution of the seis mic data, When geologists attempted interpretations of seismic attributes, they often did not have a sound ynder- standing of the limitations of seismic data, and thei physicist teammates did not do a good job of educating ‘hem in the pitfalls of seismic resolution, Over cstributes 7 Tran on to 3D data ‘Thice-dimensional seismic surveying arose in th 1970s, and by the middle 1980s it had emery cial technique for imaging many onshore and offshore areas, around the world. It improved resolution enormously and ed w fewer dry wells. Even though it was considered &x- pensive at the time, 3D seismic interpretation proved to be much better at making successful exploration predictions ‘than seismic stra 1D seismic data was. ‘That dampened the ase af 2D seismic stratigraphic inter pretation of attribute sections. Use of attribute techniques, resumed only when workstation tools were develiped 10 apply the technology te 39 data carly “dats a benef phic analysis of Careles s drilling choices Afier the energy crisis of the 1970s and the accompa: ying rise in oil prices, in the early 1980s of companies, serambled 19 drill prospects and were not cursfil to drill only the good ones. In essence, exploration management allowed too many poor prospects 10 be drilled, and seismic stratigraphy and attribute analysis wok the blame for fail~ tures that really should have been blamed oa management's, poor judgment or overly optimistic predictions of the price of vil AL that time, other experts also voiced concems about the limitations of seismic attributes, many of which were summarized by Barnes 2001) Two-dimensional attributes Ry the middle 1980s, considerable improvements in schniques fad enkanced the infor- Imation content of seismic data used for stratigraphic interpre- tation, During that time, several b9o-dimensional continuity and dip attributes also were developed that were employed in procedures for defining and analyzing seismic facies (Con- ticini, 1984; Vossler, 1988}. Tian (1986) anticipated the need for 3D estimates of dip and azimuth by applying a 2D sem: blance estimate of apparent dip on surveys of 2D intersect ing lines, Although nove! and interesting, such procedures recording and processing did not evoke an enthusiastic response. "The results could be subjective, and 2D surveys simply contained too many arti~ facis from out-of the plane reilections. Horizon and interval attributes During the middle 1980s and later, horizon attributes (Dalley et al., 1989) and interval attributes (Sonneland et al., 1989) were introduced. Those attributes demonstrated that interpreted horizons not easily observed on vertical seismic sections. Areal varie hibited reflector characteristiesSeismic inversion revisited The original recursive or trace-integration seismic inyer- sion technique for acoustic impedance also evolved during the late 1980s and the 1990s, with developments in mode! based imersion, sparse-spike inversion, stratigraphic inver- sion, and geostatistical inversion providing accurate resulis (Chopra and Kuhn, 2001). Early techniques used a local optimization method that produced good results when pro- vided with an accurate starting model, Local optimization techniques were followed by global optimization methods that gave reasonable results even with sparse well contra Connolly (1999) introduced elastic impedance, which computes conventional acoustic impedance for finite angles of incidence. That technique was further enhanced by Whit- combe (2002) {0 reflect diferent elastic parameters, such ay Lame’s parameter 4, bulk modulus K, and shear modu- lus 4 Crossplotting of attributes Crossplotting of attributes was introduced to display the attributes’ relationship visually among two or three var lables (White, 1991). Verm and Hilterman (1994) used crossplots in AVO analysis, and such erossplots have been used Since then as AVO anomaly Indicators. When sppro- priate pairs of attsibutes are crossplotted, common litholo- ics and fluid types often cluster together, thereby provid= Ing straightforward interpretation. Also, mast rocks in many Tertiary basins are cither shales or water-saturated sandstones and lie along an easily recognized trend in the crossplot. Rocks that anomalously plot off of those two lnends may indicate the presence of hydeoestbons, thereby meriting further AVO analysis, Extending crossplots. to Overview of Seignic Atiributes 11 Figure 7, Time slices at 1.184 s through a (a) seismic vol ‘um and ils corresponding (b) coherence volume, Notice the clarity with whieh the individual channels show up on the coherence time slice three dimensions is henefiefal because data clusters hang ing in 3D space are more readily diagnostic and result in more accurate and reliable interpretations In Figure 10, we illustrate the use of modem cressplot- sing software for thee attributes that help identify a gas, anomaly —2-p on the canis, 2-9 on the y-axis, and fMuid stack on the zaxis. In Figure 10a, a blue patch indicates a gas anomaly on a time slice through the J.p volume. We then draw a red polygon on the time slice (outline) to select Tive data points to he displayed in the crossplot, The red cluster of points in Figure 10D corresponds to the red poly gon and five time slices (wo ahove and two below the one: shown). The yellow and magenta clusters are the corre Figure 8. Time slices through (a) scismie and (b) cuberenve data volumes. (¢) Overlay of volictence on scismie dat, Nowe thatthe coherence slice not only reveals faults with clarity bat also shows the intensively fractured region w the right. Afer Chopra (2002of approximately 10-20 waveforms (classes) that best ex- press the data, The most useful of such classifiers is based on self-organized maps, or SOM (Coleou et al, 2003). Un- like k-means and most other classifiers, SOM. algorithms order the classes such that similar classes are adjacent to each other ona 1D or 2D topology. By assigning a contin- ous ID or 2D color spectrum to those ordered classes, SOM Algorithms produce maps whose appearance is relatively insensitive to the number of classes. Although the results can be calibrated to well control through forward modeling, and although actual well classes ean be inserted, this tech nology is particularly well suited to a geomoxphotogy- measure the maximum and minimum bending of the surface at each point (Lisle, 1994) Note that our sign convention f° yin All Bpay FIOWS that used by Sigismondi and Soldo (2003), Those authors also showed that the maximum curvature allows one to easily determine the relative movement of fault Blocks from time slices. In Figure 2, we show reflector shapes as a function of the sign of the most-positiv hive curvature curvature and mostnega- Relationship between Curvature and Fractures Clearly, fractures occur when brittle rocks bend, How ver, the relationship between open fractures and our dif- ferent curvature measures is quite complicated and depends on lithology, previous faults and fractures, the paleostress regime, pore pressures, and the present day stress regime We feel confident that curvature maps (or volumes) derived from seismic data provide an accurate map of the present ay subsurface structure. If deformation was simple and the current stress regime has remained unchanged since the time Volumetric Cur ure and Reflector Shape 78 Of structural deformation, we ean use such maps predict copes we need frachites. Much more cammonly, however, 10 reconstruct the paleostresses and paleostrains through fone or more phases of palinspastic reconstruction, We also ‘may need to consider whieh faults and fractures have been diagenctically filled or altered. Ezeh stage of deformation adds a fabric’ or anisotropy to the rock, thereby making its subsequent sponse lo stress more complicated. Flecures in brittle rock cause fracture swarms (Figure 3). Hart et al. (2002) showed how production can be en= hhanced by faults and fractures associated with flexures in a ‘ight sandstone (Figure 4a-c). Specifically, they noted that the best-producing wells are correlated 10 intersections of feaures, such as those seen in yellow in Figure 4e. Inspec tion of Figure 4b shows that the flexures also would be well represented as zones oF high curvature, Lisle (1994) made one ofthe catliest correlations of re flector curvature and fractures and predicted that Gaussian curvature, Agayss (hich is positive For bowls and domes, negative for sudiles, and zero for planes, ridges, and val leys), shoul ‘0 open fractures (Figure 5). More recently, Hart (2002) found that strike curvature com strongly 10 open fractures over a producing field in north- western New Mexico, U.S.A. (Figure 6). In contrast, other ‘workers found that te dip component of curvature corre lates to open fractures in the Austin Chalk Formation of contral Texas, Ineach af these eases, variability in lithology and the presentalay stress regime deiermine whiet frac~ tures are open and which are closed, Such stress can be measured of inferred by borehole breakouts, hori zontal image logs, tracers, and production history. We will return to the use of production data to calibrate fracture prediction using curvature in Chapter IS. Although direct prediction of open fraetures using eur vature requires a significant amount of geologic unraveling and calibration through production data, eurvanure are @ powerful aid (© conventional structural and strati- Figure 3. (2) A map viewand th) a cross-section view of an idealized fracture sw with a feauze or faulted the trend oF more pers al. (2002), im. The fracture swarm is associated may be orienied regional fractures, Aner HFigure 11. The sirat cubes seen in Figure 10, disphayed here ata different angle and looking more like map views. Data courtesy of Areis Corporation, Marfurt and Kirlin (2000) calculated a crude measure of mean curvature (equivalent 9 equation 4.2 ifthe tangent to the reflector surface is flat such that a 0) using the formula ba=3(2+ 4) a) 4s well as tne rotation given by ja B, as) We apply equation 4.14 to the horizon-time-pick esti- mates of dip shown in Chapter 2, Figure 12a and b, and display the results here in Figure 15a, We note the very poor estimates of mean curvature in the two older portions of the merged surveys indicated by the dotted and solid white ree= tangles. Next, we apply equation 4.14 to the entire volume of p and g obiained using the multiwindow search technique deseribed in Chapter 2. Results of that estimation process were shown previously as a horizon slice along the Caddo Limestone horizon in Chapter 2, Figure 13a and b. We now plot the horizon slice through the mean-curvature volume in Figure 15b. Note the improved signal-to-noise ratio in the area of the two older surveys indicated by the dotted Volumetric Curvature and Reflector Shave 79 4) Figure 12. Curvedness calculated at different scales using apertures of [a) 3% 3,(b) 55, (c) 7 7, and (d) 9x8, respectively. In this example, the longer wavelength (lar aperture) estimates provide a clearer image of the fault-block Recall thatthe detinition of curvedness shows that it er Sigismondi ard is always greater than or equal 10 2er0, A Solio (2003), and Solid white rectangles. White zerows indicate Faults and. Fractures that we were able © carry through these areas and that we were unable to interpret in Figure ISa. In general, equations 4.14 and 4.19 provide somewhat noisier images of edges than those provided by structure coherence algorithms discussed in Chapter 3. Sinc Marfurt and Kirlin’s (2000) work, more-robust estimates of dip, structurally oriented ilk ‘multispectral curvature estimates (or the estimates of cur vature at different wavelengths) have provided significan- tly improved curvature volumes, eigen 2, and_most important Multispectral Estimates of Volumetric Curvature cour discussion of multispectal estimation of curvature by simply showing that geologic structures often have curvature of different wavelengths. In Figure 16, we show a zoomed portion of a vertical line BB’ that was, shown previously as Figure 10 in Chapter 2. Wehave drawn four black circles, each with a different diameter, to ap- proximate the curvature seen on the Marble Falls horizon, These circles have rai of curvature of approximately 00, 250), 100, and 50 m, Depending on our exploration objective, we may wish to generate separate maps that emphasize each of these different wavelengths. Tight, short-wavelength curvature may correspond 10 intease but highly localizedivature and Reflector Shave Volumetric a) ») Figure 20. A time slice, at = 0.8 s, corresponding to Figure 18 and showing most-negative curvature, fy. For spectral com ponents shown in Figure 175 defined by a = (a) 2.00, (b) 1.50, (1.00, (4) 0.75, (e) 0.5), and (£0.25. After Al-Dossary and Marfurt (2006), 8386 Seismic Attibutes for Prospect Identification and Reservoir Characterization pharmaceutical and genome work, and biometric indicators, ‘such as face and fingerprint identification. We show stich a biometric identifier in Figure 25, where Woodward and Flynn (2008) scanned 400 human hands using a laser dis scanning device and were able to identify a specific hand in question with 95% confidence. A close inspection Of the authors” hands shows a series of ridges, ruts, and flat spots exacerbated by the worry of typing this manuseript under a tight deadline. In Figure 26, we display a composite image of the shape index (rep sented by hve) and curvedness (represented by lightness) using a 2D color bar described in Chapter 9, The value c¢ = 0.25 enhances longer-wavelength features. Fig ure 26b, at r= 0.855 is at the same level as images displayed in Figures 18 and 20. Figure 23c, at ¢= 1.2 s,is atthe same level as images displayed in Figures 21-24. As we s the vertical slice shown in Figure 19, the deeper image here has more structural deformation and greater curvedness, and thus is somewhat brighter. Collapse features appear as shades of blue (indicated by the blue arrow) to cyan, and a few isolated domes appear as red (red arrows). The ridge indicated in Figure 19 is indicated here by the yellow arrow. 2 Faz Ss and (b) a range (what geoscientists would call a topographic) image of a human hand, (¢) is Figure 25. (a) An intensity imag Shape indices of the middle three digits, using equation 4.11. Look at y and valleys around your knuckles to thos: After Woodward and Flyan (2004 b) s ¢) 8 ¢ 4 o 8 a 8 Figure 26. (a) A 2D color table used to display the shape index against curvedness, Planar features will have a curvecness near bbe displayed in black. Highly curved features will be plotted asa pure color, beaween reflector shapes. We plot curvedness, c, and shape index, s, calculated using equations 4, 10 and 4.11 and using this 2D color table, at (b) = 0.8 s, corresponding to Figures 18 and 20, and at (c) r= 1.2, corresponding to Figures 21 to 23. Note that the shallow section is less comorted and yields adarker image, The cyan valley indicated by white arrows in (b) corresponds toa channel, The yellow ridge indicated by the black arrows corresponds to the ridge seen in Figure 19. Other cyan valley and xyellow ridge lineaments correspond to deformation about faults and joints. In contrast, (c) is dominated by collapse ihe Ellenburger Formation. Collapse features appear as blue eiseles— such as the one indisated by the blue arrow — ringed by yellow ridges and green saddles, Red arrows indicate two domal features, one of which corresponds to the dome indicated by the gray arrow seen in Figure 19, After Al-Dossary and Marfurt (2005) our own fingers and try to mateh the dota sof ridges ze10 and us wo dif hereby allow atures inShape Components ne goal in seismic attribute analysis is to statistically conrelate & given attribute with well logs, production data, or features of interpreiational interest. Whereas the Iuuman mind can readily identify the shapes displayed in the eon- posite image shown in Figare 25, 2 computer may lack that ability. Iust as dip ing only if the dip magaituse is nonzero, the shape index has interpretations value only when it is used with eurvedaess or some other measure of the intensity of deformation. In principle, non- linear neural networks should be able to comrelate multiple tributes with validation data. However, cther powerful sechniquss, such as geostaisties, may not. Because we in iuitively expect correlations between domes twith four-way closure) and hydrecarhon accumulations, or between howls and greater aevonimodation space for send deposits, of be tween a particular shape and open fractures, we think its valuable i generate shape components Such as those shown jgure 27 that correspond to the sare time slice as Fig vrimarth has mm Figure 27. Long-wavelength (a 0.25) curvature anomalies at 1= 1.200 s, decomposed according to shape. First, we Volumetric Curvature and etiectorShace 87 ture 26, All we have done is multiply the curvedness at every paint in the volume by a simple Filter (shown in Fig lure 27f) that passes a shape component of interest lineaments A bumuan interpreter can readily identify several linea. ‘meats in Figure 20f. By using the shape index, curvedness, and azimuth of maximum curvature (perpendicular (© the strike of Tinear features), we ean both enhance and count she number of pixels that have a significant line ponent, In Figure 28, we display the filter that is bssed om, the shape index and that will pass features that have a shape best represented as a ridge or valley. In Figure 28b, we dis play a 2D color table in which the strike oF the ridge or val- ley (perpendicular to the azimuth of maximum curvature given by equation 4.12) is represented by nue, and the cury= edness multiplied by the shape index filter is represented, by lightness. In Figure 28, we display the composite image comesponding ta the time slice Bs and @= 0.25, Jae the corvedness using equation 4.10. Nest, we multinly the eurvedness by a filter designed to passa particular shapes (8) bowl, (b) valley, (e) saddle all he f lores that may core: forms Unat are based on a weighted linear sum of atributes. After Al-Dossary and Ma (d) ridge, and (e) dome. Corresponding contained in these five images also is contained in Figure 26c. However, tersare shown in (). In principle, all the information tecause the sum of @ bow] ard valley image includes 1nd a thicker accommodation space, we expect better performance irom statstieal tans furt (2006. Copyrighted material88 Seismic Attibutes for Prospect Identiication and Reserveit Characterzation Figure 28. (a) Generation of a lineament attribute, obtained by applying a fier that passes only those values of eurved ness that correspond to ridges and valleys. (b) A 2D calor bar to display the lineament attribute against curvedness. (c) A time slice at After Al-Dossary and Macturt (2006 1.2 s through the Hineament attribute volume. that was shown in Figure fault (strike of 90° and colored yellow) runnin OF, Note the strong east-west through the + of the image. Magenta lineaments strike o the north east, green lineaments strike to the northwest, We do not ay significant north-south- trending Fineaments in blue. Examples Central Basin Platform, west Texas, U.S.A. Our first example is from a survey over the Central Basin Platform of west Texas, U.S.A. The major hydrocar ton production in this area is from the Devonian Thirtyone Formation, which is a chert deposit carried by turbidity flows from the shelf in the north into tls deeper part of basin. The reservoir is highly compartmentalized and is enhanced by fractures. (In Chapter 18, we present a case sudy of the same formation imaged by a neighboring sur; vey.) In Figure 29, we show an image of the mos.