Download as pdf or txt
Download as pdf or txt
You are on page 1of 106

Frequency-Dependent Traveltime

Tomography for Near-Surface Seismic


Imaging

Claudio Jordi

Master of Science Thesis

for the degree of Master of Science in Applied Geophysics at


Delft University of Technology
ETH Zürich
RWTH Aachen University

Supervisors:
Dr. Cedric Schmelzbach
Prof. Dr. Stewart Greenhalgh

Zürich, August 8, 2014


Abstract

Frequency-dependent traveltime tomography does not rely on the high frequency (asymptotic
ray theory) assumption made in classical ray-based tomography. By incorporating the effects
of velocity structures in a the nearby region (called the first Fresnel volume) around the
central ray, it offers a more realistic and accurate representation of the actual physics of
seismic wave propagation more accurately and thus, enhanced imaging of the subsurface
is expected. Improvements in imaging include the recovery of additional information on
the subsurface model, finer resolution and better detection and delineation of low velocity
zones. The fat ray or Fresnel volume tomography algorithm developed within the framework
of this thesis uses the sum of source and receiver (adjoint) traveltime fields, calculated by
finite-difference modelling of the eikonal equation, to determine the Fresnel volumes and
the sensitivity kernels. Using different scale exploration synthetic data examples, the imaging
capabilities of the fat ray tomography algorithm were investigated and compared to the results
of classical ray tomography. The of the velocity fields used to generate the synthetic data
were chosen to emulate two real field data sets, to which the fat ray tomography was also
applied. The first real data example is a large scale data set (profile length > 10 km) acquired
for hydrocarbon search; the second data set was recorded for high-resolution near-surface
imaging of a Quaternary valley (profile length < 1 km). Resolution of the tomograms was
assessed on the basis of checkerboard tests and a column sum of the sensitivity matrix. For
the synthetic data examples as well as for the large-scale real data example, it was possible
to obtain modest, additional information on the subsurface model when fat ray instead of
classical ray tomography was used. However, for the small-scale field example this was not
the case, because ray coverage was very limited in the deeper parts of the inverted model.
Testing different frequencies as input to the fat ray tomography algorithm, it was observed
that the choice of the input frequency should not only be based on the frequency characteristics
of the source wavelet but also partly on the scale of the experiment. Compared to classical
ray tomography, it was generally possible to reduce the smoothing for the fat ray tomographic
inversion. From the synthetic data studies it was not possible to unequivocally demonstrate
superior imaging of low velocity zones using fat ray tomography.
iv Abstract
Acknowledgements

First of all I want to express my deep gratitude to my supervisors Dr. Cedric Schmelzbach
and Prof. Dr. Stewart Greenhalgh. They were always dedicated to help me overcome any
problem I encountered during this project. I always had the feeling of being well guided by
their competent supervision.
I also want to thank OMV for providing me with a data set which was used for this project.
I am thankful for the scholarship provided by E.ON AG. I very much appreciated the financial
support.
I also wish to thank my family for supporting me.
Cheers to my Joint Master’s colleagues, especially the ones I shared the office with. The
balance between silent working, inspiring discussions and fun created a pleasant working
environment.
Special thanks go to Ylenia for her visits while I studied abroad and for her patience when I
wasn’t in the best mood.

Swiss Federal Institute of Technology Claudio Jordi


Zürich, August 8, 2014
vi Acknowledgements
Table of Contents

Abstract iii

Acknowledgements v

1 Introduction 1
1-1 Frequency-dependent traveltime tomography approaches . . . . . . . . . . . . . 1
1-2 Thesis objectives and outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Theory 5
2-1 Fundamentals of ray-based traveltime tomography . . . . . . . . . . . . . . . . . 5
2-1-1 Forward calculation and model parameterisation . . . . . . . . . . . . . . 6
2-1-2 Solution of the inverse problem . . . . . . . . . . . . . . . . . . . . . . . 6
2-2 Fat ray tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2-2-1 Concept of fat ray tomography . . . . . . . . . . . . . . . . . . . . . . . 7
2-2-2 Sensitivities for fat ray tomography . . . . . . . . . . . . . . . . . . . . . 8

3 Implementation of fat ray tomography 11


3-1 Forward and inversion grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3-2 Jacobian matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3-2-1 Fat ray width and weighting scheme . . . . . . . . . . . . . . . . . . . . 12
3-2-2 Sensitivity calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4 Synthetic data examples 17


4-1 FDSIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4-1-1 Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4-1-2 Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4-2 Model A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
viii Table of Contents

4-2-1 Ray-based traveltime tomography . . . . . . . . . . . . . . . . . . . . . 21


4-2-2 Fat ray tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4-2-3 Comparison between ray-based traveltime tomography and fat ray tomog-
raphy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4-3 Model B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4-3-1 Comparison between ray-based traveltime tomography and fat ray tomog-
raphy at different frequencies . . . . . . . . . . . . . . . . . . . . . . . . 28
4-4 Model C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4-4-1 Comparison between ray-based traveltime tomography and fat ray tomog-
raphy at different frequencies . . . . . . . . . . . . . . . . . . . . . . . . 33

5 Exploration-scale data example 37


5-1 Data acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5-2 First break picking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5-2-1 Traveltime reciprocity analysis for data error estimation . . . . . . . . . . 42
5-3 Traveltime tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5-3-1 Ray-based traveltime tomography . . . . . . . . . . . . . . . . . . . . . 43
5-3-2 Fat ray tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5-3-3 Checkerboard tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5-3-4 Comparison between ray-based tomography and fat ray tomography . . . 52
5-4 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5-5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

6 Small-scale high-resolution data example 57


6-1 Traveltime tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6-1-1 Variable velocity gradient initial model . . . . . . . . . . . . . . . . . . . 59
6-1-2 Results of ray-based and fat ray tomography at different frequencies using
the variable gradient initial model . . . . . . . . . . . . . . . . . . . . . 59
6-1-3 Herglotz-Wiechert initial model . . . . . . . . . . . . . . . . . . . . . . . 62
6-1-4 Results of ray-based and fat ray tomography using the Herglotz-Wiechert
initial model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6-2 Resolution assessment: Checkerboard tests and column sum of Jacobian matrix . 65
6-2-1 Checkerboard tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6-2-2 Column sum of Jacobian matrix . . . . . . . . . . . . . . . . . . . . . . 66
6-3 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6-4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

7 Conclusions and Outlook 71


7-1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
7-2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

Bibliography 75
Table of Contents ix

A Fat ray tomography algorithm 79


A-1 Matlab code of forward solver . . . . . . . . . . . . . . . . . . . . . . . . . . 79