-positive curvature extracted along the yellow Thirtyone Formation horizon posted on the vertical slice through the seismic data, In Figure 29b, we display an enlarged view of the scismic data corresponding to a producing part of the reser voir indicated by the green box in Figure 29a, Green arrows indicate synclinal and red serows anticlinal features in this structural high In Figure 30, we display the time-structure map of the hirtyone Formation pick shown in Figure 29a, well as the corresponding coherence extraction (horizon slive through the coherence volume). In Figure 31, we show corresponding horizon slices through the most negative curvature and most-positive-curvature volumes. Green ar rows (indicating synclines) and red arrows (indicating ant. lines) correspond to those shown in Figure 29b. Note that this subtle warping seen on the vertical seismic image can be carried along the eatire horizon, providing constraints fon the palcostiess envitonment and on possible fractures The coherence extraction is relatively featurele the zone of interest, whereas the curvature yolume is not This observation reinforces our major point that coherence and curvature volumes are different because they are mea suring different attributes of the input seismic volume. In particular, curvature shows subile (unbroken) flexures not seen by coherence because coberence is sensitive only to lateral discontinuities In contrast, we will show images in Chapter 12 (on elas: tie depositional environments) in which coherence will de Iineate channels and (in the absence of differential compac on) curvature will not The curvature computations are volumetric rather than along a surface, which we illustrateVolumetric Curvature and Reflector Shave 89 Figure 29. (a) A vertical slice through the seismic data and a horizon slice along the yellow Thirtyone Formation, from the most positive curvature volume, (b) Enlargement of the seismic data shown in the green box in (a). Green arrows indicate syn- clinal and red arrows anticlinal features within this structural high. Note how the flexures seen on the seismic can be ¢ through the volume on the horizon slice through the most-pesitive curvature volume, Data ccurtesy of Burlington Res Figure 30, (2) A time-structure map and () Figure 31. (a) Most-negative-curvature and (b) most- zon slice through the coherence volume alon positive-curvature extractions along the yellow Thirtyone Thiryone Formation pick shown in Figure 29a, Formation pick shown in Figure 29a and corresponding to the images shown in Figure 30. Note the added detail pro- ‘vided by the curvature volumes in areas where the coherence is featureless90 Se&smic Alibules for Prospect Identification and Reservelr Charac! in Figure 32, Figure 32 displays time slices at = LOS through the most-neyative-curvature and mast-positive- curvature yolumes; the posted yellow picks correspond 10 the intersection with the structurally deformed Thirtyone Formation. To illustrate more explicitly the correlation of these curvature computations to the original seismic data, ‘we display the seismic data, coherence, most-negative eur- ‘ature, and most pesitive curvature in Figures 33 and 34, ‘These curvature images give the interpreter a means of mapping local highs (domes) and lows (howls) and of car: tying subtle flexures across the Vinton Dome, Louisiana, U.S.A. In Figure 35, we show a time slice through (a) a coher cence volume and (b) vertical slice through the seisinie am plitude volume aequised over Vinton Dome, Louisiana, USA, ‘We note that several ofthe faults easily interpreted on the seis a) Figure 32. A time slice at 11s through (a) the mostmega- uiye-curvatace volume and (bythe mostpostive-curvatare volume from Pigure 31, The yellow picks correspone to Inersestions of the time slice withthe Thistyone Formation pick stown in Mpure 29% Polis and Mexures preted on these ‘ime slices prior to picking any horizons. ibe inter sation inie section in Figure 3b (indicated by arrows) do not show up clearly on the have been prestack time migrated, the velocity structure is sufficiently complex that the fault edges are blurred Ge. overmigrated or undermigeated), with residual hyperbola: rendering the faults more continuous and less detectable by coherence measurements. Short-wavelength estimates of cur ‘vature see these poorly migrated edges as subtle flexures and provide marginally improved images of the faults (Figure 36). In contrast, long-wavelength estimates of curvature pro: ‘ide stoong, continuous fault plane images, partially compen ‘sting for the inaccurate migration (Figure 37), In addition, the long-wavelength estimate of rotation in Figure 37¢ shows several faults that have a seissorlike rotation about diem, thereby defining a nonquadratic surface that was not defined by conventional curvature measures. coherence volume. Alkhough these dats Figure 33. Cube-cdge displays solded at |.0 slong line AA’ from Figure 30b, showing seismie data on the vertical section and (a) seismie and (b) coherence data on the time slice, The zone of intersection is relat y featureless on the coherence sliceDelaware Basin, New Mexico, U.S.A. As our last example, we retumn to the survey over the Delaware Basin, New Mesico, U.S.A., that we presented earlior in Chapters 2 and 3. In Figures 38 and 39, we dis- play time slices through the most-negative-curvature vol- ume (Figure 38) and the most-positive-curvature volume (Figure 39), which correspond to images through the volu metric dip and azimuth volume displayed in Chapter 2, Figure 20 and the coherence volume in Chapter 3, Figure 37. These curvature ealeulations are made directly from in formation displayed in Figure 20 of Chapter 2. The most- positive-curvature time slice at.8 s in Figure 39a shows a Fexure that clearly delineates the platform edge (yellow arrow). This platform edge appears to be at a more-north= erly position on the 10-8 time slice. We clearly delineate Figure 34. Cube-calge displays folded at 1,0 slong line ‘AA from Figure 30, showing seismic deta on the vertical section and (a) most-negative curvature and (b) most-posi tive curvature on the time slice. We can easily track folds 1nd floxures seen on the seismic image into the survey along the curvature slices Volumetric Curvature and Reflecior Shave 91 the platform edge, ‘The east-west Hexure indicated by the blue arrow shows up prominently an the deeper most-posi tive-curvature slices and is somewhat offset to the north on the deeper most-negative-curvature volume, The broader, rmose-dilfuss, nomh-south flexure indicated by tre may arrow does not show up #s clearly as the east-west flexure, indicating that we need t use a Tonger-wavelength curva ture operator The comples areas of folding indicated by the cyan and orange circles (and discussed in Chapter 2) ap- pear asa patter of spatially periodic flexures, Figure 40 displays vertical slices through a composite image of the most-negative-curvature volume and the seis- mie data volume, and Figure 4 shows vertical slices through 44 composite image of the most-positive-curvature volume ») a8 spe) GEM Figure 36. (a) Atime slice at |.0 volume and (h) vertical slice through the seismic data irom 1 survey acquired over Vinton Dome, Louisiana, U.S.A. The arrows correspond to the sume faults shown in Figures 36 and 37. Note that some of these fauks appear to be diserete ‘ough a coherence discontinuitios, but others appear to be folds, either result ing from drag al Faults having a folded appearance show up more clearly in long the faslt, or from inaceuzate migration the curvature image than in the coherence volumes. After AL Dossary and Mariurt (2006).92 Seismic Attibutes for Prospect Identification and Reservoir Characterization Figure 36. Short-wavelength (= 1.00) estimates of (a) most-negative curvature hyeg, (b) most-positive curvature pale? pos and (c) reflector rotation, r., corresponding Wo the 20.10 same time slice shown in Figure 3a. After Al-Dossary ‘and Marfurt (2006). 00 0.10 b) sin? sin 0.40 1050 09 00 10.10 050 Figure 37. Lonz-wavelensth 1 =0.25) estimates of (a) ost-negative curvature, Kg (b) most-postive curvature, py and (€)feflector rotation, 1, corresponding Wo the same Lime slice shown in Figures 35a and 36. White arrows depict faults that are not seen as clearly on the coherence time slice shown in Figure 35b. Gray arrows correspond to Faulls where reflectors have significant rotation about the fault plane. Such rovations are not represented by the quadratic surface used in curvature ealeulations, The black aerow indicates noisy, low. fold areas ofthe data. After ALDossary and Marfart 2006). ») ° stn? 025 | 00 +025 +028 00 =028 10 00 10and the seismic data volume. Here, we note that the folded structures have sharper erests and broader valleys, resulting in the spatially periodic, most-positive-curvature image Volumetric Curvature and Reflector Shave 93 in Figure 34e and f, In contrast, the ‘wavelength synelines appear to be more diffuse on the eor- responding most-negative-curvature slices. entler and longer Figure 38. Time slices through the most-negative curvature at(a) 08, (b) 1.0, (6) 1.2,(@) 14, (e) 1.6, and (F) 1.8 s,cor- responding to images shown in Chapter 2, Figure 20, and Chapter 3, Figure 37. We have etained the same arrows shown in Chapter 2, Figure 20 t aid in the comparison,94 Se&mic Aitioules for Prospect Ideniffcation and Reservoir Charactervation 5 ° 5 Figure 39. Time slices through the mos-positive curvature at (a) 0.8, (0) 120, (©) 1.2,(@) 14, (€) 1.6, and (6) 1.8 5, corresponding to iniages shown earlier ia Chapter 2, Figure 20, Chapter 3, Figure 37, and this chapter, Figure 38 We have retained the same aartows shown in Chapter 2, Figure 20, to aid in the com- parison, Copyrighted mVolumetric Curvature and Reflector Shave 98 Figure 40. Vertical slices through a composite image of he most-negative curvature and the seismic data, corresponiling to Chapter 2, Figure 21 and Chapter 3, Figure 28, Line locations are indicated inthis chapter, Figure 388, ° Tine (6) Copyrighted material96 Sesmic Altibules for Prospect Identification and Reserveir Choracterzation Figure 41. Vertical slices through a ‘composite imaige of the most-pos tive curvature and the seismic data, corresponding to Chapter 2, Figure 21 and Chapter 3. Figure 28. Line locations are indicated in Figure 39a. Copyrighted materialChapter Summary This chapter shows well-established methods for gen- erating curvature measurements from interpreted horizons applied © volumetric estimates of dip and azimuth. Curva- lure calculations performed directly on the dip and azimuth cubes provide goud edge-detection images for high-quality data. However, the true power of this technique is achieved by calculating curvature of broader featares hy altering the wavelength of the caleulation. Longer-wavelength estimates oF curvature can be obtained by successively smoothing maps or dip und azimuth volumes, by explicitly fikering the di in the k-&, wavenumber domain, or by explicitly filtering the curvature aperators. These later two approaches are mone general than simple smoothing, thereby allowing the inter preter to explicitly define wavelengths of interest. In addi- tion to enhancing longer-wavelength flexures, olds, and compaction Features, long can improve our ability to map poorly focused faults that are contaminated by residual dilfraction hyperbolae Volumetric estimates: of reflector curvature and rota 1h estimates of curvature tion are hased on, and are only as good as, the volumetric stimates of reflector dip and azimuth discussed in Chapter 2. In contrast, volumetric estimates of curvature and rote tion are independent of the coherencs in Chapter 3. The different images si coherences are equally dependent ematies. Certain features, such as well-focused normal faults, show up equally well on coherence and curvature volumes, However, in the absence of differential compac- tion, curvature attributes provide poor images of ch In contrast, curvature provides excellent images of subtle Flexures, folds, and collapse features that are not commonly measures discussed en by curvature and fon geology and on mathi= seen on coherence volumes. Reflector rotation can be a good indicator of data quality as well as of scissor move- ment along a faut. The mostnepative and most-positive curvatures are the most unambiguous oF the curvature images in highlighting faults and folds. Several authors have shown a strong cor tion between Gausan curvature, dip eurvsture, and strike curvature, and the presence of open fractures. Currently, such calibration work using image logs, production history, trac- er dita in conjunction with physical and nurmerical model, geostatistics, and neural nets is an active area of research, Wwith Fischer and Wilkerson (2000) relating the orientation of fractures to fold shape and Stewart and Wyan (2003) pre dicting strain from spectral estimates of curvature, The shape index, when it is corendered with curved ness, allows us to visualize 3D reflector morphology on simple time slices, Shape indices have proven to be a valu- able too] in facial recognition and olher biomarker technol- as well as with molecular docking analysis. We fully Volumetric Cur ure and Reflector Shape 97 ‘expect shape indices or a successorbased on curvature mea sures to form a key component in future computer-assisted aphy analysis. If yolumetrie estimates of curvature have a major Timi- tation, it is in the restrictions applied to the calculations themselves. The vector dip is a true 31) calculation centered, fon the analysis point that would follow the best reflector as, it ascended and descended in time or depth. However, be cause we wish to calculate curvature even when only piece- culate the wise-continuous reflectors are available, we ci derivatives of dip on constant-time time slices (or constant depth depth slices). For steeply dipping horizons and/or long-wavelength estimates, we expect that these measures, undesirably mix geology of different formations. Under such, conditions, predictions of fracture intensity are linked not tw thin-plate theory, but rather to some more-complicated composite-plate deformation, Although we anticipate there will be near-term improvements in our estimates that will betier follow the local dip, we know that such improve iments will not be Wivial to implement References AL-Dossuny, S.,and K. J. Marfurt, 2006, 3-D volumewic multispectral estimates of reflector curvature and ro- tation: Geophysics, 71, 41-51 Bergbauer, S., T. Mukerji, and? Hennings, 2003, Im= proving curvature analyses of deformed horizons using scale-dependent filtering techniques: AAPG Bul- letin, 87,1 Cooper G.R., snd D.R. Cowan, 2003, Sunshading geo physical data using fractional order horizontal gradi- ents: The Leading Bilge, 22, 204-205. Dalley, R M., EE, A, Gevers, G. M. Stampli, D.J. Das vies, C. N. Gastaldi, P. R. Ruijetnberg, and G. J D.Vermeer, 1989, Dip and azimuth displays for 3D seismic interpretation: First Break, 7, 86-95. Briesson, 1. B., H.C. MeKeon, and 8. J. Hooper, 1988, Facies and curvature controlled 3-D fracture models in « Cretaceous carbonate reservoir, Arabian Gul, G. Jones, QJ. Fisher, and RJ. Knoipe, eds., Fault- ing, fault sealing and fluid flow in hydrocarbon res crvoirs: Geolog: lication 147, 299-312 Fischer, M. P., and M. S. Wilkerson, 2000, Predicting the ‘orientation of joints fiom Told shape: Results of pscudo~ three-dimensional medeling and curvature analysis, Ge ology, 28, 15-18 Hart, B. S., 2002, Validating seisinie atuibutes: Bey statistics, The Leading Fdge, 21, 1016-1021 Hart, B.S.,R. A. Pearson, and G. C. Rawling, 2002, 3D seismic horizon-based approaches to fracture-swarm, al Society of London Special Pub: ond98 —_Se&mic Aitibules for Prospect Identification and Reservalt Cheractervation sweet spot definition in tight-gas reservoirs: The Lending Fége, 21, 28-38 Lisle, 8. 1., 1994, Detection of zones of abnormal strains in Structures using Gaussian curvature analysis: AAPG Bulletin, 78, 1811-1810. Luo, ¥., S. ALDossary, M. Marhoon, and M. Alfaraj, 2003, Generalized Hilbert transform and its applica. dge, 22, 198-202. sophysies: The Leadi Marfurt, K,J, and R. L. Kirlin, 2000, 3-D bmadband es. timates of reflector dip and amplitude: Geophysics, 304-320, Mitsova, H., and J. Holierka, 1993, Interpolation by re- gionalized spline with tension I: Application to ter metry anillysis: Math rain modeling and surface ematical geology, 25, 657-669. Roberts, A., 2001, Curvature attributes and their applica tion to 3D interpreted horizons. First Break, 19, 85-99. Sigismondi, E, M., andl C, J, Soldo, 2003, Curvature at tributes and seismic interpretation: Case studies from Argentina basins: The Leading Edge, 22, 1122-1126. Stewart, S. A,, and TJ. Wynn, 2000, Mapping spatial variation in rock properties in relationship to seule dependent structure using spectral curvature, Geol ogy, 28, 691-694. Wood, I. D., 1996, The omorphological characteriza ton of digital elevation models: Ph.D. thesis, Uni versity of Leicester Woodwanl, D. L., and P. J. Plyan, 2004, Finger surface aya biometric identifier: Bion ference, Arlington, VA. Wyna, T.1., and 8. A. Stewar curvature mapping characterizing subsurtace strain dis tributions, 7 M. Ameen, ed. Fractures and in-situ stress characterization of hydrocarbon reservoirs: Geolog ical Society of London Special Publication 209, 127- 13. srie Consortium Com 2003, The role of spectralChapter 5 Lateral Changes in Amplitude and Pattern Recognition + predict which geal + apply best practices for stratigraphic interpretation + identify channels and other thin stratigraphic Features on amplitude Chapter Objectives Aner reading 1 YOU Will be able To + interpret lateral changes in amplitude that occur because of Hhin-bed tuning slic im: features ean be soon est by amplitude gradients, curvature, and coherence altsibwtes + use gray-level en-occurrenee matrices to quantify seismic textures for visual and computer-assisted interpretation Introduction In Chapter 2, we examined how dip and azimuth can quantify lateral changes in reflector time (or depth). In Chapter 3, we examined how coherence can quantify laters .e8 in reflector waveform. In the first section of this ‘we will examine the third family of geometric at- that measare race to-teuce variation — those attr butes that are sensitive to lateral changes in amplitude. A primary goal ofthis book is to clearly define the phys ical and mathematical basis of the seismic attributes com monly used in interpretation, so that we ean relate them bet- ter to geologic lithology and fluid properties via geostatistics We also will show that many of the mathematically in dependent attributes, some of which we have discussed al: ready and some that we introduce in this chapter, may be linked through the underlying geology, Whereas any one of these three Families of attibutes may be seusitive to thick channels and faults, their responses to subtler features var} Ins particular, many thin, Tithok charged reservoirs can be characterized by strong reflection events that have a constant waveform. Such hetemgeneities are not seen by crosseorelation or jcally heterogeneous of echerence, but they are expressed thecuth subtle changes in waveform amplitude. Likewise, subtle changes in thin-hed sickness also cause variations in Waveform amphiuide 99 Geometric attributes that estimate dip, azimuth, curvae ‘ure, and energy-weighted coherent-amplitude gradients (dis cussed in this chapter) provi coherent reflectors. Coherence {tributes provide excellent images of diserote discontinuities, hut do not diseriminste between h y low-colierence zones [such as a mass transport complex (MTC) or a slump] and low-energy low coherence zones (such as often oveur with shale-on shale reflections) e address these others kinds of -xtures in the second section of chapter, This latter type of analysis Teads us into the area of statistical pattern recognition, which, along «with the morphological pattem recognition discussed in Chap= ter5 2 through 4 (angular unconformities detected using dip and azimuth, meandoring-channel edges using coherence, etc.) provides the building blacks for compater-assisted in- ‘etpretaion Seismic Resolution and Thin-bed Tuning ‘To potter understand the effect of hed thickness on seis mic amplitude, we examine the simple wedge model shown in Figure 1. In this fi mped- ance layer encased in a higher-impedance material sueh as, ture, we assume a simple low100 ‘one we might encounter in a sand chaanel that cuts through, ‘and is later buried by shale. The impedance changes in Fig ure 1a generate a negative reflection coefficient at the top and an equal but opposite reflection coefficient at the bot tom of the wei (Figure 15). We generate the synthetic seismogram in Figure Ie by convolving these two reflec tion coefficients with a band-limited wavelet, As the wedge thins toward the left, the positive side lobe of the m reflection from the top constructively in rerferes with the positive main lobe of the reflection from the bottom. Simi- larly, the negative side lobe from the positive reflection Figure 1. (a) Thin-hed resolution and the classic we model of a low impedar ance mattis. (b) ases in thickness from 1 layer encased ina higher imped cflectivity and thickness: the wedge o 50 ms from left 10 right, tion on the top and a positive reflec tion on the bottom. (©) Seismic synthetic data generated by convolving (b) with an 8-10-4)-S0-Hy Ormsby filter. (4) Instantaneous envelope (also ealled reflection st puted from the seismic data in (¢), The peak enetg the one-quatter-wavelength (one: period two-way tay chime) tuning thickness indicated hy the white aerow Below he tuning thickness (which is indicated by the green ellipse), only the amphtude enanges, se and frequency remain fixed. Such subtle changes in amplitude allow us to m J/changes in bed thickness using lateral in amplitude vanity Gand es we discuss in at faer 96 inthi (2001 book, using spectral decomposition). After Partyka mic Altrioutes for Prospect Identification and Reser i Characterization from the bottom constructively interleres with the negative main lobe of the reflection from the top. The maximum constructive interference occurs when the wedge thickness is one-quarter of the eflvetive source wavelength, or, when measured in swo.way traveltime, the wedge thickness is ‘one-half the thickness of the dominant period (indicated by arrows in Figures [e and 1d), For thicknesses smaller than this, the waveform's shape stabilizes (Figure 2a) and only the seismic amplitude changes with thickness. Widess (1973) showed thal when we are well below the one-quar ter-wavelength (tuning) thickness, omplitude changes lin early with thickness, as indicated by Figure 2h and as de: fined by ar. ROULT 1 2)0(1)— RAG—T Dwi RT. G1) where R is the reflection coefficient, w(i) is the seismic wavelet, and 9 (F+/2) and 6 (t—T/ 2) represent spikes at the top and bottom of the thin layer of two-way traveltime thickness, T. Many rewlers n the finite difference approximation to the Fist derivative, One well-nown property of Fourier transforms is that if recognize equation 5.1 as wine We) 2 denotes the Fourier transform pair, where «is the temporal requeney, then Wo), 63) From equation 5.1, we see that for thin beds, amplitude Tinearly with thickaess cha In addition, the low fre quencies are attenuated and the high frequencies are ampli fied (determined by the e term in equation 5.3), whereas the phase is rotated by 90° (the ‘term in equation 5.3), thickness, we do not expect to detect any significant changes in coherence using algorithms sem sitive only to waveform (and aot © amplitude), such as Ue rosscorrelation algorithm given by equation 3.4 and th structure algorittim given by equations 3.15 and 3.16. Be: cause these particular coherence algorithms are insensitive to amplitude changes, they will not detect lateral change: that occur in lithology within a thin bed below the tuning thickness and that do not cause a polarity reversal, such as may occur in asand-filled thin channe! cutting tnrough a shale matrix. In contrast, we do expect such subtle features to be detectable by attributes sensitive to lateral changes in am plitude. That motivates us to consider atteibutes that will be Sensitive to such changes,Lateral Changes in Ampliude and Pattern Recognition 101 The Sobel Filter The simplest amplitude-sensitive multitrace attribute is the Sobel filter, or spatial firstderivative operator, which tn a seismic time slice would have the form ‘faeil ult.ry-+ Ay) ul 2Ay Equation 5.4 is used routinely in photographie digital image processing, By adding three steps, Luo et al. (1996) modified the Sobel filter to make it applicable to seismic data, Firs, they took the derivatives along a plane defined by reflector dip and azimuth. Second, they vertical analysis window to improve the signal-to-noise ratio, Third, they normalized the result hy dividing the dif, such that they ached those derivatives alon, ferences by the energy of the input traces, mneasure relative changes in amplitude rather than absolute notation used in equations 3.1 and 3.3, changes. Using Jn which the analysis trace trace is) and thenext adjacent crossline trace is tain Luo et al.'s (1996) original edge-detection algorithm, is uy, the next adjacent inline we ob- elt yt, )= |S fowezan wee kare tluiebean-wiettarr, }) {5 tnc-tane tuyertéar—r 91 + [uber il} where 7, andr, ate the delays associated with the reflector {ip in the inline direetion, x, and the erossline direction, » As with the crosscorrelation algorithm discussed in Chap- ter 3, the edge computation given here by equation 5.5 is offset from the target trace in the analysis. In Figures 3 and b we display one of Luo et al.’s (1996) first publishes -quation 5.5 and note that Figure 3b Iooks re- mmarkably like the semblance-based and variance-based co: results using herence algorithms. Although we suspect that Chevron has improved on this algorithm, we have not seen any details in the pub lished literature. However, we display the results of the present-day implementation as Figure 4. In Figure da, we show a dip magaituds map generated from the water bot tom pick fora survey acquired over the western Niger Delta slope at water depths of about [300-1500 m. The sealToor is ruptured by feulting e ‘and diapirisr, ansl has mud digpirs and mud voleanoes (Ade- ogha et al., 2005). Two mud diapiss are Seen. One is on the related to avity-driven extonsion ‘west central part of the image, which also has a concentra tion of mud voleanoes at its crest and a few more an its eastern flank, An eastern mud diapir is barely visible be- cause most of i€ falls outside the study area, Figure 4b is a horizon slice 250 ms below the water bottom through an attribute cube generated using a modem implementation of the Sobel filter defined by eyuiation 5.5. In this image, we are able to observe fault pattems, channel forms, fan geometries, mud mounds, and ges-esrape craters ‘on diapirs, channel Kaickpoints, mudflows, slumps, bypass incisions, and large-scale scours. These nearly as well defined! on conventional displays such as in- lings, erosstines, and time slices from the seismic volume, eatures were not Sensitivity of Variance and Semblance Measures to Amplitude Variability Semblance and variance definitions of coherence pend on lateral amplitude changes, whereas the eigenstruc~ ture and crosscorrelation definitions of coherence do not. IF a) Tamasieen oo 4 Figure 2. ‘The sensitivity of (a) rellector trough-to-peak thickness and fh) the trough amplitude For the simple wed mode! shown in Figure 1. Note th sion of Ioagh amplitude with thickness for small values ot th 10 ms). After Partyka (2001 in turn after Widess (1973). ness102 Seismic Altioutes for Prospect Identification and Reservoir Characterization ‘we Substitute the classical definition of variance for the computational definition of variance given by equation 3.7 in Chapter 3, we obtain S13 (RAI psa) wtp] SS a crsar ps0] elspa) 6.0) 1(t,p4) is the mean trace slong the reflector dip and ‘azimuth dotined by the apparent dip pair, (pp Clearly, if cach trace is identical along the reflector dip and azimuth, then each of those traces is also equal to their mean, such, that the vatiancs estimate of coherence, cy given by equa tion 5.6, is zero, In contrast, if the waveform of each trace is identical but the amplitude of that waveform for each trace varies (¢-2. ifone tae ean he represented by a 30-He Ricker wavelet of peak amplitude 1.0 and an adjacent trace can be represented by a 30-Hz Ricker wavelet of peak am plitude 1-1), then the variance estimate of cok zor. Because by equation 2.10 we showed that the sem lance and variance estimates of coherence are directly related to each other (with c, = I~ ¢), we could state that the semblance estimate of coherence also is sensitive to amplitude. In contrast, the erosscorrelation and eigenstruc~ ture estimaves of conerence given by equations 3.4 and 3.13 only to changes in waveform, not to changes are sensit in amplitude. Hilbert Transform Estimates of Amplitude Variability Implementation of the Sobel Blt SA uses omly the adjacent traces and produces what we lven by equation Figure 3. (2) A cedee-detection fils fated along rel seismic time slice and (b) a corresponding tribute obtained on the hasis of the Sobel jon 5.5. Amplitude de dip and azimuth, After Luo et al. 1980). atives were ealea- called a short-wavelength derivativ Our processed seismic d arid of infines and erosslines separated by distances Ar and Ay. Clearly, we could have binned our data at some coarser cr finer grid and applied the Sobel filer defined by equa, tion 5.4. In Figure 5, we display four such possible differ tence operators, where the grid operator spacing corresponds to 25 m, 75 m, 125 m, and 175 m, Note thet because we estimate in Chapter 4 1 heen binned into a regular sees Figure 4, (a) A conventional dip-magnitude map genersied from water-bottom picks. (b) A horizon slice, 250 ms below the water bottom through ube generated using & aration of eqaation 5.2. After Adeogba st al. 2005), AAPGO20DS. Reprinted by permission of the AAPG whose permission is requited for further use.Lateral Changes in Amplitude and Pattern Recognition 103, divide by max in equation 5.4, the magnitude OF the operé~ length that is much longer than any of the distances used in tor dbereases in each image the difference operaors, stich as in the ease displayed in In Figure 6, we display graphically what such opera Figure 6b, all of the derivative estimaies will be identical tors do. Many reaiers will recall fondly their pleasant days If our horizontal wavelengths are long and our data are in caleulus class, where a derivative was defined 28 the noisy, we may consider stacking the results of all these de- limit, as Ax approaches zer0, of the chord (ie, the secant) rivative estimates to obtain a more robust estimate. joining two points on a curve, which we plot in Figure 63, Mathematically, the éerivative operator given by equa- As Av bovomes smaller and smaller, the slope ofthe chord tion 8.6 is a Tincar operation, which means we obtain the more and more accurately represents the slope of the curve same resull if we stack the operator hefore applying it to being analyzed. If the curve being analyzed has a wave- the data or if we stack the results of each operator after ap- PERAOOE EDS A PIEDSSFAPESE LEE SA maxes maxa7s PRORPREPEPOSOLIS game PEEDEREROSOOP ODS pay msx= 105 mana 178 five defined as a wetman) Figure 5. Alternat weomae), where Ax isthe seismie bin size and m is the number of bias away from the center analysis point. finite-difference approximations to the first deri a a 8 1 Hae SPRPARID OER ED we PEROREPP ESOP ELIS Figure 6. Graphical display of the approsimations made to the first derivative (or lope using the operators shown in Figure 5 for (a) anamplitude variation that has as moderate wavelength, on the order of 10 scisiie bins, ad (5) an amplitude variation that has a very long wavelength compared with the seismic bin size At. For very long wavelengths, each of the altemative finite difference approximations provides the same estimate ofthe frst derivative, Copyrighted material104 Seismic Altioutes for Prospect Identification andi Reser plying it to the data, Stacking the operator gives us. |: (5) which we display in Figate 7 By now, the astate adr may recognize that we have built up a Hilbert transform peretor from a suite of Finite difference operators. Unlike the Hist ransorm operator used in conventional complex-tice analysis, we Wish to apply the operator in Figure 7 along the horizontal axis (across traces) rather than along the vertical axis (actos samples) Luo et al. 2003) recognized that many statgraphie features do not have abrupt edges. but steal are repre- cated by changes that have afinte wavelength, sub asthe amplitude variation over the idealized channel shown i Figure 8, The conventional Sobst filler produces a 008 image of the euthark channs! cdpe t the right of Figure 80, bu! a poor image of the point bar channel edge to the lef In conmast io ths, a Figure Se we show how the Hi bert transform filkerimages both sides of the channel equal ty well, although the steep eutbank Iooks overly smeared Applicaton of the derivative operator given by equation 5.7 enkanees the lorg-wavelonph variation an diminishes the short-wavelength variations, as we observed in Figure 6. Luo et al. (2003) added an additional twist. Instead of aiding each term equally, fist they raised the absolute Value ofeach tm to the power r, summed the results, and then normalized by taking the rth root, The reader il re. call these kinds of rmorm mewsremen's when we dis essed the Manhatian distance estimate of coherence in Chapter 3, “The result ofthis weighting i similar to the Fractional decivatives presented in the discussion of curvature in Chapter 4— the seismic interpreter ean now view his o an Satomi) ie mac mAs her lateral variation (of amplitude in this chapter, of vector dip in Chapter 4) ina manner that enhances short-wave~ Teagth, modlerate-wavelength, or long-wavelength varia J, te Hilbert transform displayed in Figure 7 is Figure 7. The Hilbert transform displayed as a linear combi raion of the difference operators displayed in Figure 5 vol Characterization nearly identical to the very-long-wavelength (a = 0.01) ive operator plotted in yellow in Figure fractional deriva 14a of Chapter 4. Luo et al. (2003) call this variation of the Sobel filter the generalized Hilbert transform, or GHT, For readers not enamored with the mathematical beauty of the Hilbert ‘vansform, we suggest simply interpzeting this method as @ suite of weighted difference operators, each of which oper ales at a different length and thereby enhances edges that have different wavelengths. The result of Luoet al.'s (2003) weighting is to push the sensitivity of their generalized Hil bert transform to intermediate wavelengths, thereby pro: viding @ good balance between the cuthank and point-bar amplitude edges, ts shown in Figure 8d In Figure 9c, we display the results of the generalized Hilbert transform applied to seismic data shown in Figure 9a, The channel images are superior o those obtained using the eigenstructure-coherence algorithm shown in Figure 9. The differences result from three causes. First, the GHT edge detector is a long-wavelength operavor und is able « extract subtle, long-wavelength changes in amplitude pra, dients unseen within the smaller (typically five- t0 nine: trace) eigenstiucture colterence-anaiysis window. Second, se \ vont tenstom b) cena ‘i Figure 8. The sensitivity af alternative edge detectors to abrupt changes and togradational lateral changes in ampli- lude. (a) A eross section of athin channel, with the inner hank on the leit and the outer (or eut) bank on the Fight Following Widess (1974), the amplitude response is linearly proportional to the channel thiek1ess.(b) The response of ‘three-point Sobel (or frst-derivative) filter, The steep cut hank shows up well, bu the gradatioral inner bank shows up poorly. (e) The response of a 61-point Hilbert unsto filter. Both sides of the channel are imaged equally well, but, wwe feel thatthe steep cutbank is overly smeared. (d) The response of a 21-point generalized r= 2 Hilbert transform filter. Both sides ofthe channel are well imaged. After Luo stal. (2003), pyrightedLateral Changes in Ampliude and Pattern Recognition 108 ‘operators use more traces and will do a beter job of cking out errors caused by the difficule near-surface eon. ditions. Third, the eigenstructure- (and emsscorrelation-) co- hiereace operators are insensitive to lateral changes in amp tude of a fixed waveform. [F such channels are well below ‘he tuning thickness, their waveform will not change (Fig- ure 1) but their amplitude will Ideally, caleutations of lat- ral changes in reflectivity operate along the dip and ai muth of the seismic reflector at each point in the seismic volume. Usually, for small analysis windows containing five to 25 traces, we ean approximate reflector dip by a local plane as described in Figure 1 of Chapter 2. However, for larger windows containing 100 or more traces (such as the generalized Hilbert transform shown in F approximation may break down except in r ot have ah © alo re 9c}, this gions that do h degree of structural deformation, In areas with structural deformation, before we apply long-wave length amplitude edge detectors we recommend first fat toning a volume shout the zone of interest to remove most of the impact of regional eurvature, Energy-weighted Coherent- amplitude Gradients Although the previous algorithms have measured am plitude variability of the total seismic data regardless of the » alternative approach is to measure amplitude variability of just the coherent component of the seismic data, We discussed eigenvalues and eigenvectors in Chap. tee 3 as partof our coherence estimate. We seview that pro- cess now by redisplaying here, as Figure 10a-d, the graphic explanation of the eigenstructure caleulatioa shown ia Fig ure 17 of Chapter 3. In Figure 10d, sve plot the amplitude of he amplitude-variability volume generated using th Peninsitla. Amp fuseand dif c anomalies associated with channe is insensitive to these changes in amplitude. Io contrast, th a seismic data volume and () the corresponding alized r ill and changes in thickness ieult 9 delineate on the seismic amplitude time slice show in (a). The sstrueture coherence volume and norm Hilbert transform from a survey on the Arabian clearly present bit ars somewhat lif= eralized Hilber transform in (¢) is direct measure of amplitude variability and delineates the channels clearly. After Luo et al. (2003), Figure 10. A schematic diagram surnmarizing the steps in te ei 3, Figuee 17. The steps are to (a) cal waverorm that best best fits the input trace. The amplitudes of the five wavelets i component eigenvactor, vi", To caleulate energy wel shown by the dotted line ofthe input traces within an analysis window. proximates the waveform of each input tras d coh d) and weight it by the sum of the coherent en nsrueture estimate of coke nice shown earlier in Chapter b) ealbulate the seismie sealed version of (b) that ch race with sand (6) repli ©) define the components of the five-clement-long principal araplitude gragkents, we take the derivative of the eurve ‘within the analysis window shown in (c).106 Seismic Altioutes for Prospect Identification and Reservoir Characterization the coherent wavelet that fils each trace bes: within the analysis window. These five data points form the five ments of the prineipal-component (or first) eigenvector v! given in Chapter 3 by equation 3.15, To analyze the lateral variation in those values, we fit a eurve through them, as, shown by the dotted line in Figure 104. This concept is readily generalized tw 3D data, Hei cnvector, vl, as bel we can think ofthe mathematical e 2 a discretely sampled map, v(1,y) that represents the lateral variation of the coherent energy within the analysis along the structural dip and azimuth of the re Two variants of the ampliude-gradient calculation cist. The simplest estimate is to take the derivative of the unit-length principal-component eigenvectors themselves, vil We will call those derivatives eigenvector gradients, Unfortunately, such gradients are overly sensitive to noise flector (Figure 11). in low enewy areas of the seismic volume. We have found that a more useful estimate is to weight these derivatives hy the principal-component eigenvalue, Ay. First, for simplicity, we assume that each wace in the analysis window has the exact same waveform (but possi bly different ampliwdes), such as we see in Figure 10¢. Given that waveform, we can express all the taces shown in igure He with the si 10d, Next, we meognize that each element of the diagonal oF given by equation 3.17 is @ measure of energy based fon the real and quadratare components of the seismic trac ILisa well-established mathematical fact that the sum of the diagonal elements of the covariance matrix C is equal Figure 11. lne:pretation of the eigenvector, vi, cortespond- ing (0 nine-trace analysis window, as the representation of a map of coherent amplitude, 1,9). From this map, we c caloolte gradients, 2° (xy) and 28 (sy), whi wil ind cea lateral changes in the coherent amplitude. These gradi- cents are paricularly aseful ia mapping lateral changes in layer hickness when the layer thickness is less than one-quacter wath, For suck thin layers, the armplitude variation is portional to thickness an wavek constant. After Marfurt (2006). tw sum of the sigenvalues, Thus, if all of the ener resented by just on wenvector has been normalized to be of unit length, then Ay is equal to the energy of the traces in the analysis win, dow. If the tnices in the analysis window have different waveforms, the amument iy somewhat more complicated and requires first that the data be decomposed into princi pal components (which we will do in Chaptsr 8) seb that the condition of waveform similarity holds and reproduces the argument made earlier In summary, 4; is a measure of the energy within the analysis window, whereas v! is a measure of the lateral variation of amplitude across the analysis window, There fore, we will all the spatial derivatives of 29" ener weighted coherent-amplitude gradients. The energy-weight ed coherent-amplitude gradients are large when there is rapidly varying, high-amplitude coherent energy, and dhe ‘gradients are small when the reflectivity is either smoothly varying, low energy, or incoherent. We have found the im ages from cnergy-weishted coherentampliude gradients to be quite complementary to coherence images and to pro. Vide needed detail in reservoirs represented by a strong, consistent waveform, Tio minimize the impact of acquisition footprint (acqui Siion Footprint will be discussed later in Chapter 8), Mar furt (2006) avoids using simple derivatives ealculsted slong the ¥- and y-axes (the inline and evossline direction) by using a more general derivative operator of the Form 8a) and (5.80) where (5.80) and where x and y are measured from an origin at the analy Sis point in a centered J-trace analysis window, Note that the center tence (j = 1) is aot included in the caleul Inspection of equations 5.88-c shows that they ean be inter preted 10 be an unweighted average of x- and y-components of directional derivatives obtained by pairs of points strad dling the analysis point in the ana Altheugh gradients of coherent energy ean be quite ef fective in mapping faults and fractures, we have found them to be most useful in delineating thin channels for which they emphasize subtle lateral changes in tuning. Stati graphic Features such as channels are examined best by using horizon slices that better represent a fixed point in geologic time. In Figure 12, we eexamine the Vinton Dome data volume shown previously in Chapter 2, Figures 8 and Sand in Chapter 4, Figures 36 and 37. Here, we display ex sis window.Lateral Changes in Ampliude and Pattern Recognition 107 iractions, along the Hackberry horizon, of amplitude, e0- heren and the east-west and north-south components of the energy-weighted coherent-amplitude gradient, Because of the complex lateral variation in velocity and our use of prestack time instead of a more appropriate presiack migration, resolution of the coberence image is poor. How- ever, we still are able to clearly delineate meandering chan- nels that formed before and during salt diapirism, Idontifi jon of such channels in a structural setting can offer inte esting exploration opportunities da cohe Figure 12. (a) A vertic ume along line AA’. Horizon slices alon amplitude (¢) the coher cent-amplitude gradient, and (e) the north-south energy-weighted coher- fent-amplitude gradient, Faults have a wormy appearance hee shave been ontrast, the components of the ¢ cent away from a by arrows) and other stat sce is very high (white), Seismic data courtesy of OPEX. Arter Marfurt (2006), In Figure 13, we display horizon slices of east-west (Figure 13h) and north-south (Figure 13a) components of the energy-weighted cohereat-amplitude gradient come sponding fo the coherence slice shown in Chapter 3, Figure 33b. Overall, the two images express the same geolog However, subtle differences do occur, which we indicate with arrows in Figure 13, Yellow arrows indicate narrow channels, which we also interpret to be thin the tuning thickness and thus having no lateral change in waveform, The two channels indicated by magenta arrows, well below slice through the seismic amplitude vol: The dotted line indicates the Hucknerry horizon. 