B Exploration-scale data: Supplementary information 81


B-1 Alternative initial models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
B-2 25 Hz and 75 Hz fat ray tomography . . . . . . . . . . . . . . . . . . . . . . . . 84
B-3 Differences between results of different tomographic inversions . . . . . . . . . . 85

C Small-scale high-resolution data: Supplementary information 87


C-1 Herglotz-Wiechert initial model: 50 Hz fat ray tomography . . . . . . . . . . . . 87

D Herglotz-Wiechert method 89
x Table of Contents
List of Figures

2-1 Depiction of fat ray concept. The fat ray is constructed by summing the source
and receiver traveltime fields. All points having a summed traveltime smaller than
tsr + T /2 define the fat ray (from Husen and Kissling, 2001). . . . . . . . . . . 8

3-1 Forward and inversion grids. The fat ray (red area) is discretised by forward mod-
elling grid cells. The inversion cells, each having a constant slowness, are defined
on the coarser grid. The number of forward grid cells covered by the fat ray within
the kth inversion cell is denoted ak and A is the total number of forward grid cells
covered by one fat ray. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3-2 Traveltime fields. (a) Source traveltime field (source at position S). (b) Receiver
traveltime field (adjoint source at receiver position R). (c) Sum of receiver and
source traveltime field. The black line is the ray path. The yellow, orange and
red contours delineate the outer boundaries of 200 Hz, 50 Hz and 10 Hz fat rays,
respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3-3 Weight distributions: (a) 200 Hz fat ray, (b) 50 Hz fat ray, (c) 10 Hz fat ray.
Sensitivity distributions calculated according to Eq. (3-3): (d) 200 Hz fat ray, (e)
50 Hz fat ray, (f) 10 Hz fat ray. . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3-4 Fat ray sensitivity distributions for differing frequencies of rays passing through a
low velocity anomaly (left square) and a high velocity anomaly (right square). (a)
2D velocity model. (b) 200 Hz fat ray, (c) 50 Hz fat ray, (d) 10 Hz fat ray. . . . 16

4-1 Velocity models used to generate synthetic shot gathers. (a) Model A: Small scale
two gradient model (blue line on right panel) with localised vp anomalies and vp
jump. (b) Model B: Small scale two gradient model with undulating vp anomaly
(green line on left panel). (c) Model B: Large scale two gradient model with
undulating vp anomaly (green line on left panel). . . . . . . . . . . . . . . . . . 18
4-2 Example of a typical shot gather generated with FDSIM. The first break picks are
marked green. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4-3 Ray-based traveltime tomography for Model A. The black squares mark the posi-
tions of the embedded anomalies (see Figure 4-1). (a) Initial model. (b) Inversion
result. (c) Difference between inversion result and initial model. (d) Difference
between inversion result and true model (shown in Figure 4-1a). . . . . . . . . . 23
xii List of Figures

4-4 50 Hz fat ray tomography for Model A. (a) Inversion result. The black squares
mark the positions of the embedded anomalies (see Figure 4-1). (b) Difference
between inversion result and initial model (Figure 4-3a). (c) Difference between
inversion result and true model in Figure 4-1a. . . . . . . . . . . . . . . . . . . . 24
4-5 Comparison between fat ray and ray-based tomography for Model A. (a) Difference
between the result of fat ray tomography (Figure 4-4a) and ray-based tomography
(Figure 4-3b). The black squares mark the position of the anomalous block inserts
(see Figure 4-1). (b) 1D velocity-depth profile through low velocity anomaly at
x = 725 m. (c) 1D velocity-depth profile through high velocity anomaly at x =
275 m. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4-6 (a) Pattern of undulating velocity anomalies. The arrows indicate the positions of
the 1D velocity-depth profiles shown in Figure 4-8. The polarity of the anomalies is
indicated with + and − signs. (b) Velocity gradient of the initial and background
model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4-7 ”Amplitude percentage patterns” calculated using Eq. (4-1). The undulating ve-
locity anomalies (Figure 4-6a) are perfectly recovered for values of ±10%. The
’+’ and ’−’ signs indicate the true polarity. (a) Pattern for ray tomography. (b)
Pattern for 100 Hz fat ray tomography. (c) Pattern for 50 Hz fat ray tomography.
(d) Pattern for 25 Hz fat ray tomography. . . . . . . . . . . . . . . . . . . . . . 30
4-8 1D velocity-depth profiles through the ray tomogram and the 100 Hz, 50 Hz and
25 Hz fat ray tomograms. (a) Velocity-depth profiles at x = 450 m. (b) Velocity-
depth profiles at x = 550 m. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4-9 (a) Pattern of undulating velocity anomalies. The arrows indicate the positions of
the 1D velocity-depth profiles shown in Figure 4-11. The polarity of the anomalies
is indicated with + and − signs. (b) Velocity gradients of the initial model and
background field. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4-10 ”Amplitude percentage patterns” calculated using Eq. (4-1). The undulating ve-
locity anomalies (Figure 4-9a) are perfectly recovered for values of ±10%. The
’+’ and ’−’ signs indicate the true polarity. (a) Pattern for ray tomography. (b)
Pattern for 75 Hz fat ray tomography. (c) Pattern for 50 Hz fat ray tomography.
(d) Pattern for 25 Hz fat ray tomography. . . . . . . . . . . . . . . . . . . . . . 34
4-11 1D velocity-depth profiles through the ray tomogram and the 75 Hz, 50 Hz and
25 Hz fat ray tomograms. (a) Velocity-depth profiles at x = 2’500 m. (b) Velocity-
depth profiles at x = 7’500 m. . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5-1 Source and receiver elevation along profile (vertically exaggerated by a factor 3). 38
5-2 Sample source gathers. The upper panel in each figure shows the receiver elevation.
Red: First break picks. (a) Source: Vibroseis. (b) Source: Dynamite. . . . . . . 40
5-3 Traveltime residual maps. The traveltime residual is defined as the difference
between the picked traveltime data and the calculated traveltime data for a given
source-receiver pair. (a) Traveltime residuals after a first test inversion using
51’600 traveltime picks. Note the regions of systematic misfits marked by the
two arrows. (b) Traveltime residuals of the result shown in Figure 5-5a (using a
reduced data set of 5’338 first break picks). . . . . . . . . . . . . . . . . . . . . 41
5-4 Initial model for the ray-based and fat ray tomographic inversions. (a) 1D variable
gradient velocity model calculated according to the Evjen function. (b) 2D initial
model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
List of Figures xiii