1¢ Hackberry horizon through (b) the seismic 4) the east-west enerey-weighed coher: use the n interfingerin -weighted amplitude gra sccurate images of channels tindieated raphic features, including areas where the of reRectors, ated poorly, esultin108 Seismic Altibutes for Prospect Identification and Reservoir Characterization in Figure 13a are wider but have a much gentler lateral In Figure 14, we show a similar comparison but in a change in amplitude, Although oar eye ean see them easily carbonate terrain, Roth the eigenstructure coherence and in Figure 13a, the local nine-irace coherence algorithm used the energy-weighted coherentamplitude gradients were lo compute Figure 33b in Chapter 3 camot, Such slowly computed using a nine-trace, =1D-ms-analysis window. The varying. long-wavelength changes in amplitude would be a ima good candidate for Luo et al.’s 2003) generalized Hilbert res transform, whieh we discus es are horizon slices of a chert ressrvoir through the spective attribute cubes of the Devonian Thirtyone For odin the previous section, mation from the Central Basin Platform, Texas, U.S.A. The Figure 13. Horizon extractions corresponding te Chapter 3, Figure 33h: fa) north-south and (b) east-west components oF the jenergy-sweighted eoherent-amplitude gradients described in Figure 11. ‘The vertical analysis window was: 32 ms, Yellow arrows indieate narrow channels not readily seen on the ceherence images shown in Chapter 3, Figure 33. Statistically, narrow channels also are thin. If ey are thinner than the tuning thickness, they will be expressed only asa lateral change in amphiude, fot in wavelorm, Magenta arrows indicate a wider channel that is not seen in the coherence horizon slice in Chaprer 3, Pipere 335, We interpret these channels to have a longer-wavelength gradational change in amplitude such as that shown in Figure 8, so they are not wel illuminated by our nine-point coherence algorithm. After Marfust and Kirin (2000). Figure 14, Horizon slice along the Devonian Thistyone For ff (a) eigenstrueture coherence, and (b) the eas the reverse faults indieated by on fromm a survey in the Central Basin Plauform, Texas, USA, ‘est component of the energy-weighted coherent-amplitude gradient, Note that cllow arrows show up as low-coherenee black zones in the coherence image. Because th sweighted coherentamplivud almost tually incoherent, they appear as bla ay zones ia the ene adient image. We can sce channels in the northern third of both images, Although the southern one-third of the coherence image is relatively feature Tess, we still see chaannel-like features int outhern one-third of the amplitude gradient image. These channels cross.an east- ‘west fault indicated hy green arrows on the horizon extraction, further suggesting that they are sedimentary features. After Blumentritt et al. (2003),Lateral Changes in Amplitude and Pattern Recognition chert was deposited as turbidite flows of sponge spicules that grew on the platform to the north; these flows gave rise to a highly productive but highly compartmentalized reses- voir, We can see hints of these channels on the coherence horizon extraction in the northern one-third of the image shown in Figure 14a, but there is little hint of the reservoir heterogeneity in the southern one-third of the image. The ‘wo magenta arrows indicate broad zones corresponding to major reverse faults cutting the survey area. Because the data in those zones is highty incoherent, they appear as black bands in Figure Ha In contrast to the amplitude-gradient work developed by Luo etal. (1995, 2003) and applied to the complete seis- volume, the east-west energy-weighted coerent-ampli- tude gradient shown in Figure Hb is applied only t0 the eo hereat component of the data. For that reason, the reverse faults indicated by magenta arrows appear as blank cones (pure gray on a black-gray-white single-gradational-color bar). In the nerthem one-third, we see the same channels on the amplitude gradient that we see on the coherence, but in slightly more detail. In the southern one-third, the ener2y- weighted coherent-amplitude gradient is superior to the co- herence image anc provides strong evidence of reservoir het crogeneity. The east-west gradient preferentially highlights north-south-trending features that we interpret to be turbi- dite deposits from the north. Indeed, these features are quite continuous across the east-west fault indicated by the gre arrows, which has several hundred milliseconds of throw. The continuity of the north-southetrending features sug- jests that they are stratigraphic rather than structural We sce @ similar algorithmic behavior in Figure 15, which exhibits channels of approximately Pennsylvanian age seen on atime slice from a survey acquited in the Mid 109 continent arca (Oklahoma) ofthe U.S.A. Again, the images resemble each other, but we ean trce the path of the chan- nels considerably farther using the energy-weighted coher- cent-amplitude gradieats In Figure 16, we display time slises at (a) 0.800, (b) 1.000, (¢) 1.200, 1d) |.400, (e) 1.600, and (f) 1.800 s. These mages correspond to the dip and szimuth images shown in Figure 20 of Chapter 2, the coherence image in Figure 35 of Chapter 3, and the most-negative-curvature and most- positive-curvature images from Figures 38 and 39 of Chap- ter 4. To facilitate the comparison of these images in Figure 16, we have retained the arrows from Figure 20 of Chapter 2. A significant acquisition footprint in the shallow tim slices ereates the subtle erosshatehing appearance that is especially evident in the upper, right-hand postion of the cross section. This acquisition footprint arises from the ‘varying fold and azimuth disuibution ia the CMP bias. The edges of the slump features in the southern part of the sur- vey are indicated by the yellow arrows in Figures 163-4, but we conclude that the coherence images shown in Figure 35 of Chapter 3 provide better resolution. Once again, how- ever, the energy-weighted coherent-ampiitude gradients show better delineation of channels, which ae indicated by the red arrows in Figure 16d. Note that the channels are only partially resolved by the coherence. Also, because no differential compaction overlies these channe's, they are not resolved at all by the curvature images in Figares 38 and 39 of Chapter 4 Finally, the north-south-trending chan- nel indicated by the magenta arrow in Figure 6d appears to be structurally controlled ‘The time slice images at 1.600 and 1.800 s are some ‘what less clear to interpret. First, at this level, we know from Pigures 38 and 39 of Chapter 4 that we have consider Figure 15. Time slices at 1= 0.832 through the (a) east-west and (b) north-south components of the energy-weighted coher- cent-amplitude gradients, and (c) a coherence time slice through a survey st approximately Pennsylvanian level from the mid continent, U.S.A. Although the channels can be seen on the col ‘gradient images. + time slice, they can be traced farther using the amplitude110 Seisrric Altrioutes for Prospect Identification and Reservoir Characterization able folding, so we do not expect 10 see clearly defined stratignaphie features on a time slice. In Figures 16e and 16f, we do see lineaments that follow the erests of the folds, seon it Figure 39 of Chapter 4 and that we interpret to be changes in amplitude that result cither from deposition on. paleotopography or from differential compaction. The red and blue arrows in Figures 16d-f indicate song fold aves, in the dip and azimuth, coherence, and curvature im ages. These fold axes do not appear on the enengy-weighted coterent-amplitude gradient images, which implies that any. thinning or thickening about these axes is insignificant. igure 16, Time slices through the east-west ‘componentof the coherent energy-weighted ampli= tuule gradient at (a) 0.800, (b) 1.020, (2) 1.200, (d) L400, fe) 1,600, and (f) 1.800 + corresponding to images shown ia eater in Chapter 2, Figure 20, Chapter 3, Figore 38, and Chapter 4, Figures 38 and 39, We have retsined the same artows shown in Chapter 2, Figure 20 toaid in the comparison, ere, we indicate chine nels at¢= 1.400 and 1.600 with red arrows, The structural feature indicated by the magenta arrows ‘appears vo vont a aorth= south-trending channel We also note strong lin laments within the eitele: fon the 1.600 and 1.800 s images. These features dooper than our welleor trol and are more ambigu: ‘ous to interpret. We inter pret them go be changes in ariplitade that result form eposition on paleotnpog- raphy oF from differential compaction, Second Derivatives of Amplitud Amplitude Curvature Ii we think of structural curvature, which we discussed in Chaptes 4, as being lateral second derivatives of the Gime structure (or phase) component of seismic data, it seems reasonable to expect information also t0 be contained in the lateral second derivative of the amplitude behavior oF seis ic data. In Figure 17a, we provide an idealized cross see, ‘ion across a lateral change in amplitude. In Figure 17h, we calculate the frst derivative in «of the amplide shown inLateral Changes in Ampliude and Pattern Recognition 111 Figure 17a, In Figure 17, we display the amplitude’s see- ond derivative in ¥. In three dimensions, we fit a quadratic surface to the amplitudes and calculate the curvature mea- sures discussed in Chapter 4. Computationally, all the egua~ sions prevented in Chapter 4 that were applied to the ialine and crossline components of reflector dip and avimuti also can be applied to the inline and crosstine components of the energy weighted coherent amplitude gradients (or to other directional measures of amplitude variability) AS was the case with structural curvature, we have found that loag-wavelengih estimates of amplitude curva- lure provide additional information. Figure 18 is a time- thickness map of the area between the top and base of & carbonate ree, generated from a survey acquired in south eastern Saskatchewan. The figure includes a representative vertical seismic slice through the reef, We note that there are considerable lateral changes ia amplitude shat result from changes in thickness and (possibly) changes in lithol- gy, such as dolomitization. Amplitude is weakest where ihe reef Is thickest In Figure 19a, we display the amplitade extraction along the top of the picked reel from Figure 18, The weak~ est amplitudes are light blue, In Figures 19 and 19e, we display the most-negative and most-positive amplitude eur vatures, The most-negative-curvature anomalies in dark een track primarily what we identified as low-amplitude {light blue) anomalies in Figure 19a. In contrast, the most- positive curvature in Figure 19¢ tracks the edges ofthe rect. (Quantitative analysis of such second/-derivative maps awaits calibration with geostatistics. Far now, we reognize such images as another means of measuring subtle changes in seismic character Figure 17. Aschematie di sly, (b) its firs: spatial derivative, and (c) its second s Lerivative. The extrema of (c) demarcate the limits of the anomaly m of fa) an amplitude anom= sata Pattern Recognition of Seismic Data: Texture Mapping Interpretation of seismic data for the ukimate objective of finding hydrocarbons has always been driven by pattern recognition. Interpretation of seismic wiggles and wavelets is done in terms of th about subsurface geology. Recognition of the patterns of amplitude on a horizon extraction, for example, hss been used for a Jong time, but itis not enough. ‘The human mind is highly skilled at recognizing patterns — ask any parent looking for their child in a crowd. Imitating this skill with computers requites two steps: (1) extseting distinguishing features in our ssismic data and (2) assigning subvolumes Of seismic data to classes or clusters. Step two isthe subject of human interpretation and ongoing developments in sco statistics and neural networks, neither of which will be ad- dressed in this book. The first step falls cleanly into the do~ main of aitibutes. A good auuibote or texture should be robust and insensitive to local distortions. In this section, ‘we address three types of pattern recognition used in seis= mnie interpretation. wayerarm clasification, seismic geo- ‘morphology, and texture mapping, We touch only briefly oon the first two. A photographic or computer image consists of a pata arrangement of colors and their intensities. The repeating patiern of local variations (in the small neighborhood con- sidered) in image intensity is called texture, Its a feature used to partition images into regions of interest and to clas: sify those regions. So in that sense, testure is an important delining characteristic of an image patterns, which yield information Figure 18, \ time thickness map and a verti slice through a Winnepeposas teef from southeastern Saskatchewan, Canada. The thickness is calculated between the blue pick (op) and the green pick (base). Nowe thas here, the maxisnun amplitude corresponds to the thinnest part of th plex, Data courtesy of Talisman, ef com-112 Seismic Altioutes for Prospect Identification and Reser Textures consistof texture primitives, of elements, which, sometimes also ane referred to as tevels, Textures usiilly are described as fine-grained, coarse-grained, smooth, and. the like. These features ean be found in the fone and struc dure of a texture. Whereas tone refers to the pixel intensity | ce | pitas Poste Figure 19. (a) Amplitude, (b) long-wavelength most-negt- tiye-anplitude eurvatue, and le) long-wavelength most powitivesamplitide curvature, along the top af the reef (lee pick) shown in Figure 18, Here, the strong negative values Of the mostnegative-curvature, seen ck the low-ammplitude light-Alue zrends in the amplitude horizon slice, In contrast the strong po of the most pes tive-curvature, seen a8 dark red, track the high-amplitude red trends in th delinoste the edges oF the reer. The rectiline in both (by anal (6) are the resul of aequistion foompeit. Two hundred traces were used inthe curvature calealstions. Data courtesy of Talisman tive values amplitude horizon slice. ‘Paese latter anomalies features seen i Characterization properties in the texel, structure represents te spatial rela tionship among texels. Thus, if texels are s differences between them are large, we talk of a fine tex lute, Likewise, if texels are large and consist of several pix els, a course texture results An example of natural textures is rock texture. Rock textures usually are nonhomogeneous and strongly diree: tional, In addition, th all but tonal rain size and color of the texture can vary significantly in some rock texture types, so their analysis and classification can be dificult Visual textures contain variations of intensities, which form certain repeated patterns. These patterns may be caused by physical-surface properties such as roughness or reflec tance differences. Because of their stochastic nature, rock textures can be characterized into different orders of statis: ties, First-order probability distibution of the amplitude of the quantized in tribution. These statistics are essy to compute and include the mean, standard deviation, skewness, and kurtosis. Instead of distinguish possible to distinguish features via their shape or their struc, tural characteristics. These help us in extracting features that in tum help describe textural properties of classes. Co. occurrence matrices have heen used quite successfully in ‘many pattern-recognition applications, inckiding remote sens, ing. Three measurements are used commonly for analyzing ctr be used to entrtet the statistical dis the vasiation in intensity, iis textures: spectral, structural, and statistical measures Spectral measures of texture Light reflected from a rock surface forms part of the sisible spectrum. Thus, characterization of the rock essen dially becomes characterization of itsteflected light. Achto, atic light has only one attribute — its intensity. The scalar measure of reflectance is gray-level. All colors ean be seen as variable combinations of the three primary colors: red (R), greea (G), and blue (B). Although these colors and their combinations are used in measurement of the visible spectrum, another scheme, called the HLS model, is imple: mented to exiract spectral information trom the textural image. In this model, hue (H) describes a pare color, and saturation (S) gives a measure of the degree to whieh the patre color is diluted by white light. Lightness (L) is decou pled from the color information of the imag Structural measures of texture Brick-wall-type regular textures are composed of simi lar patteras that are repeated structurally. By defining tex ture primitives. whereby each primitive is deseribed by 2 tonal value, we can use two displacement vectors, alongLateral Changes in Ampliude and Pattern Recognition 113 which primitives are formes, todetermine the spatial disti- hution of those primitives in a asighborkood. For example, we may assume texture primitive to be a parallelogram nd the primitives to be arranged along the periodicity an- ples of the parallel Statistical measures of texture Textures can be regular, or they ean be random but with consistent properties. Thus, a logical way to describe such textures is through their statistical properties. Co-oceurrencs matrices are calculated locally within a small window thal glides across the image. The choice of window size is a judicious one: ‘The window should be small enough that the Jow-pass-filtering effeets from mask= ing remain tolerable, yet large enough that the extracted in- formation exhibits statistical Significance. ix (COM) is a square whose elements correspond to the relative frequency of oe- currence of pairs of gray levels of pixels that are separated by Phe im mediate neighbors of any pixel can lie on one of four pos. sible directions: 0°, 45°, 99°, and 135°. The COM is con- structed for observing pairs of gray-level values at distance J from cach other in the direction considered, Thus, four matrices can he derived for each distance, with P represent- ing @ COM: PO), PASE), POA, and P35. the spatial relationship is derived in terms of distance o and angle 0. certain distance and lie in a given direction. Let us consider an image that hay the elements shown in Figure 20, We will refer to this image as a test image This test image has four gray levels (1, 2, 3, and 4). IF we 1 1 2 1 1 2 1 3 | 3 | 3 3 |3a | 4] 4 Figure 20. A:est image that will be used to form a co-e jee matrix, aking into account the relationship hetween the different eo-0 the adjacent elements. This schematic show's ad x 4 window Df Seismic samplitades that have heen Sealed t fall within four bins. value of 1 indicates the lowest (typically nega ive) amplitude, wheress a value of 4 indicates the highest amplitude, Typical’, scaling unis will hea froction of the rms amplitude of the seismie data volume, fix the spatial relationship (or the position operator) as the element to the right, then the pray-lovel co.oceurrence ma trix (GLCM) becomes, 2210 0200 69) 0031 ooo) The el ment in the first row, second column of the ma- trix given in equation 5.9 indicates how many timesin Fig ture 20 the number 2 occurs to the right of aumsber 1. This, happens exsetly two times, so GLEM(,2) = 2. Similarly, the number of times that the value 4 cccurs to the right of value 4 is only one time, so GLOM(4) = L. There are no. occurrences of 4 appearing to the tightof 1, so CGLM(4) =O. The matrix in equation 5.9 is not symmetrical (e.g., the valuc of the cell 1,2 isnot the same as the value of 2,1). This matrix can be made symmetrical by counting each, pixel pa that is, once to the right and once to the lett wie Analtemative and easier way to make the matrix sym- metrical is to add the matrix to its transpose (interchanging the rows and columns oF the original mattia), The transpose of the GLOM matrix given by equation 5.9 is 2000 ouemt= | 2.2 00 (5.10) 1930 oo1d ‘Ths, adding the transpose to its GLCM gives the symmet- sical matrix J42 10 GLCMgmm= | 2 4 0 S.1b, 1061 0012 21 shows, we will use four differ: tent axes, each of whieh will generate its own COM for an intersample distance, , that is equal 10 1 The four matrices in Figure 2b ean be combined into In general, a Figur xed co-vecurrenve matrix because no directional variations in texture are expected lo 4 60 GLEMgym= | 412 5 0 6) 65126 00 62 ‘The nest step is to seale the GLCM such that it looks like a probability table. 1s an approximation, because a true probability would ne continuous values, whercas the gray:14 cls are integer values and So are discrete. This process is seed t as normalizing the matrix and consists of divid ing the elements by the sum of values. Thus, the e0-oceur- rence matrix represents the joint probability occurrence oF ray levels j and j for two pixels with a defined spatial rela tionship in an image. ‘That normalization results in 0.19 0.05 0.97 0.00 pe 005 0.11 0.06 0.09 5.13) 0.07 0.06 0.14 0.07 0.00 0.09 0.07 0.02, ‘where we now use the symbol P to denote that cur GCLM is, an estimate of probability. Each element of equation 5.13 ep. resents the probability of going from gn jj when the intersample spacing is ¢ and the direction is @. In practice, instead of computing a simple average, usw: ally we perform a weighted average — that is, each value in the mattis is multiplied by a factor (weight) before summing, and dividing by the number of values. This allows us to ae: level fto gray level GLCMs for seismic data The GLCM has dimensions » Xn, where nis the num: ber of gray levels. For application to seismic data, the gray levels refer to the dynamic range of the data, For example, B-bit data will have 256 gray levels. A GLCM computed for these data would have 256 rows and 256 columns (or (65 536 elements). Similarly, 16-bit data would have & mar trix of size 65 $36 x65 536 = 429 496 720 elements, which, could he overwhelming even for a computer (Haralick et al, 1973; Reed and Hussong, 1989; Gao, 2003). Usually, the seismic data are rescaled to be 4-hit data (resulting in a 16 x 16 matrix} or five-bit data (nesulking in a 32 x 32 ma- trix). In practice, this does not result in any significant ail ferences in the computed propertios, The structure of GILCMSs as applied to seismie data ean beunderstood easily, Figure 22b shows ther gion segments Seismic Atiibutes for Prospect Icientification and Reservoir Characterization 1, 2, and 3 selected for GLEM computation, and the com: pitted GLCMs are shown to the right. For strong contin ous reflections, the GLOM eshibits a tight distribution along the diagonal, The matrix size chosen is 32 and the pa rameters chosen are 4, 3, and 1, respectiv directions. Window 2 has lower amplitudes as well as inco- herent reflections, so the GLCM shows a scatter about the di occurrences farther away from the diagonal (window 3). For a matrix size 16, we see a smaller number of ele- ments in the GLCM (Figure 22a), whereas for a matrix size 64, there isa higher population of points (Figure 2c) Although GLCMs give us all this information, they are fot suitable for direct interpretation by a human being Haralick et al. (1973) demonstrated the derivation of 4 dif, ferent measures of textural features ftom the input GLCMs, sin the x,y, and maul, Discontinuous or incoherent reflections have more Each of those features represents specific image propestics, such as coarseness, conurast.or textural complexity, The Four such measures used most commonly are energy, entropy comurast, and homogeneity Energy Energy is a measure of textural uniformity of an image. Mathematically, itis given as Dae Py donotes the ith row and the jth coliusn of the GCLM matrix P, such as the one in equation 5.13. Energy is low when all elements in the GLCM are equal, and it is useful for highlighting geometry and continuity (6.14) where Entropy Entropy is a measure of disorder or complexity of the Entropy = >) Py le P, was bs 50° as (24 0) rot) (fore) (as oor at looi a) loorojloo: o Figure 21. Computation of a itseight neighbors. [a this masner, we will compute preliminary GLCM matrices at 0% 45 iay-level co-oecurrence matrix (GLCM) by comparing each element shown in Figure 20 with ‘90°, and 135° from the vertical. The final GLCM mairix will be an average of these four preliminary GLCM matrices.Lateral Changes in Amplitude and Pattern Recognition 15 for images that are not uniform texturally fh acase, many GLCM elements have low values. En lopy measures the organization of pixels. It is large when the values of GLCM are uniform, which is possible in the ease of an equiprobable seenario. Entropy is Ia Inst Contrast Contrast is a measure of the image’s contrast oF of the amount of local variation present in an image. Conat= SP0-7'*, Contrast, oF inertia, is high for contrasted pixels, and its ho- mogeneity (defined below) is low, When used together, both inertia and homogeneity provide discriminating information 6.16) Matrix size 16 Metin azo 64 Homogeneity Homogeneity is a measure of the overall of an image Homogeneity achieves iis maximum valus when most of u CM are concentrated near the main diagonal. Highly contrasting pixels result in Jow-ho- ‘mogeneity values, so homogeneity is inversely proportional to contrast, Homogeneity is useful for quantifying reflec= tion continuity For 3D seismic volumes, computing GLOM texture at- tributes at one location yields the localized features at that point. If we repeat the computation of those attributes in a sequential manner throughout the volume, we transform the input seismic volume into trib which we discuss below. For seismic data with significant dips, itis advisable to dip-steer aGLCM calculation using techniques discussed in Chapter 2. West et al. (2002) ad- opted @ gradient-based dip-steering method, wherein the horizontal (dx) and vertical (dy) gradients of pixel values, ‘are computed in two orthogonal directions and then used to, cestimate the true dip and azimuth for dip and directional (5.17) least four text Matix size 32 Figure 22. (a) Computation of GLCMs for different reflee tion cheracteristics selected by polygons marked 1, 2, and 3 ‘The size of the matrix used for computation is (a) 16 % 16 (b) 32% 32. and (c) 64 » 64, Notice that with a matrix size of 16, fewer elements are seen populating the GLCM, compared with the 32-element GLCM. The bi ou. ighest density of elements ‘occurs with a 64-clement116 Seismic Altioutes for Prospect Identification andi Reser steering of the GLCM calculations, This method is the basis, for the gralient structure tensor estimate of dip and avi Imuth, also discussed in Chapter 2. High-amplitude continuous reflections, which general ly are associated with marine shale deposits, have relativ low energy, high contrast, and low entropy (Gao, 2003), Low-amplitude discontinuous reflections, which generally are associated with massive sand or turbidite deposits, have hhigh energy, low contrast, and high homogeneity. Low-fre quency high-amplitude anomalies, whieh generally indieate hydrocarhon accumulation, generally exhibit high energy, low centrast, and low entropy, relative to noniydmearbon sediments, Case study 1 Our first ease study focuses on a 3D surtace-seismie survey in southern Alberta, Canada. ‘The target zone was Lower Cretaceous glauconitic Nuvial deposits that have been, productive in the area. The 3D seismic survey was xequired to create a stratigraphic model that would be consistent swith the available well control and would mateh the pro- duction history. Our ultimate goal was to locate the remain ing undeveloped potential in the known fluvial deposits and to find additional drilling targets in untested fluvial sand stones and in the area in general. The model we developed. was based on a comprchensive geophysical and geological interpretation wherein the geological data were integrated closely into the geophysical model. That integration was, required because the Lower Cretaceous system in the study area has exhibited a very complex fluvial environment through time. That complex environment makes si sle-diseipline interpretations very difficul: in tems of new drilling locations, compared with these that could be de cided on the basis of the present analysis. Figure 23.4 Cretaceous glauconitie fuvial highlighted zone, neat ofa seismic section showing Lower epesits, seen tracked in the i Characterization Because cur objective was stratigraphic, we processed the seismic data to preserve relative amplitudes. Prestack time migration improves our ability 10 resolve stratigraphic objectives and extract high-quality seismic attribules, 0 we ran iton the data, That migration improved the stack image in terms of frequency and lateral definition of features by better energy focusing and improved image positioning prior to stack (Reilly, 2002), Figure 23 is a segment of the seismic section showing the level of production from the slauconiti sandstone. Stratal volumes allow snic interpreters to study ob jects in a 3D perspective, which in turn sheds fight on those ‘objects’ origin and spatial relationships. Stratal cubes are subvolumes of seisinie data (or their attributes) bounded by two horizons that may or may not be parallel. Figure 24a is astratal volume display from the original seismic data, here covering the zone of interest at the level of the reservoir Gust below the horizon shown in Figure 23). Figure 246 shows a strata slice through a coherence volume, processed using a semblance algorithin, Although a beer definition of some of the subsurface features can be interpreted here versus in the migrated stack, this display does not yield much information about the areal extent oF the productive sands. Texture: tribute analysis was run on the subyolume covering the broad zone of interest, and Figure 24¢-f de, picts the energy, eatropy, homogeneity, and contrast-attr bates, respectively, Figure 24c shows high values of enerey associated with the fluvial depesits, and the areal distribu tion is depicted as we would expect it to be. However, this inference needs corroboration from the other texture attr. utes, We sce that corroboration in Figure 24d-e. The high ray in Figure 24¢ 1s associated with low entropy and high homogeneity. Another observation is interesting bere Well W3 (lo the northeast of W2) has a different pressure and apparently does not share the stme producing forma, tion with well W2, The low coherence (Figure 24h) indi cates an islandlike feature surrounding well W3, and the texture attributes confirm that observation, Although itis possible to interpret the productive sands con gamma-ray logs for wells WI and W2 (which have val ues lower than approximately 50 APL units), the texture-at, tribute displays provide a more intuitive presentation of the geology and show the areal spread of these productive sands. Figure 25 shows the 3D view of the stratal cube ex tracted from the energy attribute Case study 2 Case Study 2 is from south-central Alberta, Canada, ‘This 3D survey was acquired to explore the possibility of deciding on reservoir pockets that could be drilled. ‘The field has been producing for about a year, OF theluce and Pattern Recognition 17 ‘al Changes in Figure 24, strata slices through (a) aseismic ampli tude volume and (b) a corre- sponding conerence volunie a) atthe level of a producing sandstone, The seismic amplitudes indicate the sandstone distribution. The the channel edges clearly, and there are indications o el sands in the top left cation on the in sto the productive sands. (€)(F) Strata slices are shown from (c) ener (d) entropy, (e) homoge ity, and (F) contrest atribuies atthe same zone of interes. The energy attribute indi ceates h sinter with the same productive sands seen in wells WI valles that ied to be associated and W2. Corresponding to these Jow values of entropy and high values of homogeneity are seen, which we would expect of fluvial deposits. The low-contrast feature seen around well W3 indi separate block that is confirmed by a different cate observed pressure in this Figure 25. A 3D view of the stratal cube extracied from 1 in energy attribute that are inter 1 24e, The top surface indicates high values of ene preted to be associated with productive sandstone deposits at this level118 Seisrric Altioutes for P >spect Icentificas swells, and one is ahandoned. The 3D seismic ampli-- discontinuities at the reservoir level of interest, therehy signatures that consistently showing that the reservoir’s producing formations donot tudes were expected t0 ind characterized the two subsurface gas formations, whereas form a uniform blanket, although no distinet faults or chan. the oil formations could not be detected solely on the basis nel edyes are obvious ture attributes (F ‘ofamplitudes. \lso, the pressure data in the wells indicated The te that wells Gas-1 and Oil-1 should drain different pools. The calibrate nicely for the hip im this d low entropy, and bi sis Figure 26a) neity. All four oil-bearing formations show moderate values Figure 26, A strtal slice at eservoir level From (a) the seismic b) the corre 24, The strat slice characterize the wo gas formations ating fragmented formation and explain ing the different peocuetion pressures, bserved in Gas-1 and Oil wells, coo Figure 27. A strata slice from (a) seismic, (b) energy, (¢) entropy, and hontogeneity attributes. Notic hat the S wells penetrate fand the dry well penetrates low energy pockets, High- to moderate low entropy and high homogeneity, as we would expect for uvial deposits,Lateral Changes in Amplitude and Patton Re Other examples Figure 28 shows a salt canopy detected from a seed in wergy cube, The texture energy of salt is signifi- her than that in surrounding areas, so the whole salt body can be detected, isolated, and mapped effectively by propagating the seed within the salt, Because amplitude samples within the salt body are similar 10 and connected with those in the surrounding areas, seed-based propaga- tion may cause bleeding across the salt houndary and is not offcctive for sutomatie salt detection, Figure 29 shows (a) an original amplitude section and (b) a texture ene jon. An energy cube with opacity applied (Figure 29e) isolates the high-energy Festure (red) along a channel system by rendering transparent the low homogeneity features. It is difficult to isolate the same fea- lure using the original amplitude volume because the am- plitude is limited in discriminating channels from other geologic features. 5 ofamplitude and energy at the same stratigraphic level. Notice that the chan- nel and levee deposits can be recognized, mapped, and de- tected more effectively from the energy volume than from the amplitude volume. sure 30 compates the horizon s Figure 28. A salt canopy detected by using a seed in a tex: ture energy volume. The energy alt are higher than the suercunding areas, can be detected m than by using a seismie volume. Use of seismic volume alues corresponding to the snd the salt body re effectively from the energy volume would result in bleeding across the boundaries and would prevent acrisper definition of the salt canopy. After Gao 2003). ognition 19. Seismic facies application of texture attributes Seismic facies ean be defined as stratigraphic units or regions with characteristic reflection pattems that are dis- tinguishable from those of other areas on the basis of re- flection amplitudes, continuity, geometry andlor internal configuration of reflectors hounded by stratigraphic hori zons (Mitchum, 1977}. Auempts to analyze seismic facies, usually involve two steps. First, seismic facies patterns are defined in terms of their lateral and vertical extents. See ‘ond, those defined seismic facies are interpreted in terms of their lateral and vertical associations and their calibration with wells all of which give insight into the geological and depositional settings. This step is significant because the relationship between seismic data, seismic facies, and dep, ositional environment is not unique. Conventionally, an interpreter delineates seismic facies between mapped horizons, This entails examining the dom- inant seismic facies on vertical sections through the seismic Volume and posting that information on a map. Usually, a 2D map is produced that generalizes the distribution of seismic facies vertically within a mapped interval, Besides being laborious, it may be difficult to map different seismic facies consistently in large and complex areas, especially if multiple mapping units are involved West et al. (2002) have demonstrated a probabilistic neural-network approach to mapping seismic facies quanti {atively in 3D surface seismic data. In West et al.'s ap- proach, interactive training of neural networks begins with, definition of analysis parameters such as calculation vol ‘ume, window size, analysis distance, and the like. The in texpreter selects polygons on ke: ‘out of 3D seismic volumes that exhibit different seismic fa: cies. The computer then computes the associated GLCMs, and in this way several examples of each class of facies are specified. After the taining procedure, several quality con ‘ols are run. The result of this exercise is a seismic facies, classification volume in which each truce and sample has & seismic facies classification, Figure 31 shows the definition of taining polygons on 4 seismic section. The interpreter has selected the polygons, oon the basis of their reflection character. The classification scheme used is ASC (amplitude semicontinuous), HAC (high- amplitude con LASC (low-amplitude semicontinuous), MASC (moderate amplitude semicontinuous), MAC (moderate-amplitude con- tinuous), and HAS -amplitxle semicontinuous). The result of the textural analysis isa seismic classification vol ‘ume, which is examined and calibrated with the available well and core information, Figure 32b is an example from a seismic lines extracted uous), LAC (low-amplitude continuous),120 Seismic Attioutes for Prospect Identification and Reservoir Characterization channelized deepwater reservoir, where a slice from the seismic facies classification volume is seen heing compared \ith an equivalent slice from the coherence volume (Figure 32a), The coherence slice (Figure 32a) highlights the lateral edges of the channel (Fed lines), a broad, older sinuous ele ment (1), and a narrower, younger sinuous element (2). The Figure 29. (a) A segment of the original seismic section from Figure 28, (b) The corresponding texture energy section. (c) The texture energy subcube, with an opacity filter applied. ‘The high-energy feature (red) along a channel system is revealed clearly by rendering amplitudes os transparent. The same would be difficult © visualize and iso- late from the original amplitude volume. Afier Gao (2003). the low-energy a) sheng “Tine (e) ») Figure 30. A horizon slice comparison atthe same strati- graphic level from (a) a texture energy volume and (6) a seismic amplitude volume, Notice how the chanrel ard levee deposits can be recognized, mapped, and detected moze effectively from the seismic amplitude volume. After Gao exture energy volume than from the 2003) ‘whereas sinuous clement 2 is composed primarily of HAC seismic facies. An interesting fact is that be tural analysis seismic facies classification yields a volume, this type of analysis can be applied at different stratigraphic levels in an interval of interest, regardless of whether those intervals are hounded by mapped horizons. Calibration of different features with the well data results in an interpret ton map of the environment of deposition (Figure 320). use the tex ey Figure 31. Atypical seismie facies classification using an interpreter-trained probabilistic neural network, in which multiple facies classes have been identified. The seismic classification scheme on the right consists of high amplitude (HA), moderate amplitude (MA), low amplitude (Li tinuous (C), and semicontinious (SC) seismic facies. After Westet al. (2002).Lateral Changes in Ampliude and Pattern Recognition 121 a) b) ‘Seisme dlscomtruty " Envronmrant of deposition and geobsie interpretation Sosmictacios 3 Channelaxis. (igh net gross Nonnet coverban ‘Overoank \lownetgross) Crate axis (moderate nelgross) Figure 32. (a) A coherence slice highlighting the lateral edges of a channel (red Lines). 