5-5 Classical ray tomography. (a) Result of ray tomographic inversion. The colour
coding of the velocity values is given in Figure 5-5b. (b) Ray coverage. (c)
Difference between the inversion result shown in Figure 5-5a and the initial model
shown in Figure 5-4b. (d) Traveltime residuals (picked - calculated traveltimes)
versus offset for the initial model (grey) and the inversion result (blue). The black
lines mark the mean traveltime residuals of the inversion result within 200 m offset
bins. The red line indicates zero residual. . . . . . . . . . . . . . . . . . . . . . 45
5-6 Amplitude spectrum of first arriving phase. . . . . . . . . . . . . . . . . . . . . . 46
5-7 Inversion results of fat ray tomography at frequencies of (a) 25 Hz, (b) 50 Hz, (c)
75 Hz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5-8 (a) Difference between the inversion result of the 50 Hz fat ray tomography shown
in Figure 5-7b and the initial model shown in Figure 5-4b. (b) Traveltime residuals
(picked - calculated traveltimes) versus offset for the initial model and the inversion
result. The black lines mark the mean traveltime residuals within 200m offset bins.
The red line indicates zero residual. . . . . . . . . . . . . . . . . . . . . . . . . . 49
5-9 Checkerboard test for 50 Hz fat ray and classical ray tomography. (a) Checkerboard
added to preferred inversion result of the 50 Hz fat ray tomography (Figure 5-7b).
Checkerboard recovered (in percentage terms) using (b) 50 Hz fat ray tomography
and (c) classical ray tomography. The checkerboard would be perfectly recovered
for values of ±10%. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5-10 Comparison between ray tomography and 50 Hz fat ray tomography. (a) Contours
indicating depth range of the model covered by fat rays superimposed on the result
of the ray tomographic inversion. Black: 25 Hz fat ray; Grey: 50 Hz fat ray; White:
75 Hz fat ray. (b) Difference between final model of classical ray and 50 Hz fat ray
tomographic inversions (Figure 5-5a and Figure 5-7b, respectively). . . . . . . . 53
5-11 (a) Tomogram obtained using 50 Hz fat ray tomography. The black bar indicates
the depth of the well (used for the VSP) and it’s projected location on the profile.
The true well location is ∼ 1.5 km away from the seismic line. (b) Geological
cross-section over the anticline (provided by OMV). The depth information is only
approximately known, the length of the cross-section is unknown. (c) Interval
velocity from VSP (red). Blue: Velocity from the tomographic inversion at the
position indicated with the black bar in Figure 5-11a. Green: Initial model. . . . 55

6-1 Source and receiver elevation along profile (vertically exaggerated by a factor 7). 58
6-2 Sample amplitude spectrum of first arriving phase (frequency range: 0 - 250 Hz). 58
6-3 Initial velocity model used by Reiser et al. (2014). (a) 1D velocity function at
position x = 600 m. (b) 2D initial velocity model. . . . . . . . . . . . . . . . . . 59
6-4 Inversion results using the velocity field in Figure 6-3b as initial model. (a) Clas-
sical ray-based tomography. (b) 100 Hz fat ray tomography. (c) 50 Hz fat ray
tomography. (d) 25 Hz fat ray tomography. . . . . . . . . . . . . . . . . . . . . 60
6-5 Velocity residuals (inversion result - initial model) for (a) classical ray tomography
and (b) 50 Hz fat ray tomography. (c) Traveltime residual (picked - calculated
traveltimes) shown for the initial model (light grey), final model from classical
ray tomography (blue) and final model from 50 Hz ray tomography. The red line
indicates zero residual. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6-6 Initial velocity model calculated using the Herglotz-Wiechert (H-W) method. (a)
1D velocity depth profile. Note the maximum depth obtained from the H-W
inversion is at z = 210 m. (b) 2D H-W derived initial velocity model. . . . . . . 62
6-7 Inversion results using the Herglotz-Wiechert initial model. (a) Classical ray to-
mography. (b) 100 Hz fat ray tomography. . . . . . . . . . . . . . . . . . . . . . 63
xiv List of Figures

6-8 Velocity residuals (the inversion result minus the H-W initial model) for (a) clas-
sical ray tomography and (b) 100 Hz fat ray tomography. (c) Traveltime residual
(picked minus calculated traveltimes) shown for the initial model (light grey), fi-
nal model from classical ray tomography (blue) and final model from 50 Hz ray
tomography. The red line indicates zero residual. . . . . . . . . . . . . . . . . . 64
6-9 Checkerboard tests. (a) Checkerboard added to the preferred result of the 50 Hz
fat ray tomography (Figure 6-4c). Checkerboard recovered (in percentage terms)
using (b) classical ray tomography (variable gradient initial model), (c) 100 Hz fat
ray tomography (H-W initial model) and (d) 50 Hz fat ray tomography (variable
gradient initial model). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6-10 Column sum of Jacobian matrix for (a) 50 Hz fat ray tomography (variable gradient
starting model) and (b) 100 Hz fat ray tomography (H-W starting model). (c)
Faded-out 50 Hz fat ray tomogram (Figure 6-4c; variable gradient starting model)
(d) Faded-out 100 Hz fat ray tomogram (Figure 6-7b; H-W starting model) . . . 67
6-11 (a) Tomogram obtained using 100 Hz fat ray tomography (Herglotz-Wiechert ini-
tial model). (b) Tomogram obtained using ray tomography (from Reiser et al.,
2014). The black bar indicates the location and depth of a borehole. (c) Inter-
val velocity from check-shot experiment in the borehole and velocity-depth profile
extracted from tomograms obtained using the variable gradient initial model (left
panel), and the H-W initial model (right panel). The middle panel displays an
interpreted stratigraphic column (from Schmelzbach et al., 2014). . . . . . . . . 68

B-1 1D velocity-depth profile of initial model (left). 2D initial model (right). . . . . . 81


B-2 (a) Inversion result. (b) Difference between the inversion result (Figure B-2a) and
the initial model (Figure B-1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
B-3 1D multi-gradient of initial model (left). 2D initial model (right). . . . . . . . . 82
B-4 (a) Inversion result. (b) Difference between the inversion result (Figure B-4a) and
the initial model (Figure B-3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
B-5 (a) Difference between the result of the 25 Hz fat ray tomography (Figure 5-
7a) and the initial model (Figure 5-4b). (b) Difference between the result of
the 75 Hz fat ray tomography (Figure 5-7c) and the initial model (Figure 5-4b).
(c) Traveltime residuals vs. offset for 25 Hz fat ray tomography. (d) Traveltime
residuals vs. offset for 75 Hz fat ray tomography.. . . . . . . . . . . . . . . . . . 84
B-6 (a) Difference between the results of classical ray and 25 Hz fat ray tomography.
(b) Difference between the results of classical ray and 75 Hz fat ray tomography. 85
B-7 (a) Difference between the results of 50 Hz and 25 Hz fat ray tomography. (b)
Difference between the results of 75 Hz and 25 Hz fat ray tomography.(c) Differ-
ence between the results of 75 Hzand 50 Hz fat ray tomography. . . . . . . . . . 86

C-1 Final model obtained using 50 Hz fat ray tomography and the H-W initial model. 88

D-1 Sketch of a ray path and T − X plot. . . . . . . . . . . . . . . . . . . . . . . . 90


D-2 (a) Raw T − X data (blue). Red: line of best fit. (b) Ray parameter p (dT /dX
values) versus X. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
List of Tables

4-1 List of input parameters for FDSIM modelling. . . . . . . . . . . . . . . . . . . . 20


4-2 Inversion parameters for Model A. . . . . . . . . . . . . . . . . . . . . . . . . . 25
4-3 Inversion parameters for Model B. . . . . . . . . . . . . . . . . . . . . . . . . . 28
4-4 Inversion parameters for Model C. . . . . . . . . . . . . . . . . . . . . . . . . . 33

5-1 Acquisition parameters of the exploration-scale data. . . . . . . . . . . . . . . . 38


5-2 Inversion parameters for classical ray tomography. . . . . . . . . . . . . . . . . . 43
5-3 Inversion parameters for fat ray tomography at 75 Hz, 50 Hz and 25 Hz. . . . . . 47

6-1 Acquisition parameters of the small-scale high-resolution data . . . . . . . . . . . 58


6-2 Inversion parameters (variable gradient initial model). . . . . . . . . . . . . . . . 60
6-3 Inversion parameters (Herglotz-Wiechert initial model). . . . . . . . . . . . . . . 63

B-1 Inversion parameters for classical ray tomography (anticlinal initial model). . . . 81
B-2 Inversion parameters for classical ray tomography (multi gradient initial model). . 83

C-1 Inversion parameters for 50 Hz fat ray tomography (H-W initial model). . . . . . 87
xvi List of Tables
Chapter 1

Introduction

This thesis is a study of the application of frequency-dependent traveltime tomography to


surface seismic data at the exploration and high-resolution scale. Following the approach of
the so-called fat ray tomography, an algorithm was developed and implemented within an
existing ETH classical ray tomography software package. The performance of the algorithm
was examined on both synthetic and real field data examples and was compared against
classical ray-based tomography.

1-1 Frequency-dependent traveltime tomography approaches

Classical ray tomography is based on the high frequency assumption of asymptotic ray theory
that the wavepath is infinitely thin and does not take into account diffraction and other effects
caused by velocity variations away from the ray. Various attempts to take account of the
finite frequency characteristics of wave propagation in tomography have been introduced and
investigated since the 1980’s.
Following the early work of Slaney et al. (1984) and Wu and Toksöz (1987) on diffraction
tomography, Pratt and Goulty (1991) and Woodward (1992) introduced the concept of wave-
equation tomography. Woodward (1992) describes the calculation of wavepaths using the
first-order Born and the first-order Rytov approximations and compared wave-equation to-
mography to standard ray-trace tomography. The latter projects traveltime delays (difference
between observed and calculated times) back over source-receiver ray paths, whereas wave-
equation tomography projects monochromatic, scattered wave fields (total wavefield minus
the incident wavefield, the latter often being difficult to separate in geophysical applications,
as opposed to medical imaging) back over source-receiver wavepaths (Woodward, 1992). In
theory, wave-equation tomography offers better resolution than classical ray-based traveltime
tomography, because it does not rely on the high frequency assumption made in ray tomogra-
phy and furthermore increases the angular coverage of the target through scattering (Schuster
and Quintus-Bosz, 1993; Sheng and Schuster, 2003). Wave-equation tomography is a subset
of full waveform inversion in which the source wavelet needs to be sufficiently well known for
2 Introduction

the calculation of the wavefields (Woodward, 1992). Because both forward and time-reversed
propagation of the wavefield is needed for wave-equation tomography, the method has rather
high computational costs. Other drawbacks of this method are the strong dependence on the
initial model and it’s highly non-linear behaviour (Bai et al., 2013).
A compromise between classical ray tomography and wave-equation tomography is Fresnel
volume or fat ray tomography (e.g. Vasco et al., 1995; Watanabe et al., 1999; Husen and
Kissling, 2001). Fresnel volume tomography takes into account the influence of the first
Fresnel volume on seismic wave propagation. In traveltime tomography this influence is
described by the sensitivity kernel or Fréchet derivative (Tarantola, 1987). The sensitivity
kernels can be derived in different ways. One approach is to calculate the effect of a
velocity perturbation in a given subsurface cell on signal amplitude and phase, according
to the first-order Born or Rytov approximation (Snieder and Lomax, 1996; Spetzler and
Snieder, 2004). Based on this work Jocker et al. (2006) and Liu et al. (2009) propose
monochromatic as well as band-limited amplitude and phase sensitivity kernels expressed
in terms of the source and receiver (adjoint source) Green’s functions. This approach
follows the wave-equation tomography approach and requires finite-difference wave-equation
forward modelling (Liu et al., 2009). Another approach of Fresnel volume or fat ray
tomography is to calculate the traveltime sensitivity kernels using the source and re-
ceiver (adjoint) first arrival traveltime fields (Watanabe et al., 1999; Husen and Kissling,
2001). This is the approach I use in this thesis and it is described in detail in Chapters 2 and 3.

1-2 Thesis objectives and outline

Most fat ray or Fresnel volume tomography approaches have been developed for and applied
to either cross-hole surveys (e.g. Vasco et al., 1995) or imaging of the earth at the global or
regional scale (e.g. Husen and Kissling, 2001; Bai et al., 2013). The application of fat ray
tomography to surface seismic experiments at scales for near-surface seismic imaging is not
so common (e.g. Gance et al., 2012; Chen et al., 2013). Depth scales for near-surface seismic
imaging range from a few tens of metres in engineering and environmental investigations to
several kilometres in hydrocarbon exploration.
Besides the development and implementation of a fat ray tomography algorithm, the main
objective of this thesis is to examine its capabilities and to compare fat ray tomography
with classical ray tomography. The expectation is an improvement in seismic imaging of the
subsurface. In the context of traveltime tomography, improvements in imaging include better
detection and delineation of low velocity zones, an increase in resolution and possibly the
recovery of additional information on the subsurface model. Furthermore, the tomograms
obtained using frequency-dependent traveltime tomography may be used as (better) starting
models for full-waveform inversion (Chen et al., 2013).