4 broad, older sinuous element (1). and MASC seismic facies, whereas sinuous element younger sinuous clement (2). (b) A seismic facies slice showing that singous element (1) is composed of HAC 10 is composed primauily of HAC seisinic facies. Because dhe textural-analy'sis seismie-ficies classification isa volume, this type of analysis ean be gpplied at different stratigraphic levels within an interval whe of interes not the interval is bounded by mapped horizons, (c) Considering the conceptual relationships between ssociated geologie fill, the net-1o-gi0ss environment ean be understood and 3D regions ean Uelireate differing depositional and geologic properties. After West eta. (2002) Chapter Summary Because of the oscillatory nature of the seismic wave- let, and hence of the seismic trace, all good measures of changes in reflector amplitude should be calculated along the dip and azimuth of an assumed reflector: Lateral changes in bed thiekness, lithology, and poros- lay result in lateral changes in thin-bed ning. Eigenstruc ture and erosscorrelation coherence algorithnis are designed Jo be Sensitive only to reflector waveforms and will not see coherent lateral changes in reflector amplitude. Other co- herence algorithms, including semblance, variance, and Man- hattan-distance estimates of reflector similarity, are sensi- tive to both amplitude and waveform. In a similar mannes, algorithms designed ta measure changes i reflector ampli= ude — such as Luo et al.'s (1996) derivative algorithm — enerally also are sensitive to reflector waveform. Unfor Tunately, all sich algorithms comnionly are referred to 2s coherence or edge detectors, which Further obscures their dil- fer Whenever possible, we recommend using algorithms that are as mathematically decoupled from each other as possible, thereby providing the interpreter with orthogonal [independent) views of his or ber data. Obvious end mem bers exist that we could separate, For instance, a lateral change in porosity of a thin reservoir may give rise only 10 a subtle amplitude vaviation, with the waveform remaining consiant, In contrast, & continuous reflector may be eross- cut by backscattered ground roll. The amplitude of the co: herent part of the reflector would remain constant, w the waveform and coherence me: ure would not, Simpler cedge-detection algorithms would sce both of these features as edges or discontinuities, In contrast, energy-weighted coherentamplitude gra- dients are sensitive 10 lateral changes in amplitude for 3 fixed wavelorm, Such gradients are particularly effective in deseeting thin channels that are not seen by and principal-compenent coherence algorithms, and. the gralicnis provide insight into re Be cause they work only on the coherent component of the data, energy-weighted coherent-amplitude gradients high- Tight channels that have stronger reflectivity and deempha- size incoherent fault zones that may confuse our strat. gtaphie interpretation Long-wavelength estimates of amplitude variability, such as Luo et al.'s (2003) generalized Hilbert transform, ‘can enhance lateral changes in reservoir thickness and po- rosity that are more gradational and that have scales larger than the typical five- or nipe-trace coherence algorithm. Long wavelength second-cerivative estimates of amplitude varie ability, achioved by applying the most-positive-curvature and most-negative-curvature algorithms to amplitude gra- dients, also can enfiance such subtle changes in reflectivity AL present, Iong-wavelength estimates of changes in amplitude should be applied with care. Reflector dip varies, slowly in the vertical direction (except near unconformi- ties), thereby allowing long-wavelengti estimates of struc tural curvature on time slices. In contrast, seismic ampli tude varies rapitlly in the vertical direction and renders, long-wavelength estimates of amplitude changes challeng ing in structurally complex terrains. Although long-waxe~ length derivative calculations ean be applied readily 10 dip- eigenstructure oir heterogene!122 ping planes, such calculation reflector has produce artilacts if the any structural curvature falling within the (rather large) analysis window. For that reason, we recom- rend that long-wavelength estimates of amplitude gradi cents be restricted to subvolumes of data that have been flat tened fo remove most of the structural curvature Texture-attribute studies usually are nol associated with scismic attribute studies, We consider texture analysis, in its most general form to be a superset of the geometric attribuies that comprise the major Hocus this book. Texture analysis uses second onder statistics t@ extract aitributes that in turn help deseribe the textural properties of classes. Textures based on GLOMs work well so long as the gram lusty of tentures being examined is of the order of the pixel size, which commorly occurs on seisinie data. On the basis, oF our analysis, we conclude that 1) texture attributes enance our understanding of the res ervoir hy providing 4 clearer picture of the distribution, volume, and connectivity of the hydrocarbon-hearing, facies of the reservoir, ) texture attributes are a quaniitative suite that aids the hrough in using the visual process an interpreter goe: conventional atiributes, 3) the simultaneous and exhaustive analysis that gener aes texture auutibutes gives insights into the linkages ofthe wservoir’s geology and peophysies, and in ceases of its engineering properties, and 4). application of texture analysis for seismic Facies elas sification, using neural netwerks, holds promise in that it can provide seismic facies information thet is not provided by any conventional seismic approach. References Adeogba, A. A., TR, McHargue, 8, A. Graham, 2005 Transient fan architecture and depositional controls, from near-surface 3-D seismic data, Niger Delta con- tinental slope: AAPG Bulletin, 89, 627-643. Blumentritt, C. H., E. C. Sullivan, ané K. J. Marfurt, 2003, Channel detection using seismic attributes en. the Ceatral Basin Platform, west Texas: 73rd Annual International Meeting, SEG, Expanded Abstracts, 406-409. Seismic Atiibutes for Prospect Icientification and Reservoir Characterization Chopra, S., and V, Alexeey, 2005, Applications of texture utributes to 3D seismic data: CSEG Recorder, 30, 28. 32. Gao, D., 2003, Volume texture extraction for 3-D seis, mic visualization and intemretation: Geophysics, 68, 1294-1302, Haralick, R. M., K. Shanmugam, and I. Dinstein, 1973, “Toxtural features for image classification: IEEE Trans actions: Systems, Man, and Cyberneties, SMC-3, 610 2 Luv, Yi, $. al Dossary, NM. 2003, Generalized Hilbert transform and its applica: on in geophysics: The Leading Edge, 22, 198-202. Luo, ¥., W. G. Higgs, and W. S. Kowalik, 1996, Edge detection and stratigraphic analysis using $-D seis. mic data: 66th Annual International Meeting, SEG, Expanded Abstracts, 324-327. Marfurt, K. J., 2006, Robust estimates of 3D reflector dip and azimuth: Geophysies, 71, 29-10, Marfurt, X.J., and R. L. Kislin, 2000, 3-D broadband es Uimates of reflector dip and amplitude: Geophysics, 65, 304-320. Mitchum, R. M. Marhoon, and M. Alfaraj 1977, Scismie stratigraphy and global € 1, Part I: Glossary of terms used in seismic stratigraphy, in C. E, Payton, ed., Seismic stratigraphy: Applicatons to hydrocarbon explora Hon: AAPG Memoir 26, 205-212, Partyka, G., 2001, Seismic thickness estimation: three ‘approaches, pros and cons: 71st Annual Lnternational Meeting, SEG, Expanded Abstracts, 308-506. Reed, T. B., and D. Kussong, 1989, Digital image pro cessing techniques for entancement and classifica ton of SeaMARCII side-sean sonar imagery: Jour nal of Geophysical Research, 94, 7469-7490. Reilly, J., 2002, 3-D prestack data mining to mest emery, ing challenges: 72nd Annual Intersational Meeting, SEG, Expanded Abstracts, 476-479, West, B.. 8. May, J. E. Eastwood, and C. Rossen, 2002, acies classification using text raland neural networks: The Leading Edge, 21, 1D42~ 1049. Widess, M. B., 1 ies, 38, 1176-1254 973, How thin is @ thin bed: Geophys:Chapter 6 Spectral Decomposition and Wavelet Transforms Chapter Objectives Aner reading this be able fo + identity she geologic Features highlighted by spectral decomposition and wavelet wansforms + interprct spoetral anomallos in the context of thin-bed tuning + analyze singularities of seismie data for structural and stra wphie details + evaluate the use of spectral information as a direct hydrocarbon indicator Introduction Since the beginning of digital recording, geophysical data processors have decomposed the measured seismic signal into its Fourier (or specteal) components 10 attenuate low-frequency ground roll, 50- or 60-Hz cultural noise, and high-frequeney random noise. Also, data processors have routinely balanced the source spectrum throug! deconvolution and wavelet-shaping techniques 10 account for the input source signature, spectral changes resulting from shost period multiples, and attenuation of the overburden Any time series ean be represented in terms of a sum- mation of ather time series, For example, in Fourier analy- sis, we eam represent any time series by a weighted summa tion of selected sinusoidal functions. In this example, the set of the selected sinusoidal time series is termed basis fonctions bocause they are the units from which we ean re create the original time series If one of those sinusoidal functions is croxscorrelated with another one of a different selected frequency, the crosscortelation will be zero. Mathematicians would state that such selected sinusoidal functions are onthogonal (per pendicular) c each other. Thus, such & set of sinusoidal functions is not internally redundant. In recreating the orig- inal time series, no particular sinusoidal funetion can re~ place another one in the summation, In our context, a spec tral or wavel component is simply the exosseorrel coefficient of given basis funetion with seismic data. 122 If we choose to amplify or mute a given component and reconstruct the data from new weighted sums, we ob- tain an altered (filtered) version of the original data. In our example of Fourier analysis, because the basis functions or the selected sinusoidal functions are orthogonal to each other, mathematicians would term the orthogonal Fourier transform to be an orthogonal transform, Such orthogonal transforms provide a minimum number of computed com ponents that represent the measured seismic data. Orthogo- ‘ral implementations of wavelet transforms in particular are clfective al dala compression, Because many of the wavelet components are very small, often we ean represent a trace consisting of a thousand or more seismic samples with only fone-temh as many wavelet components Running-window spectral filtering and spectral-balane ing techniques have been applied to seismic data at least since the later hal of the 1970s, mic trace is broken into a suite of shoxter, overlapping traces centered shout the output sample. Longer windows, provide more-robust statistics, with typical windows bein, 1000 oF 500 ms in length. and with shorter windows of 250 Ins providing poorer results. In pat, the poor results are be~ cause of an assumption that tie underlying reflect vity las, a white spectrum. If the underlying reflectivity is not white it would be distorted during the speetral-balsncing. step, Because of this focus on speciral balancing, spectral ana sis oF shorter windows was overlooked until the middle those applications, each, 1990s. The impetus for using spectral analysis of shorer124 Seismic Altibutes for Prospect Identification and Reservoir Characterization windows ultimately came from seismic interpretation rather than from ssismie processing. In general, seismic interpret ers are quite content with relative spectral measurements, such as the frequency at which tuning occurs, and do not require the absolute value of each frequency component de sired by processors, which are needed t0 reconstruct origi nal data Because we only wish to interpret spectral components rather than to filter components and (efficiently) reconstruct the data, we no longer are bound by orthogonal transforms, As an example, if we wish to analyze seismic data within 100-ms window, Nyguist’s sampling criterion states that swe only need to decompose the data at 10-Hz increments, say, at 10, 20, 30, 40, 50, and 60 Hy for band-limited seismie data with a highest-fequeney contribution between 60 and 70 Hz and no amplitude at 0 Hz. Knowledge of the spectral components at these seven frequencies completely ‘and uniquely describes this time series. Past Fourier trans. Forms (FFT) provide a particularly effective means of gen erating such components Unfortunately, the efficiency of FFT is so ingrained in processors’ heauls, they forget that instead they could use a simple slow Fourier transform — thatis, they could simply crosscorrelate any sine and cosine with the data. In that manner, we will use slow Fourier transforms (more proper ly called diserote Fourier transforms) to decompose the data into sinusoidal componeats separated by Lor 2 Hz. thereby’ allowing the interpreter (0 inspect a more finely sampled spectrum for features of interest. Because the 10-Hz fre quency interval will do the job, the use of the finer frequen cy sampling (1 or 2 Hz) is extra work that provides no ad digonal information Although in theory we could reconstruct any given spectral component at the desired 1-Hz interval from the minimal number of components defined by Nyquis’s exite ion (at 10-112 intervals), itis simpler computationally: to calculate them directly at the desired (1-Hz) sampling, Likewise, orthogonal wavelet-transform components used in data compression also ean be oversampled, allow ing an interpreter to inspect a more Finely sampled spectrum for features of inte We begin this chapter by reviewing hasie concepts of seismie resolution and thin-bed tuning, Then we briefly re ‘view the properties of Fourier transforms, emphasizing the impact of the shape and size of the seismic an dow. That will give us a quantitative understanding of the similarities and differences between two different methods: Shiort-window discrete Fourier ransforms (commonly called spectral decomposition) and wavelet transforms (called in stantaneous spectral analysis). After defining the terms and. esiablishing the thearetieal basis, we provide a suite of ex- amples that exhibit these types of analysis in terms of both ‘workilog and the features that are amenable to the analysis We clase the chapter with a brief overview of the recently in troduced SPICE. (spectral imaging of correlative events) algorithm, showing its zelationship to both instantaneous-at twibute analysis and wavelet decomposition analysis Seismic Resolution and Thin-bed Tuning To better understand the effect of bed thickness on thin-bed resolution, we return to the simple wedge mode! sown in Chapter 5, Figure 1. Generally, the top and bot tom reflections from a thin bed do not have the equal and ‘opposite values routinely ased in seismic modeling of @ thin-bed response, Castagna (2005) recently demonstrated that this overly simplified model, in which reflection coef. ficients have equal magnitude but opposite signs, is patho: logical and that forthe more general case the limits to vert cal resolution are less severe, Accounting for attenuation also may allow us 10 increase the resolution (Goleshubin et al,, 2002; Korneev et al., 2004), Nevertheless, we will follow Widess (1973), Kallweit ‘and Wood (1982), and Roberison and Nogami (1984) and use this wedge model to understand why we can detect thin-bed anomalies even for this worstease model. ‘The maximum constructive interference occurs when the wee thickness is one-quarter of the effective-source wavelength or, when measured in two-way traveltime, itis one-half the thickness of the daminant period (indicated by arrows in Figures le ani Id of Chapter 5). For thickns than that, the wavelorm stabilizes first and then remains constant, and only the seismic amplitude chi thickness. Thus, below tuning, the frequency spectram’s shape does not change with changes in thickness because the waveform does not change shape. For this (worst-case) model (waich may not be valid in actual practice), Widess (1973) showed that when we are Well below the one-quar ter-wavelength tuning thickness, the amplitude changes carly with thickness (Chapter 5, Figure 2) Kallweit and Wood (1982) examined this problem of resolution with @ model consisting of two reflectors thal have equal reflection coefficients of the same sign. The au, thors showed that when the thickness falls below that given by Rayleigh’s criteria (Figure 1), it camnot be estimated from seismic data alone. Kallweit ani Wood’s (1982) paper had a profound impact on the seismic processing commu nity, Although the authors wok eare 10 point out the differ ses smaller ses with ence between detection of relative changes in thickness and resolution (determination) of a given thickness, a whole generation of seismic processors (including the Second au thor of this book) apparently confused the issue and feltthat spectral mapping of reservoirs thinner than one-quarter wavelength would be fruitless, Not until the 1990s did Pav. tyka and his colleagues revisit Widess’s observations and show that amplitude variation with frequency, or more ap- propiately, amplitude variation thickness for a fixed frequency, interpretation tool ross strata of vary ould be used as « powerful Relevant Concepts of Fourier Analysis Fourier analysis is simply crosscorzelation of the seis- mic data with a suite of sines and cosines at predctermined frequencies. Each erosscorrelation coefficient between & given sine and cosine pair and the data is called a frequeney component. Commonly, we use Euler's theorens = cose) sino). 6.) where 0 = 2etf and is the radial frequency measured in radi ans per second, isthe temporal quency measure in lem, and = 41, We then express the erosscomelation coeffi lens of ses and cosines with the data as a complex nut- her, Alen): AMo)= Sieh) = LeosokAndibAny +7Y sineokANdAN (6.2) whore £ is the sample index, Ar isthe time-sampl ment, and «(AA is the seismic data atime #= KAY. It also 's common to detine A(e) in terms of its amplitude, a. vad ts Mo) = a(whe'** 63) Until now, research efforts have focused more on the umplituds component of the spectrum, a(n), However, we pect that changes with respect to Frequency ofthe phase, #(), may allow us to differentiate between upward-fining and upward-coarsening Sequences, both of which have the same amplitude speetram, Although sines and cosines are used for specteal de- composition, other transforms also ean be used 0 decom- pose wavelorms for interpretation, One major software vendor decomposes seismic data using orthogonal Teachy = chey polynomials as a basis function, and another vendor derives (nonorthogonal) basis functions defined hy the dat themselves using cll-organized maps, oF alte ively, using Spectia! Becampostion and Wavelet Transforms 125 (omhogonal) prineipal-component analysis, Regardless of the hasis functioas used, the transform coefficients are ob tained by simply erosscorrelating each basis funetion with a window of the dati in a Fashion analogous to obtaining, the Fourier coefficients through equation 6.2. In this chapter, we limit ourselves to basis Funetions that are windowed sines and cosines and we use either the short window discrete Fourier transform (SWDFT) or the wavelet transform basis functions. WDFT and the wavelet transforms, the basis functions are tapered sines and cosines. They differ from each other in the application of their respective tapers, For the SWDFT, the tapers are independent of frequency and are the same for all sines aad cosines, For wavelet transforms, the fapering windows are proportional to the Frequency of the sines and cosines and are shorter for high- cr frequencies It should come as no surprise that the unwindowed Pourier transform of @ sine!) or coset function has a single nonzero value at o = 0%. However, with the use of these new, tapered-windowed basis functions, we will ob- serve nonzero amplitudes over a range oF frequencies. This smearing in the frequency domain occurs because the basis functions are tapered. Thus, we need to understand the spec~ {rum ofthe tapered window itself omen A Lit al neo = manors «exes mee oe cos Figure 1 with the sam The definition of resolution between 240 reflectors that also has the same I hy the vertieal lives). The effective frequency reflection cveffic of the soutee wavelet is defined as Ih, where b isthe wave jet readth, We can interpret the separation as being ether 1 time [for this chapter on thir-bed resolation) or in spice Uor la resolution occurs when images are sepatated by the peak + chepters on lateral resolution). Rayleigh’s limit of to-trough time interval, wheres Rieker’s Himit nesors when mages ate separated by a time interval equal to the separ tion between inflection points. After Kallweit and Wood126 Seismic Altibutes for Prospect Identification and Reser For the SWDFT, we choose a tapered analysis window, w(t), with length by af the fom 1D cou AAMC DHE 2 re roe ska wikAN= w= 1 ff /bASb-E 0 if lkAM> A) where the window fanetion w(s 1) = w(kA0) will be eon tered about each analysis point at time 7. By tapering the sharp comers of the analysis window, we minimize unde~ sited side lobes in the frequency spectrum (Figure 2b). ‘Typically, in the case of a tapered-windowed DPT, the length of the window, b, will be a fixed S0or 100 ms for all Frequencies, and the length of the taper, § will be 20% of the window length, The same slab of data (of fixed length 50 or 100 ms) will be analyzed by each frequency. We dis. play a suits of SWDPT wavelets and their corvesponding a) _ Mh: al ey Tee te Figure 2. Scismie wavelets and their Fourier spectra at (a) 10, (6 short-window discrete Foutier transform (SWDFT) spectral scomposition analysis. W window tape cosine and sine wavelets, respectiv ), andl (e) 40 Hz, representative of those used in iow width fsa constant 010M ).020 s, Solid and dashed lines indicate the ly. Note the side lobes of the spectra, Also note thal the spectrum of the 10-Hz wavelet extends into negative frequencies i Characterization spectra in Figure 2 and the spe The result of higher or lower frequenci the analyzed spectta up or down, In coatrast, the window functions w(1) that are used in wavelettransform decomposition are typically Gaussian and have the form 1 [etant Vo?