In Chapter 2, I review the basic aspects of ray-based traveltime tomography, introduce the
concept of fat ray tomography and the sensitivity calculation based on fat rays. In Chapter
3, the implementation of the fat ray tomography algorithm is described. A substantial part
of the implementation is to adapt the forward solver of the existing ray tomography software
1-2 Thesis objectives and outline 3

package such that the Jacobian matrix containing the sensitivities is set up according to
the concept of fat ray tomography. In Chapter 4, the algorithm is tested on synthetic data
generated from three different synthetic P-wave velocity models. The performance of fat
ray tomography at different frequencies is compared against classical ray-based inversion. In
Chapters 5 and 6 fat ray tomography is applied to two real data sets. The data set described in
Chapter 5 was acquired during a hydrocarbon exploration survey, whereas the data in Chapter
6 were recorded in a small-scale seismic experiment aimed at mapping a Quaternary valley
in Northern Switzerland. The conclusions and an outlook to possible further investigations
given in Chapter 7 finalise this thesis.
4 Introduction
Chapter 2

Theory

In this chapter I review the basic theory of seismic traveltime tomography. The focus is on
the methods used in the 2D traveltime tomography software package ”inv2dm” (Lanz et al.,
1998), which was used for inverting the synthetic and real data examples. Furthermore, the
fat ray tomography scheme which I implemented in the inv2dm algorithm is described in
detail.

2-1 Fundamentals of ray-based traveltime tomography

In seismic traveltime tomography, a set of model parameters m (P-wave velocity) is recon-


structed from a set of observed data dobs (first arrival traveltimes). The model parameters
are related to the data through the equation

d = g(m) (2-1)

where g can be interpreted in the following ways: for the true model (i.e. the true earth)
mtrue , g would be the kernel representing the true physical process within the subsurface
resulting in the observed data dobs , whereas for a synthetic or estimated model mest , g is the
forward operator and g(mest ) is the model prediction. Calculating the synthetic data dpred
for a given model m in Eq. (2-1) is generally referred to as solving the forward problem, or
forward calculation. Reconstructing a subsurface model from the observed data is referred
to as solving the inverse problem. While the forward problem has a unique solution, the
inverse problem does not because the data may be incomplete, inconsistent and subject to
error (Menke, 1989). Furthermore, the inverse problem is usually mixed-determined (over-
and underdetermined model parts). The general strategy to obtain a solution to the inverse
problem is to minimise the difference between observed data and predicted data by adjusting
the model m, subject to certain constraints, regularisation and assumptions on the degree of
model complexity (Rawlinson and Sambridge, 2003).
6 Theory

2-1-1 Forward calculation and model parameterisation

In classical (asymptotic) ray theory the traveltime ti of the ith ray along ray path Li in a
medium of slowness (inverse of velocity) s(r) is given by the line integral:

Z
ti = s(r) dl. (2-2)
Li

One way to parameterise the model is to approximate the slowness field s(r) by m cells, with
each cell having a constant slowness value sk (k = 1, ..., m). With this parametrisation Eq.
(2-2) can be written in discrete form as (Lanz et al., 1998)

m
X
ti = lik uk = Lu (2-3)
k=1

where lik is the segment length of the ith ray in the kth cell, L is the matrix containing the
segments of the ith ray in the ith row and u is the model slowness vector. Common methods
for calculating traveltimes in Eq. (2-3) are ray-tracing and first-arrival wavefront tracking
schemes e.g. eikonal solvers (Rawlinson and Sambridge, 2003).
For calculating the first-arrival traveltime field, the finite-difference solver of the eikonal equa-
tion of Podvin and Lecomte (1991) is used in the inv2dm-algorithm. The ray segment lik
contained in matrix L describes effectively the sensitivity of the traveltime ti to a change
(perturbation) in slowness in cell k. For calculating theses sensitivities and constructing
matrix L the ray paths are determined by ray-tracing.

2-1-2 Solution of the inverse problem

The dependence of the ray path on the unknown slowness makes the inverse problem in
traveltime tomography non-linear. One way of solving non-linear inverse problems is to
linearise them and iteratively solve the linearised inverse problem using, for example, a Gauss-
Newton method (Oldenburg and Li, 2005). Moreover, most inverse problems are under-
determined (fewer independent traveltimes than model parameters) so additional information
needs to be brought to bear. For the under-determined problem infinitely many solutions
exist, i.e. any number of models can explain the observed data equally well, but not all are
geologically plausible. In order to select a particular model the following cost function has to
be minimised (Oldenburg and Li, 2005):

Φ(m) = Φd + βΦm
(2-4)
= k WD (dobs − g(m)) k2 +β k WM (m − mref ) k2

where Φd is the data misfit, Φm is the model objective function subject to regularisation
controlled through parameter β, mref is a reference model (often the initial model), WD is
the data weighting matrix and WM is a combined model weighting matrix consisting of a
2-2 Fat ray tomography 7

smoothing and a damping matrix. The goal of the inversion is to find the model estimate
mest that minimises the overall cost function Φ(m).
For obtaining a model estimate the following system of equations has to solved (adapted from
Lanz et al. (1998) and Oldenburg and Li (2005)):

WD (dobs − dpred ) + Jm
   
J
 αs mini  =  αs I  mest (2-5)
0 α x Wx

with J being the Jacobian matrix containing the partial derivatives ∂ti /∂uk , mini the ini-
tial model, I the identity matrix, Wx the smoothing matrix, αs and αx are damping and
smoothing factors, respectively. The initial model, damping and smoothing are the addi-
tional constraints for finding a solution to the problem described in Eq. (2-5). The damping
factor controls how much the model estimate deviates from the initial model (Marquardt,
1970) and smoothing controls the model roughness (Constable et al., 1987). These two pa-
rameters should be chosen to satisfy a predefined target data misfit, which usually is an
estimate of the error in the data (e.g. picking error). The data weighting matrix WD is often
a diagonal matrix, with the elements being the inverse of an estimated error associated with
each datum.
Writing Eq. (2-5) in a more compact form:

d = Gmest (2-6)

it can be seen that the solution to the inverse problem is given by:

mest = G−1 d (2-7)

where G−1 is the generalised inverse matrix. The system of equations can be solved using
the LSQR method (Paige and Saunders, 1982).

2-2 Fat ray tomography

The 2D fat ray tomography approach presented here follows an approach described in Husen
and Kissling (2001) which combines elements of ray-based traveltime tomography and wave-
equation approaches. The partial derivatives (sensitivities) contained in the Jacobian matrix
are calculated using fat rays (Husen and Kissling, 2001; Bai et al., 2013).

2-2-1 Concept of fat ray tomography

Fat ray tomography tries to approximate the finite frequency characteristics of seismic wave
propagation. The seismic signal, and hence traveltime, is influenced by the first Fresnel
volume, where constructive interference of seismic energy is observed. In terms of traveltimes,
the first Fresnel volume is defined as (Cervenỳ and Soares, 1992):
8 Theory

|tsx + trx − tsr | ≤ T /2 (2-8)

where T = 1/f is the dominant period of the seismic wave having a dominant frequency f ,
tsr is the shortest traveltime between a source located at s and a receiver at r, tsx and trx
are source and receiver traveltime fields, respectively. Equation 2-8 effectively states that any
part of the medium at position x between source and receiver within the first Fresnel volume
surrounding the ray will contribute to the first half cycle of the arriving wave.
The concept of calculating the Fresnel volume is illustrated graphically in Figure (2-1). The
left and centre diagrams show the traveltime fields calculated from the true source and adjoint
source (receiver) positions, respectively, to all points in the medium. These traveltime fields
are obtained through finite-difference modelling of the eikonal equation. Summing the two
traveltime fields and subtracting the direct ray time yields the fat ray, as shown in the right
diagram of Figure (2-1) (Husen and Kissling, 2001).
From Eq. (2-8) it is seen that the width of the fat ray decreases with increasing frequency
(f = 1/T ). The outer boundary of the fat ray is given by:

|tsx + trx − tsr | = T /2. (2-9)

Figure 2-1: Depiction of fat ray concept. The fat ray is constructed by summing the source
and receiver traveltime fields. All points having a summed traveltime smaller than
tsr + T /2 define the fat ray (from Husen and Kissling, 2001).

2-2-2 Sensitivities for fat ray tomography

For a discretised fat ray (see Chapter 3), the sensitivity of the traveltime ti with respect to a
change in velocity vk is given by (Bai et al., 2013, Eq. 17) :

∂ti ak ti
=− i × (2-10)
∂vk Ai v k

where aki is the area of the ith fat ray within the kth cell and Ai is the total area of the ith
fat ray. In order to easily incorporate the sensitivities in the existing inv2dm-algorithm they
were rewritten as slowness sensitivities (S. Greenhalgh, personal communication):
2-2 Fat ray tomography 9

∂ti ∂ti ∂vk


=
∂sk ∂vk ∂sk
 
∂ti ∂ 1
= (2-11)
∂vk ∂sk sk
 
∂ti 1
= − 2 .
∂vk sk

Together with Eq. (2-10) it is then possible to calculate the sensitivity of the traveltime ti
with respect to a change in slowness sk , according to the following equation:

∂ti ak ti 1
=− i × ×− 2
∂sk Ai vk sk
(2-12)
aki ti
= × .
Ai sk

These sensitivities form the basis for constructing the Jacobian matrix in the implementation
of the fat ray tomography approach (see Chapter 3).
10 Theory
Chapter 3

Implementation of fat ray tomography

In this chapter I describe the implementation of the fat ray tomography approach. First, the
discretisation of the subsurface model is explained, followed by a section describing how the
Jacobian matrix is obtained. The form of the Jacobian matrix is the main difference between
fat ray tomography and standard ray-based traveltime tomography algorithms. This chapter
also illustrates the shape of the Fresnel volume (fat) rays. There will be a different fat ray
for each choice of frequency.

3-1 Forward and inversion grids


The subsurface model is discretised by two different grids: A fine grid is used for the forward
computations and a coarse grid is used for the velocity inversion. The two differently sized
grids are sketched in Figure 3-1, along with a hypothetical fat ray (red shading). In this
example one inversion cell is divided into 100 (10 × 10 grid) forward modelling cells.
The forward grid is used for the first-arrival traveltime determination using the eikonal solver.
It needs to be sufficiently small for the finite-difference scheme to run in a stable fashion, to
achieve sufficient accuracy and to avoid undersampling of the slowness model (Podvin and
Lecomte, 1991; Vidale, 1988).
In order to reduce the computational cost but also to reach an acceptable (ray) coverage
within each cell (Lanz et al., 1998), the inversion grid is invariably coarser than the forward
grid. Each inversion cell has a constant slowness value associated with it. The sensitivity
calculations described in the next section are used to define the fat rays on the forward grid.
For the inversion, the fat rays (sensitivities) are mapped from the finer forward grid onto the
coarser inversion grid (see Section 3-2-2).

3-2 Jacobian matrix


The Jacobian matrix is composed of the partial (Fréchet) derivatives of the traveltime with
respect to slowness and is of dimension N × M , where N is the number of data points (i.e.
12 Implementation of fat ray tomography

S R
6
0
*
100
A

200
z

ak
300 forward grid cell

kth inversion cell


400

500
0 100 200 300 400 500 600 700 800 900 1000
x

Figure 3-1: Forward and inversion grids. The fat ray (red area) is discretised by forward modelling
grid cells. The inversion cells, each having a constant slowness, are defined on the
coarser grid. The number of forward grid cells covered by the fat ray within the kth
inversion cell is denoted ak and A is the total number of forward grid cells covered
by one fat ray.

traveltime picks) and M the number of model parameters (slowness values or inversion cells).
The sensitivities associated with the ith ray or fat ray populate the ith row of the Jacobian.
For standard or classical ray-based traveltime tomography these are the lengths of the ray
segments within each inversion cell (Chapter 2), but for fat ray tomography the calculation
of the sensitivity follows from Eq. (2-12).

3-2-1 Fat ray width and weighting scheme

Before calculating the sensitivities, the area (in 2D) or volume (in 3D) covered by the fat
ray has to be determined. This step can be interpreted as the equivalent to ray tracing for
the ray-based algorithm. However, the width of the fat ray is determined without resorting
to comprehensive ray tracing. As shown in Chapter 2, this is achieved through summation
of the source and receiver traveltime field and subtracting the direct ray traveltime from the
summed field. Examples of a source traveltime field and a receiver traveltime field, calculated
with the eikonal solver of Podvin and Lecomte (1991), are shown in Figures 3-2a and 3-2b,
respectively. The summed traveltime field is shown in Figure 3-2c, where additionally the
contours of fat rays calculated with frequencies of 10 Hz, 50 Hz and 200 Hz (red, orange and
yellow lines, respectively) and the paraxial ray path associated with the shortest travel time
(black line) are shown. Note the increasing width of the fat ray with decreasing frequency.
The width of the fat ray is determined together with a linear weighting function w (Watanabe
et al., 1999):
3-2 Jacobian matrix 13

(a) 0 R S (b) 0 R S

200 200
z [m]

z [m]
400 400

1000 800 600 400 200 0 1000 800 600 400 200 0
x [m] x [m]

0.1 0.2 0.3 0 0.1 0.2 0.3


t [s] t [s]

(c) R S
0
0.55
100
0.5

200
0.45
z [m]

t [s]
300
0.4

400 0.35

500
1000 900 800 700 600 500 400 300 200 100 0
x [m]
Ray path 50 Hz fat ray
200 Hz fat ray 10 Hz fat ray