| 20? |’ where @ defines the width of the wavelet, For the popular Morlet wavelet transform, 0 = Lif,, where J i8 the central Frequency to be analyzed and is measured in He We plot representative Morlet wavelets and their cor responding spectra in Figure 3 Note that both the temporal window ral width are of fixed size for all Frsquencies, is simply to shift wkAD = w, = (63) 3, Note that in contrast tb the SWDPT. the higher-frequeney Morlet wavelets have short cr temporal extent but broader frequency spectra than the a — po i Figure 3, Seismic center frequencies a (2) IC lets and their Fourier spectra that have (b) 20, and fe) 40 Hr, representa tive of thoce used in the eontinuous-wavelettransfortn (CWT) Morlet wai sion and in instantaneous speetral analysis, asing the window: 1s commorly aze used in wavelet comp defined By wkAry a sample nunnber, Avis the sample inerement, and 0 sold respectively, Note thatthe bandwidth Frequency [este sin Tp dashed lines indicate the cosine and sine wavelets, increases with centercomesponding SWDFT wavelets dispkiyed in Figure Equally important, two Morler of difierent ff queney at the same analysis point will analyze diffe windows of seismic data, Thus, if an interprcter wishes to explicitly analyze the frequency content of a specie fixed temporal-window geologic interval across multiple fre- quencies, the SWDFT is more appropriate for the job. avele rent of correlative events) uses a similar but significantly longer analysis wi dow, o = fi, to provide greater frequency res lution al the expense of reduced temporal resolution (Liner ct al, 2004), We display the SPICE wavelets and their €or responding spectsa in Figure 4. Note that although the wavelets are longer in time, the spectra are narrower than those of the Morlel wavelet displayed in Figure 3. Although the window function, w(), is different for each of the SWDFT, Morlet, and SPICE algorithms, they all are imple- mented by crosscorrelating the data, (2), with windowed i(- Dyeaplion(s~ 7], OF equivalently, by crosscortelating sines and cosines, explin(s ~ 7)), with the windowed data, a7 ula: sines and cosines where Sie.r) are the transform coefficients for eae fre= quency, ©, for an analysis centered about time 7 at the jth sample, »(2) is the analysis window given by equations 6.4 and 6.5, and the seismic data are (V+ 1) samples long. Examining Figures 2 and 3, we observe the following First, as defined, the effective data analysis window for the SWDFT is fixed wheteas that for the wavelet transform is for variable — shorter for higher frequencies and lon; lower frequencies. Second, the speetra of both basis func tions are centered about the center frequency, f-. Thind, the bandwidths of the tapered SWDPT basis functions are inde pendent of frequency, whereas those of the wavelet transform ‘ue narrower at low-center frequencies aed broader at high- center frequencies, Mest important, all spectral components, whether we are using the short-window DFT or the variable window CWT (continuous-wavelet transform), have contri= butions from neighbor preters are not experienced seismic data provessons, we -mphasize that although we may see geologic features oF in- eres al @2-Hz component, rarely do we actually svcord suck low-frequency data Instead, we display the information con- ‘ent from higher frequencies (say, 8-10 Hz) that fall within the tls of the frequency spectra displayed in Figures 2, 3b, and 4b. Such a pilfall is particularly prevalent when we use an insufficiently. tapersd SWDFT, in whieh side Tobe in the frequency spectrum can mix in even more frequencies, frequencios, Because most inter Spectia! Becampostion and Wavelet Transforms 127 To further clarity the SWDET, we display Partyka et als (1900) images comparing SWDET to the spectral hal ancing routinely used in seismie processing. In Figure ‘we show a reflectivity sequence. For simplicity, we assume that the spectra of the earth's reflectivity are white. imply ing that there is an equal probability of a given amplitude reflection cuefficient anywhere within the time series I nnotal, we do notknow the source wavelet. Instead, ‘we process our data to look as though they quired using a bandfimited white source wavelet. We do this by spectrally balancing the measured seismic data, u(). One way of achieving such a balance is 19 crosscorrelate have been ac- the seismic data,» 1, with sines and cosines, ther Fourier components. Often, we compute an aver Spectrum that is representative of all the traces in the sur vey. We calculate the peak spectral amplitutle, yy, 0f that hy gon Figure 4, Seismic waveleis and thelr Fourier spectra st (2) 10, (9) 20, and (c) 40 Ha, representative of those used in the CWT for the SPICE algrith, using the window defined by wiA waa umber, Ar is the sample increment, and = 3if, Solid and dashed lines indicate the eosin tively. The seismic wavelet for 10 Hz ge of this plos 1s heyond the seale128 average spectrum and define a noise threshold, ¢, as a frae: tion ofthat peak amplitude (Figure 6a). Next, we seale each, spectral component by the value LJ{u(f) + Edyach 80 that the new spectral peak is 1.0, the noise threshold is sealed 10 (0.5, and the spectral components below the noise threshold are scaled down toward zero (Figure 6b). This new scaled spectrum is equivalent (0 the spectrum of the idealized ct shown jn Figure 5. Deeonvelution rithms are designed to achieve a similar result ate a band-limited processed-data spectrum by eliminating multiples and reverherations that otherwise would color it The SWDFT shown in Figure 7 differs somewhat from conventional spectral analysis of the ei smic trace. Fh we may sill wish assume that the reflectivity the Fact that we have only 26-S1 samples in our 100-ms analysis window implies that we have only cific realization drawn from a white speetsum. In the spectrum of that windowed reflectivity series is not white, its colored. When the resulting data window, «0, is muluiplied by the white spectrum of the source wavelet, it also has a colored spectrum. We address how to obtain a white source-wavelet spectrum from windowed data in the next section, But if we can, we want the spectrum of the seismie data to be a band-limited version of the spectrum of the reflectivity series within the analysis window. to gener fre Figure 8. Lon; tral decomposition and the convolutional model. Typically, we assume that the miultiple-tree window spec ro reflectivity spectrum has very Time lite structure (inthis ease, we display a white spectum). Next, ‘we calculate the Fourier spec- {rum of the entire (anwindowed) trace. Finally, we explicitly at- ten its spectrum, which, under the assumption of white additive noise, will Time oman fen the spectrum of the source wavelet. Similar assumptions are made i 1 deconvolution algorithms rou tinely used in seismic process. ing. After Partyka et al. (1999), ampiuce Frequency domain Frequency ite spectrum Reflectivity Seismic Altiioutes for Prospect Identification and Reservoir Characterization Spectral Decomposition Using the Short-window Discrete Fourier Transform The workflow for carrying out spectral decomposition using the SWDFT consists of the following steps, which are illustrated in Figure & 1) We select the zone of interest, which typically follows an interpreted horizon, We select the horizon and de: fine 4 constant-thickness slab of data that lie a given number of milliseconds above snd helow the selected horizon in the seismic data volume (Figure 8). 2) Next, we (or our software) extract and then flatten this sab of data (Figure 8b). 3) We (or our software) apply a DFT to the slab of data, frequency by frequency, generating a sequence of con: stant-frequeney spectral-component maps (Figure 8c). 4) We calculate an average or median spectrum represen lative of the entire slab of data, 5) Next, we assume that the geology and therefore reflec: Livity are sufficiently rendom across the entite zone of interest for the average reflectivity spectrum to be white. Source wavelet Noise Seismic data * 8D wn => fl 4 amptuse epee . - | - Band-tmited white spectrum amtuse5) We then calculate a spectal compensation factor Lf Kealft> + 82% yu, Whe {or median) spectral component at frequene where <6 ya denotes the peak amplitude of age (or medisn) speetum I this eompen is applied to the average (or median) spectra, it gener ates a result comparable to that displayed in Figure 6b. Whe this compensation factor is applied to the spectra of the individual windowed traces, it statistically re- moves the effect of the unknown source wavelet, there- hy resulting ia a colored spectrum that is representative of the colored reflectivity within the analysis window. the average f and 7). Finally, we (or our software) reload the spectrally bal- anced frequeney slices into the interpretation worksta: lion for analysis (Figure 80) The spectral decomposition (SWDET) frequency slices allow the intexpreter ( visualize inter 4s thin-bed tuning associated with channels and deltas in plan view. This is not a magical process ereating informa- tion out ef nothing, The SWDET hhas taken 25~S0 time slie- es and combined them with variable weights (the sines and cosines used in erosscorrelation) to form 50-100 frequency slices. The interpreter animates through these images and chooses the images that fit his or her geologic model. Be cause the basis functions are not orthogonal, many of the images are redundant, with much similarity between imag~ ¢s created at neighboring Frequencies, Other images contain dnnly noise. That is the case for images a frequencies at the Jow and high ends of the seismie source’s frequency spee- trum, As the following examples show, identification of tex ures and patterns that are indieative of geologic processes sence patterns, such is proportions! to the interpreters skills and understanding of the depositional environment The SWDEFT can be applied t0 a series of overlapping windows that encompass the reservoir, oe even to th seismic volume (Marfurt and Kirlin, 2001), thereby genes ating a 4D cube (2, y, Irequeney, and time/depth of window center) from a 3D cube. If he lacks 4D intespretation soi- ware, the interpreter simply loads these volumes inio the cntine workstation in multiplexed form. The interpreter ean then cither roll through the resulting slices one at ¢ tine for spectral analysis of one horizon, or pata time (where Nyis the number of frequencies) for common frequency analysis of multiple, Mattened horizons Examples Our first example comes from Partyka et al. (1999), who introduced spectral docomposition to the industry at ©. In Figure 9 we display a conventional amplitude ex traction (Figure 9) and an instantaneous envelope extrac- tion (Figure 9b) aloag a Pleistocene-age horizon from Spectial Decomposition and Wavelet Transforms 129 a) Phere de by ° i Nise tris 05 a Figure 6. The concept of specital balancing achieve a band-limited white spectrum in th. First, we caleulate each spectral component, aff ) and its Next, we maximum, dies timate & noise Level a ft tion, ¢, of the p amplitude. (b) Finally, we reseale fh spectral component by 1 /a{f) +€4aq] 9 that the new pak spectrum is 1.0, South Marsh Island, on the continental shelf in the Gulf of, Mexico. These images are of the paleo-Mississippi River, approximately 150 kat west of the river's current position, We see a complex distributary system, including bifureat- ing channels, point bars, and longitudinal bars that are con- {rolled by Falls in the southeastern portion of the image. In Figure 10 we display images of SWDFT spectral components at 16 and 26 Hz, conesponding to the images, in Figure 9. The narrow channel, A, is poorly imaged in Figure 9 butis elearly visible on the 26-7 spectral compo: nent in Figure 1b, The same channel appears fainter on the Lo-Hz image in Figure 10a, iruplying that 26 Hz is closer to the tuning frequency than 16 Hy is. In contrast, although channel B appears in ll four images in Figures 9 and 10, it hhas maximum lateral resclution in the 16-Hz spectral com> ponent (Figure 10a), in which we see diserete meander loops indicated by the white arrov: ‘The wider channel, B, is better imaged (tuned) in the Tower frequeney, 16-Hz image, whereas the nurrow ehsn nel, A, is better imaged (tuned) at the higher, 26-Hz fie quency, Thal finding is consistent with the well-established comelation between channel width und thiekness. constructed a cantoon to deseribe this phenomenon (Figure 11), in which they note that the130 Figure 7. Short-window spec~ tral decomposition and the convolutional model. Although We may wish to assume that the reflectivity has a more orless white spectrum, any short-win- owed realization of this white distribution will have only a few discrete reflection spikes and thus invariably will have & colored spectrum. If we process the data to provide a band-lim ited white source spectrum and asiume we have white noise, the colored spectrum of the windowed seismie data will be a band-limited representation of the colored spectrum of the reflectivity within the window After Partyka etal. (1999), Reflectivity a(t) — Tine a Time domain J Amplitude Frequency “i Frequency domain Colored spectrum thalweg of deeper channels is imaged by low frequencies, whereas the shallower flanks (and longitudinal and point bars) of the channel are imaged by high frequencies. The above-described model is illustrated clearly in our econd example, taken from a 3D onshore data volume cover a real channel (Figure 12) (Bahorich et al, 2002). Al though an amplitude map showed reasonable detail about the shape of the reservoir, the spectral-decomposition im wges clearly illuminated the thickest and thinnest sequences in the reservoir. Amplitude maps of certain frequencies showed the thinning of levies — information that helped the interpreter map the detailed geometry of the reservoir. That geometry was confirmed subsequently by well control Our third example comes from Peyton et al. (1998), who applied spectral decomposition and coherence to sue~ cessfully image deep (~3500 m) Pennsylvanian stratigraph- ic features in the Anadarko Basin, Oklahoma, U.S.A. The authors merged three different 3D surveys into a single sur- vey covering the area of study (Figure 13). Their objective ‘was to map multiple stages of incised valleys into three coarseaing-upward marine parasequences (the lower, mid- dle, and upper Red Fork sandstone) hounded by the resion- * Seismic Attributes for Prospect Identification and Reservoir Characterization Source wavelet s(i) Noise alt) Seismic data + u(t) => fl Fourier transform & Amplitude Arnpltud Amplitude Band-limitedcolored ‘spectrum, ally extensive Pink limestone above and the Inola Lime- stone Member of the Boggy Shale and the Novi limestone below (Figure 1). Variable sediment fill in the incised channels results in a complex internal architecture that is difficult to interpret on conventional horizon stices through the seismic data. This study focuses on the upper Red Fork incised-valley system, which is the largest such system and which images most clearly on 3D seismic and also contains the best reservoir roeks in the arca. Before acquisition of the 3D surveys, it was believed that the valley fills in this region occurred in four stag ‘with stage TIT being the most abundant hydrocarbon pro: ducer. Of the several wells drilled, some were believed to have penetrated the edge of stage Hil valley fill; however, those wells did not produce, Wells producing from stage 11 sands indicate a gep in the production (shown in yellow in Figures Lb and 15b). The fairly large width (0.75 to 1.5 km) of the Red Fork valley fill, coupled with the extensive well control defining the infill, make the valley fill an at- tractive exploration target. Thus, the area was covered with 3D seismic surveys in an attempt to reduce risk and to ex plore undiled potential in stage IT sands. The seismic dataacquired have good quality, with a dominant frequency of 50 Hz and the bandwidth extending to 80 Hz. The seismic cross section AA’ in Figure 14 shows the lower Skinner Shale marker directly above the Novi lime- stone bounding the Red Fork interval. In between these two markers, the incised valleys are characterized by discontin- uuous reflections of varying amplitudes. Obviously, inter- proters will find it difficult to use traditional interpret techniques (-g., autopicking horizons, amplitude mapping, isochron mapping, ete.) to interpret the Red Fork incised valley fill. Furthermore, it is difficult to identify the indi- vidual stages of fill. As was eventually eared, eros se tion AA’ shown in Figures 12-15 be Red Fork marine parasequences in the south, cuts through Time (6) Figute 9, (a) An amplitude extrac Louisians, U.S.A. After Partyka etal. (1999), Spectial Decempostion and Wavelet Transforms 131 three stages of valley fill, and ends in the regional Red Fork: deposits at the north (Figure 16) Unlike the simple amplitude extraction shown in our carlier Gulf of Mexico example in Figure 9a, reflectivity ‘over the Red Fork incised valley changes significantly with the valley fill, making horizon slices through the seismic data difficult to interpret, For that reason, Peyton et al (1998) used spectral decomposition and coherence to image the edges of the channels and interpret the intemal features, therein, Because the Red Fork interval was $0-ms thick. (Pigure 14), speetral components ranging from 20 Hz c0 50. Hz were computed at 1-Hz intervals within a 50 dow parallel to, but not including, the Novi limestone Figure 15a, we show the 36-lz amplitude slice from Figure 8. Basic steps in short window diserete Fourier transform (SWDET! spectral decomposition. (a) First we define an a Fow parallel to an interpreted ho zon. (b) Next we extract and then atten seismic date falling within the analysis window, (c) Then we calcu late the DFT of the window for each equency by crosscorrelatin sines and cosines. (Steps (b) and (c) are equivalent to directly applying the windowed DFT shown in Figure 2.1 (4) Finally, we spectrally balance the results as defined in Figure 6, and interpret spectral eompenenis through animation, composite dis plays (as in Figures 18, 19, and 22a) 3D visualization (as in Figure 2 statistical analysis (as in Figures and 22b).After Johann etal. (2003). ysis wine with | or on and (b) a full-handwidth instantaneous envelope along a horizon from offshore132 Red Fork spectral decomposition, That slice shows not only the ares, which, seeording to extensive well control, fave heen intespreted to be dif ferent stages of valley fill (Figure 15b). Figure [Seis a corresponding horizon sliee through the coherence volume. Note that here we can see the edges of the channels and different stages of the valley fil, although the intemal details are more pronounced on the spectral de composition image (Figure 5b). We saw in Chapter 3 that, coherence maps lateral variation in waveform and is rela tively insensitive to amplitude. In contrast, spectral decom position is a direct measure of relative amplitude within a given frequency band, We interpret the differences in these to mean that although the amplitudes differ between stage V and stage IIL, the waveform is quite similar. Images from coherence and spectral decomposition le 1996 survey (Fi alley edges but the internal show that sage IIL valley fill erossin a Seismic Altiibutes for Prospect Identification and Reservoir Characterization ue 13) conneets the producing wells in the westem half of the study area with the producing well tthe east, as inter preted before the 3D acquisition, Closer examination of the spectral decomposition and coherence images led to inter pretation of an apparently younger valley tha trends north ‘vesterly in the easiern part of the 1996 Survey and cuts out the stage III fil, but then bends to the southwest and di verges from slage IIT valley Hill, This interesting obscrva tion led workers t9 reintepret the well logs for the ares (Figure 16), and their new interpretation included a young er valley Ml sage V (Figuie 15). In summary, integration of well-log interpretation and the shapes and patteras of features of interest on spectral decompesition and coher: cence images led to a more accurate fina nterprewation. In our fourth example, from Bahorich et al. (2002). preliminary analysis of the 8D seismic data volume yielded the ampliude map shown in Figure 17a, The anomaly Figure 10, SWDFT spectral components at (a) 16 He and (b) 26 He along a horizon from offshore Louisiana, U.S.A., comre- sponding fo the images in Figore 9. Aflcr Partyks et al, (1999) ay by a 2 dere she S u a = AY 15-4 map Figure 11. A schematic diagram showing the ineerrelationship between dhin-bed tuning aad the amplitude of speetral compo nents through an idealized channel, (a) A vertical ross section, and (b) spectral components at a higher frequency and (c) at 3 Tow frequency, shown in map view: On the thinner f ss thin-bed tuning occurs at the higher-frequency components jin green} Inthe thalweg oF emer of the channel, thin-hed tuning occurs atthe lower fiegueneies (in red). After Lasghlin etal. (2002 AAPG02002. Reprinted by permission of the AAPG whose permission is required for further use.Spectia! Becamposition and Wavelet Transforms — 133 a) Figure 12. Three speciral components, increasing in frequency from (a) through (cl. with the highest speciral amplitude plot ted in yellow. Bach spectral component responds uniquely to variations in reservoir thickness. The thickest part of the channel 's seen i (2), thinner parts of the channel appear in (b), nd ihe thinnest ports, including the Nanks of he ehannel, appear ine) Data are courtesy of Apache Corp. After Buhorich etal, (2002). Reprinted by permission of Hart's E&P, Figure 13. 