Figure 3-2: Traveltime fields. (a) Source traveltime field (source at position S). (b) Receiver
traveltime field (adjoint source at receiver position R). (c) Sum of receiver and source
traveltime field. The black line is the ray path. The yellow, orange and red contours
delineate the outer boundaries of 200 Hz, 50 Hz and 10 Hz fat rays, respectively.
14 Implementation of fat ray tomography

(
1 − 2f ∆t, 0 ≤ ∆t < 1/2f
w= (3-1)
0, 1/2f ≤ ∆t

with,

∆t = tsx + trx − tsr , (3-2)

with tsx and trx being the source and receiver traveltime fields, respectively, and tsr is the
traveltime of the direct ray. The weight distributions for the 200 Hz, 50 Hz and 10 Hz fat
rays (displayed in Figure 3-2c) are shown in Figures 3-3a, 3-3b and 3-3c, respectively. On
the ray path, where ∆t is zero, w reaches a maximum value of 1. With increasing distance
from the ray path, i.e. increasing ∆t, w decreases linearly to zero (∆t = 1/2f ) and is zero
for t > 1/2f . The threshold of 1/2f follows from the definition of the Fresnel volume (Eq.
(2-8)). The width of the fat ray depends on the frequency f which is defined by the user
and the velocity distribution. The weighting is mapped directly onto the inversion grid when
calculating the sensitivities.

3-2-2 Sensitivity calculation

For calculating the sensitivity J k of traveltime ti with respect to a slowness (or velocity)
perturbation in the kth model cell, the number of forward cells ak covered by the fat ray
within cell k is determined (Figure 3-1). To account for the fact that several forward grid
points fall into one inversion cell, the weights associated with each forward cell within the
inversion cell k are accumulated to give the summed weight wk . Quantity wk is divided by
the sum of all weights W comprising the fat ray to give the expression for the sensitivity for
inversion cell k (following Eq. (2-12)):

wk ti
Jk = × . (3-3)
W sk
The sensitivity distributions within the 200 Hz, 50 Hz and 10 Hz fat rays are shown in Figures
3-3d, 3-3e and 3-3f, respectively. The corresponding weight distributions are given in Figures
3-3a to 3-3c. The fat rays are computed in a 1D medium with a linear velocity increase
from 610 m/s at the surface to 6’000 m/s at 500 m depth. The lower sensitivities close to
the source and receiver result from the lower velocity (higher slowness) values closer to the
surface. Comparing the sensitivity distributions of the differently sized fat rays, the largest
values are reached for the 200 Hz fat ray (Figure 3-3d). This is due to the smaller area covered
by the 200 Hz fat ray compared to the fat rays having lower frequencies. The consequence of
a smaller fat ray is a larger summed weight W , leading to larger values for the sensitivity J k
(Eq. (3-3)). On the other hand, a fat ray covering a larger area has a lower maximum value
of sensitivity (Figure 3-3f).
A second example of sensitivity distributions within the three fat rays, again for frequencies
of 200 Hz, 50 Hz and 10 Hz, is shown in Figure 3-4. Figure 3-4a depicts the 2D velocity model
used to calculate the fat rays (see also Chapter 4). It contains two localised block velocity
anomalies embedded in a constant velocity gradient medium. The block between x = 250 m
3-2 Jacobian matrix 15

(a) 0 1 (d) 0

Sensitivity
0.5

Weight
z [m]

z [m]
200 200
0.5
400 400
0 0
1000 500 0 1000 500 0
x [m] x [m]
(b) 0 1 (e) 0

Sensitivity
Weight
z [m]

z [m]
200 200 0.2
0.5
400 400
0 0
1000 500 0 1000 500 0
x [m] x [m]
(c) 0 1 (f) 0

Sensitivity
0.1
Weight
z [m]

200
0.5 z [m] 200
0.05
400 400
0 0
1000 500 0 1000 500 0
x [m] x [m]

Figure 3-3: Weight distributions: (a) 200 Hz fat ray, (b) 50 Hz fat ray, (c) 10 Hz fat ray. Sensi-
tivity distributions calculated according to Eq. (3-3): (d) 200 Hz fat ray, (e) 50 Hz
fat ray, (f) 10 Hz fat ray.

and x = 300 m is a high velocity anomaly (3’000 m/s), whereas the block between x = 700 m
and x = 750 m is a low velocity anomaly (750 m/s). The background velocity increases
linearly from 750 m/s at z = 20 m (ground surface) to 2’300 m/s at z = 150 m, where it jumps
to 4’000 m/s.
Two sensitivity distributions for the 200 Hz, 50 Hz and 10 Hz fat rays which pass through
the two block inserts are shown in Figures 3-4b through 3-4d. The black squares indicate
the positions of the velocity anomalies. A fat ray of each frequency encounters each of the
two anomalies. The largest sensitivity values (for each frequency) are reached within the high
velocity anomaly (right square in Figure 3-4). For the 200 Hz and 50 Hz fat rays the sensitivity
decreases within the low velocity zone. Sensitivities along the central axis of the fat ray are
smaller inside the low velocity block marked by the left square in Figure 3-4b and 3-4c than
outside it. For the 10 Hz fat ray the decrease of the sensitivity within the low velocity zone
seems to be negligibly small (left square Figure 3-4d).
To sum up, the fat ray sensitivity calculation using Eq. (3-3) follows the approach described
in Husen and Kissling (2001) and Bai et al. (2013). In addition, the weighting scheme used by
Watanabe et al. (1999) is adapted and included in the sensitivity calculation. The sensitivity
is influenced by the frequency, which determines the width of the fat ray, and the velocity
(slowness) field in which the fat rays are calculated. High velocity media increase the sen-
sitivity whereas low velocity media decrease the sensitivity. However, for localised velocity
anomalies (especially low velocity inclusions) the effect of velocity on the sensitivity seems to
depend on the size of the anomaly compared to the width (frequency) of the fat ray. In the
next chapter the performance of the fat ray tomography approach with sensitivity calculated
according to Eq. (3-3) is investigated on synthetic data sets.
16 Implementation of fat ray tomography

(a) 0 4000

Velocity [m/s]
50 3000
z [m]

100 2000
150 1000

1000 900 800 700 600 500 400 300 200 100 0
x [m]

(b) 0
1.5

Sensitivity
50
1
z [m]

100
150 0.5

0
1000 900 800 700 600 500 400 300 200 100 0
x [m]

(c) 0
1

Sensitivity
50
z [m]

100 0.5
150
0
1000 900 800 700 600 500 400 300 200 100 0
x [m]

(d) 0
0.3
Sensitivity

50
0.2
z [m]

100
0.1
150
0
1000 900 800 700 600 500 400 300 200 100 0
x [m]

Figure 3-4: Fat ray sensitivity distributions for differing frequencies of rays passing through a
low velocity anomaly (left square) and a high velocity anomaly (right square). (a)
2D velocity model. (b) 200 Hz fat ray, (c) 50 Hz fat ray, (d) 10 Hz fat ray.
Chapter 4

Synthetic data examples

The fat ray tomography approach was tested on a series of different synthetic data sets. The
synthetic data were obtained using the in-house full wavefield numerical seismic modelling
software packag FDSIM (Finite Difference SIMulation). The different true velocity models
underlying the synthetic data and the associated inversion results of both the classical ray-
based traveltime tomography and the fat ray tomography are shown and compared. The
dimensions of the models and source-receiver layouts used in this synthetic data study were
chosen to simulate the two real data examples presented in Chapters 5 and 6, so as to provide
insight into reconstruction possibilities under the most favourable (noise-free) conditions.