4 location map showing 3D survey outlines, ines of eross section, wells producing fram upper Red Fork stage [II valley fill, and n interpreted outline of the stage II valley before aequisition of 3D, seismic data, Aver Peyton etal. (1998). ea ‘ect Figure 14. Seismic eross section AA’ Ha Peyton etal. (1998), Novi horizon fa) withovt and (b) with channel interpretation. After Copyright134 Seismic Alttioutes for Prospect Identification and Reservoir Characterization marked with a black arrow seemed promising, but conven- position images not only confirmed the existence of the tional displays generated during interpretation did not clar- amplitude anomaly but also illuminated an inferred, subtle ify depositional controls aid trapping mechanisms and — swratigraphic trapping sequence running northeast-south thereby left the prospect’s potential in doubt. Spectral decom- west (Figure 17b). That sequence was not readily seen on a Figure 15. A 30 pretation, (c) A coherence slice 36 ms below the lower Skinner horizon, showing wells with faults in the Red Fork and Inola intervals, After Peyton et al (1998) amplitude slice from spectral decomposition of Red Fork deposits (a) without and (b) with current inter- Figure 16. stratigraphic well-log cross-section AA’ from Figure 13 showing stages Il. 11, IV, and V valley fill. The owpers Pear? Gains sarosen + cross-section datum is the top of the Novi limestone. Upper Red Fork valley stages and regionally correlative limes are shaded. The Griffin 1 well (perforations shown) has pro- duced I.03 bef of gas and 16,400 bbl of oil from Red Fork stage HI sandstones. After Peyton etal, (1998). Figure 17. (a) A conventional broadband amplitude map showing an anomaly. The black arrow indicates an amplitude anom- aly. The dashed blue line indicates fault. (b) Spectral decomposition imaging, which clearly shows a stratigraphic feature as the trapping mechanism to the north, indicated by arrows. After Bahorich et al. (2002). Reprinted by permission of Hart's E&P.the conventional map. Animation through discrete spectral decomposition images helped interpreters understand the stratigraphic setup of this potential reservoir and prepare to exploit the area. Visualizing a suite of spectral components Multispectral seismic images of reservoirs using spec- tral decomposition strongly resemble multispectral optical Images cbtained by remote sensing or satellite imagery techniques. One of the most common false-color image techniques (which will be described in Chapter 9) is to plot three discrete frequencies against red, green, and blue (RGA), Figure 18, from Bahorich et al. (2002), is a time slice combining three spectral frequencies over several channel sequences, offshore West Afriea. Such an image can help in reservoir charucterization and well placement Features that are tuned at a higher frequeney and that are interpreted to correspond to levee complexes appear here to be blue, whereas those tuned at an intermediate frequency are green, The thickest channels appear as orange and y low. Low reflectivity shows up as dark colors. Hall and Trouillot (2004) also plotted spectral compe: nents at 30 Hz, 40 He, and 50) Hz against RGB, where a bright response is represented as white (Figures 19a) Yellow hues in Figure 19d indicate that the red (30-Hz) map and the green (40-Hz) map have coincident high-an: plitude responses. In Figure 19e, we see a redisplay of the authors’ data using the FILS (hue, lightness, and saturation) 135, Figure 18, High-resolution spectral decomposition, show ing details of a complex reservoir image from offshore West, Africa, After Bahorich et sl, (2002). Reprinted by permission of Hart's BAP. Figure 19. Spectral decompo. sition images at (a) 30 Hz, (b) 40 Hz, ond (e) 50 Hz, Blended images using the (d) RGB and (e) HLS color models. After Hall and Trouillot (2004)136 Seismic Attioutes for Prospect Identification and Reservoir Characterization Figure 20. A schematic diagram showing a typical spectrum resulting from spectral decomposition, In general, the mode ofthe spectrum changes strongly and the mean frequency ‘changes only moderately with changes in thin-bed tuning. color model discussed in Chapter 9. Using the HLS color model, the same three amplitude slices are represented by hhue (the wavelength of the coor, Fightness (the brightness Of the color), and saturation (the amount of color tinting adkled), respectively. Then the images are combined to give the final map, in which a bright, saturated red color indi~ caies that all three frequencies are bright in that area, Because a typical seismic spectrum might look like the schematic shown in Figure 20, three components may not show sufficient detail. To address this problem, Johann etal (2003) used optical stacking of a complete suite of 16 fre- «quency components ranging from 5 Hz.t0 80 Hz (Figure 21). Alternatively, we can calculate statistical measures such as the mean, standard deviation, skewness, and kurto- sis of the spectrum, Marfurt and Kirin (2001) found that the med or peak of the spectrum is more sensitive to thin- bed tuning than the mean is. In Figure 22a, we plot an RG image of the 16-Hz, 26-Hz, and 36-Hz spectral components corresponding to the same horizon shown in Figures 9 and 10. In Figure 22b, we plot the peak frequency (mode of the spectrum) against hue and cc-blend it with an image of co- horence plotted against gray seale (c= 0.50; we discuss blending in Chapter 9), Low frequencies (blue) correspond to thicker parts of the channel, and higher Frequencies (yel- low to yellow-green) correspond to thinner channels (white arrow), @ point bar (gray arrow), and a longitudinal bar (black arrow), The combination of coherence and spectral dccomposition is particularly effective, with coherence de lineating channel edges, and the mode of the spectrum indi- cating channel thickness. The Continuous Wavelet Transform and Instantaneous Spectral Analysis Although the short-window diserete Fourie transform hhas been employed frequently with seismic data to provide lateral localization of the frequency content, the nonnegli~ Figure 21. A composite image of spectral components rang- ing between $ and 80 Hz, obtained with optical stacking. Each slice ‘opacity. After Johann etal 2003). assigned the sible-length window limits the technique’s temporal reso: lution (Xia, 1999; Sum et al., 2002; Sinha et al., 2003). In contrast, the length of the continuous-wavelet transform’s (CWT) wavekt is proportional (0 the center frequency, f- as shown in Figure 3. Thus, both narrow-band ringing and broad-band impulsive reflections can be analyzed and posi tioned beter in time. Among CWT workers, 0 in equation 6.5 is called the scale, and T in equation 6.6 is called the translation along the time axis; together they produce a 2D time-scale image for each trace. Liner et al, (2004) used this time-scale map directly in their SPICE algorithm, which we will discuss later in this chapter. However, Sinha et al. (2003) and their colleagues improved the interpretabili map by converting it to a time-frequency map. The CWT approach involves the following steps: fof the time-scale 1) Decompose the seismogram into wavelet components, S(w.1), asa function of the scale, 0, and the translation ff, z, using equation 6.6. 2) Multiply the complex spectrum of each wavelet used in the basis function by its CWT coetficient and suum the result to generate instantaneous frequency gathers. 3). Sort these frequency gathers to produce constant fre- quency cubes, time slices, horizon slices, or vertical sections As with the SWDFT, commercial CWT visualization packages can be adapted to help us interpret the results Matched-pursuit technique Although the CWT discussed above forms a set of oversampled, nonorthogonal basis functions, we stil canAdjust the coeflicients 10 reconstruct the original data, Whereas a means of representing seismic data with the fewest number of wavelet components, the matched-pursuit technigue is designed to provide wavelet components of vo the interpreter: Liu and Marfirt (2005) offered details of such an algo- the flow table of orthogonal CWT provides the most efficient rithm based on Ricker and Morlet wavelets usin shown in Figure 23. They began by precomputi complex wavelets fora finely sumpled suite of The real and imaginary components of those wavelets are simply the cosine and sine wavelets shown in Figure 3a, Next, they generated a complex trace using Hilbert trans- forms, from whieh they calculated the instantaneous enve lope and frequeney of each trace. They then searched for Ihe largest values of the envelope and its comesponding in- stantancous frequency. Next, they least-squares-fitted that suite of wavelets fo the complex trace, solving for complex coefficients that correspond 10 the amplitude and phase of the complex wavelet that best fits the data. Finally, dey subtracted that amplitude and phase-rotated complex wave- let from the current version of the data, thereby generating ‘Unew residual trace. They repeated the process on the Fe- sidual until the enerpy of the residual trace fell below a user-defined threshold Liu and Marfurt (2003) illustrate this process through the example shown in Figure 24. The result after the first iteration is shown in the left column, the result after the fourth iteration is in the center column, and the result after the six reenth iteration isin the right column, ‘The location andl mag- nitude of the wayelet envelope are shown in Figure 24d Each wavelet envelope, frequency, and phase is extracted from the precomputed complex wavelet table and added in To generate the modeled seismic data (Figure 24a). These modeled seismic data are subtracted from the original seis- mic data to generate residual seismie data (Figure 24b). As \eseribed in Figure 23, iterations cease once the residual is sufficiently small Gn this ease, after 16 iterations). Each com plex wavelet has its own precomputed complex spectrum Like the modeled seismic data, these modeled complex spee- tra accumulate as the iterations progress, resulting in a spec tral decomposition fromthe wavelet decomposition. The mag~ nitude of the 40-Hz component is shown in Figure 2: To our knowledge, Castagna et al. (2003) were the t vo develop such a matched-pursuit wavelet-decompesition technique for interpretation. To illustrate the improved res- olution of their instantaneous spectral analysis (ISA) over ihe well-established discrete Fourier transform method, Cast- gna et al. (203) applied both techniques to the synthetic trace we show in Figure 25a, The ISA time-frequency plot [Figure 2Se) shows the amplitude spectra for each time sample. The first event on the synthetic seismogram (Fig- Sa) comes from an isolated reflector and is a single Spectial Decamposition and Wavelet Fansorms 137 a) Figure 22. Alternative means of pl information correspording to the i 9-and 10, (a) blended image using the RGB color model R corresponds to 16 Hz, G 10 25 Hz, and B to 36 He (atter Wessels et al, 1996), (b) A blended image of the mode of the spectrum (plotted against hye) and coherence (plotied &s Th shown in Figures a giay scale, with low col being black| quency (blue) areas correspond to a thicker channel, whereas the high-frequency (yellow) areas corresond to a thinner ray arrow), and a longi channel (white artow), a point bar tudinal bar (black arrow). After Marfurt and Kiel (2001). wavelet (40 Hz). On the frequency spectrum shown in Fig lure 25¢, notice that the duration of the waveler is identical {0 the duration of the arrival in the time domain. Compare this with the SWDET (spectral decomposition) image in Figure 25d, in which the time duration is equal to the win- dow length of 100 ms. ‘The second event is a composite event g adding separate 1-H and 40-Hz events arriving at the same time. The ISA image (Figure quency arrival spread over time and 2 high-frequency ar rival that has a short time duration. The SWDFT result (Figure 25d) shows almost no frequency discrimination be- tween the two events. The side lobes may result from using, erated by 25e) shows a low-fre-138 Seismic Altibutes for Prospect Identification and Reservoir Characterization 4a ess than optimal (aper in an insufficiently long window in the SWDET analysis, The third event is generated by two interfering arrivals oF the same frequency (30 Hz). We can just barely resolve the temporal separation. The fourth event is a composite of four events (at 20 Hy and 30 Hz) arriving at uso different but distinet times, Altos integpret the results 10 be two (25-Hz) events instead of the four events, that were input. The fifth and final event is a composite oF three 20-Hz events that are very closely spaced in time. Although the energy is concentrated around 20 Hz, the events appew 10 bbe somewhat resolved at frequencies greater than 3D Hz. In ast, the SWDFT image shows a broader image, ranging from 10 to 50 Hz. In summary, his example shows that the ISA algorithin has not only greater temporal resolution hut also better frequency resolution than does the SWDIT algorithm. igh the time separetion is acceprable, w smeared x ie 1 for a Morlet wavelet decomposition J:pursuit technique. Afier Liu and Figure 23. 4 flow-eh ith usi Marfust (2005), Each input seismic trace originally displayed ampli tudes as @ finetion of time, but ultimately es plays amplitudes as a function of time and frequency. Simi larly, each 2D seismic line becomes a time, frequency, and CDP number cube, We show eo slices through such an out put cube in Figure 26. The first, Figure 26a, is a time-fre queney slice for a fixed CDP number like those shown in Figure 25.The! igure 26h, is atime versus CDP-numm her eube fora selected frequency. We can interpret such com ion inequeney gathers to be the amplitude ai a given time for a given f trace dis. coon jequency Mapping stratigraphic features As we do with spectral decomposition, we ean use the OWT to identify thin-bed tuning. The displays shown in Figure 27 come from a Waha-Lockridge 3D seismic vol ume from west Tesss, USA. (Sinha et al, 2003), The time Slice in Figurs 27a indicates a channel feature in blue. significant objective here was to try to determine the dispo, Sition and/or extension of the channel in the 3D volume, because that information is erucial for reservoir character ization, Application of the continuous-wavelet transform sivlded a time-frequeney-CDP volume. Figure 27 is 4 20, Hz slice that does not show the channel at all; this may be because the channel is not thick enough to tune. The same time slice for 40 Hz (Figuee 27e) shows the chanel and is Similar to die image in Figure 27a. A higher frequency (95 Hy, in Figure 274) greaily enhances the definition of the channel Feature and reveals. thin meandering channel at the bottom center, Continuous-wavelet transform spectral decomposi ion helps us analyze the data better and extract information Frequency-dependent Behavior of Hydrocarbon Reservoirs Anomalous behavior of gas sands Our first example of frequeney-dependeat behavior is from Alondra field in the Burgos Basin, Mexico (Burnett et al, 2003). Production comes from clean Midway-age low impedance sands that are approximately 20 11 thick, have porosities between 13% and 18%, and lie at a depth of ap, proximately 2000 m. In Figure 28, we display the seismic data and corresponding ISA frequency components at 20 Hz and 40 Hz. At 20 Hz, the reservoir locks similar to other per reflectors. However, at 40 Hz, the reservoir Is sig anily brighter than the other nearby reflections. This anomalous amplitude isthe result of both thin-bed tuning and gas charge. The gas charge makes the reservoir reflecSpectial Decamposition and Wavelet Transforms 139 tion coefficients larger than those in the adjacent brine- from thin, Pliocene-Pleistocene sands believed to be fluvial filled areas, and the thin-bed tuning effect of those large in origin. Reservoir pressures and water levels indicate that reflection coefficients preferentially reflects higher fre- those sands can be quite discontinuous and can have com- quencies, thus making the 40-Hz image brighter than the plex reservoir geometry. Wavelet transforms were used 10 20-Hz image. generate two frequency panels depicting the reservoir Our second example comes from the Macuspana Basin one at 25 Hz (Figure 29a) and one at 35 Hz (Figure 296). in Mexico (Burnett et al., 2003). Production here comes ‘The reservoir appears to be continuous on the 25-He. panel ‘One iteration Four iterations ‘Shteen iterations et decomposition of a seismic line from a survey acquired over the Central Figure 24. Example showing matched-pursuit way Basia Platform, west Texas, U.S.A. Columns represent algorithm results ater the first, fourth, and sixteenth iterations of (a) modeled data, (b) residual (unmodeled) data, (¢) 40-Hz component of the modeled data, and (d) wavelet location and envelope 95); data of the modeled dats. The frequency and phase of the modeted wavelets are not displayed. After Liu and Marfurt( courtesy of Burlington Resources140 but is discontinuous on the 35-Hz panel — whieh is indeed the case, because the two wells have different pressures ‘Well Bitzal 11 is now plugged, but welt Bitzal 9 is still pro- ducing. The extension of the reservoir amplitude to the left fon the 35-Hz panel suggests that the reservoir is thinning in the direction of Bitzal 9 (Figure 29b). As occurred in the previous example, the gas in this reservoir shifts the reso- 4) Fomor a) Freeney (a) Figure 25. (a) A synthetic waveform with (b) transient ariv- als (black) and constituent wavelets (color-coded by center frequency). Comparison of (c) instantaneous spectral analy~ sis (ISA or wavelet-transform analysis) with (d) short-time- ‘window discrete Fourier tansform analysis. Although the short-window DFT has excellent vertical resolution, the fee= ‘queney spectrum has been smoothed by convolution with the spectrum of the window and false events are asscciated with sa etal. (2003) side lobes of transient arrivals, Afler Cast 8) Fequensy Figure 26. (a) A frequency gather and (b) « common fe- ‘quency section obtained by sorting many Frequency gathers according to frequency. The common frequency gathers ‘can be thought of as instantaneous amplitude at given fre- ‘quency. After Castagna otal. (2003) Seismic Atioutes for Prospect Identification and Reservoir Characterization nant Frequency to the higher side, and it can be detected easily when illuminated al this resonant frequency. The ability of the wavelet-transform technique to resolve indi: Vidual reflections in the frequency domain helps us achieve such objectives. Figure 27. (a) A time sice trom Waha-Lockridge 3D data volume in west Texas, showing a channel feature in blue. (b) ), Spectral amplitudes at (b) 20 Hz, (c) Al) Hz, and (a) 95 Ha. The channel feature is not thick enough to be seen on the 20-Hr slice but show’ up well on the 40-Hz slice, At 95 Hz, wwe see that the channel extends to the bottom left indicated by the white anrow. After Sinha et al. (2003). Figure 28. Scismic data and ISA components of a Midway- age ens reservoir, Burgos Basin, Mexico, showing cifferen- tial reflectivity. The reservoir is not anomalous at 20 Hz but exhibits anomalous reflectivity at 40 Hz. After Burnett etl 2003).Our third example comes from Sun et al. (2002) and is a Fractured carbonate line, For which we show corresponding frequency pane's generated using wavelet transforms at 40 Hz (Figure 30b) and at 60 Hz (Figare 30¢), Note that the spectral amplitude below the formation at around 1.250 s is severely attenuat- ed on the 60-Hz panel relative to the 40-Hz panel, This at- tenuation anomaly shows clearly oa the difference section between the 40-Hz and 60-Hz panels (Figure 3d), where the amplitudes cancel above the reservoir formation ad at- tenuation starts at the reservoir top. Sun et al, (2002) attrib uted this attenuation to the thick overlying gas formation, Sun et al, (2002) made a similar observation that we slow in our fourth example, which compares spectral au plitudes computed using wavelet transforms at 20 Hz (Fig- ure 31a) and al 30 Hz (Figure 31b) for a thick reservoir ina different data set, In this example, attenuation at the higher frequency (30 Hz} again suggests a thick reservoir. Having seen two examples of atienuation resulting from an overlying thick gas reservoir, Sun etal. (2002) pro ide our fifth example, which examines a thin gas reservoir. Here we show a seismic line (Figure 32a) and correspond= ing spectral components obtained using wavelet anstorms at 30 Hy (Pigure 32) and at 70 Hz (Figure 32c). The target is at 1.650 s, between the white arows. In contrast to the Previous examples, no attenuation anomaly can be seen here because of the thinness ofthe reservoir. sas reservoir. Figure 3a isa seismic Low-frequency shadows Low-frequency shadows often have been observed be- neath amplitude snomalies associated with ga The term shadow refers 10 a lowering of seisinie frequen- cies seen beneath gas reservoirs. Such low-frequency shad- owsare caused by abnormally high attenuation of high-fre- querey energy in the gas reservoir tell, In relatively thick as reservoirs that offer a sufficient travel path, energy ab: sorption shifts the spectral energy from high to low fre- quencies. Consequently, reflections from just below such reservoirs exhibit anomalously low frequencies and awe been used as substantiating indicators of hydrocarbons To distinguish such low-frequency shadows in seismic data, traditionally: we look for spectral differences in the data from above and below the zone. Thus, we determine an average spectrum above the zone, which pave with an average spectrum below the zone. We ascribe the difference to attenuation within the zone. Conventional lransforms have shortcomings, as ve explained earlier, s0 wavelet tansforms are a good choiee, Figure 33a shows a segment of a seismic section from the Gulf of Mexico (Castagna et al., 2003). The characteris- tie blue trough corresponds to a reservoir, Wavelet trans form-generated frequency sections for 1) Hand 30 Hz. are e then com Spectial Decemposiion and Wavelet Fansorms 141 Figure 29, Isofrequeney panels showing a reservoir that is ‘more anomalous and resolved etter at (b) 35 Hz than at (a) 25 Hz, The reservoir pressures in the two wells are different implying thet the by the white arrow on the 35-Hz component. Novice in par ticular how much farther to the lett ihe reservoir extends on the 85-Hy section. Aflor Rarnett et al (2003) ovo separate reservoirs, as indicated shown in Figures 33 and 33c, respectively. At 10 Hz, the reservoir appears nive and bright, but interesting of abnormally strong low-frequency enemy beneath the reservoir also appears. At 30 Hz, the reservoir still shows, up clearly, but the em visible, Figure 34 shows another example, here exhibiting two Uisdnet reservolrs, fromm dhe North West Shelf uf Australia Figures 34a through © show the frequency sections for 10, 20, and 30 Hz, respectively. At 10 Hz, the brightest event fon the section is the one below the deeper gas pay, which Is, believed to be a low-frequeney shadow. At 20 Hy. the gas. reservoirs show up brighter than the shadow, which still persists. At 30 Hz, the shadow disappears, Two other examples of low-frequeney shadows in images from the Gulf of Mexico are shown in Figures 35 and 36. gy below the reservoir is no longer The diffusive-@ mode! Goloshubin et al. (2002) and Komeey et al, (2004) at- ributed such low-frequency anomalies w a diffusive-Q. ‘model, with attenuation depending on reservoir-fluid mo-