4-1 FDSIM

The parallelised finite difference code FDSIM is designed for synthetic seismogram and wave-
field snapshot computations in fully visco-elastic media based on the most general differential
equations of motion (and constitutive relations) according to the velocity-stress formulation
of Virieux (1986). 2D or 3D simulations in a variety of media are possible, but here only 2D
simulations in acoustic media were used to generate the synthetic data.

4-1-1 Input

For the acoustic wavefield simulations, both a P-wave velocity (vp ) and a density (ρ) model
need to be specified. The different velocity models investigated are shown in Figure 4-1. I
will refer to these models as Model A, Model B and Model C (Figures 4-1a, 4-1b and 4-1c,
respectively). Note the larger horizontal and vertical scales of Model C compared to Model
A and Model B. Each model entails a two-layered background (with or without velocity
gradients or discontinuities). Model A involves flat interfaces and two block inserts whereas
Models B and C involve oscillatory velocity variations (i.e. effectively corrugated interfaces
and periodic inclusions). The left panels in Figure 4-1 show the actual 2D models with vp
colour coded. The right panels of the figures show the 1D slices through the centre of each
18 Synthetic data examples

(a) 0 0

100 100
z [m]

z [m]
200 200

300
900 800 700 600 500 400 300 200 100 0 0 2500 5000
x [m] Velocity [m/s]

500 1000 1500 2000 2500 3000 3500 4000 4500


Velocity [m/s]
0 0
(b)
100 100

200 200

z [m]
z [m]

300 300

400 400

500
900 800 700 600 500 400 300 200 100 0 0 2000
x [m] Velocity [m/s]

500 1000 1500 2000 2500 3000


Velocity [m/s]
(c) 0 0

1000 1000
z [m]
z [m]

2000 2000

3000
9000 8000 7000 6000 5000 4000 3000 2000 1000 0 0 5000
x [m] Velocity [m/s]

1000 2000 3000 4000 5000 6000


Velocity [m/s]

Figure 4-1: Velocity models used to generate synthetic shot gathers. (a) Model A: Small scale
two gradient model (blue line on right panel) with localised vp anomalies and vp
jump. (b) Model B: Small scale two gradient model with undulating vp anomaly
(green line on left panel). (c) Model B: Large scale two gradient model with undu-
lating vp anomaly (green line on left panel).
4-1 FDSIM 19

model with the background vp gradient drawn in blue and the undulating P-wave velocity
anomalies in green (Figures 4-1b and 4-1c).

Model A (Figure 4-1a) simulates a sharp velocity contrast in the subsurface (e.g. bedrock
interface) and consists of two different linear velocity gradients and two localised (high and
low velocity) block anomalies. The dimensions of Model A are 1’000 m in the horizontal (x)
and 300 m in the vertical (z) directions. The uppermost 20 m of the model is the air layer
(vp = 330 m/s). Below the free surface, vp increases from 750 m/s to 2’300 m/s at z = 150 m,
where the P-wave velocity jumps abruptly to 4’000 m/s and thereafter slightly increases to
4’600 m/s at the base of the model (z = 300 m). The two localised anomalies which are
positioned 30 m below the surface are of size 50 m × 50 m. The P-wave speed is 750 m/s for
the low velocity block, and 3’000 m/s for the high velocity block.

In Model B (Figure 4-1b), sinusoidal vp fluctuations are superimposed on a background P-


wave velocity gradient. The horizontal and vertical dimensions of Model B are 1’000 m and
500 m, respectively. Below the air layer at z = 20 m the background P-wave velocity increases
from 750 m/s to 2’350 m/s within 200 m. At that point the velocity gradient (increase with
depth) is reduced such that vp reaches a value of 2’900 m/s at z = 500 m. The sinusoidal
fluctuations have amplitudes of ±10% of the background velocity value and the wavelength
is 200 m in the horizontal direction and 100 m in the vertical direction.

Model C (Figure 4-1c) is the same as Model B but on a much larger (factor of 10) distance
scale. The dimensions are 10’000 m in the x-direction and 3’000 m in the z-direction. The
sinusoidal anomalies, again with an amplitude of ±10% of the background velocity value, have
wavelengths of 2’000 m in the x-direction and 1’000 m in the z-direction. The background P-
wave velocity increases from 500 m/s at the ground surface (z = 200 m) to 5’000 m/s at
z = 2’000 m. For the bottom 1’000 m of Model C the increase of vp with depth is reduced
such that at z = 3’000 m a maximum vp value of 6’200 m/s is reached.

The density model used for the simulations is common to all simulations. It consists of an
air layer having a density of 1.2 kg/m3 and a constant subsurface density of 2’000 kg/m3 . In
addition to these input model values, various other input parameters must be specified in
a parameter input file. Table 4-1 lists the most relevant input parameters for each of the
models. This includes the cell size, number of time steps, boundary conditions, source type,
positions and wavelet shape, number of receivers, receiver increment etc. Note that a 50 Hz
Ricker wavelet was used for all simulations.

4-1-2 Output

The output of the FDSIM simulations are shot gathers, one for each predefined source loca-
tion. A shot gather consists of a number of pressure traces recorded at the specified receiver
locations (Table 4-1). In Figure 4-2 a typical shot gather is shown. The necessary first break
times (marked green in Figure 4-2) for the tomography were picked automatically, using an
amplitude threshold. The first time at which the trace reached a certain threshold value
was taken as the onset time. A threshold of 0.4% of the maximum trace excursion yielded
consistent picks.
20 Synthetic data examples

Model A Model B Model C


No. of cells (x / z) 2’000 / 600 2’000 / 1’000 5’000 / 1’500
Cell size (x / z) 0.5 / 0.5 0.5 / 0.5 2/2
No. of time stepts 30’000 32’000 30’000
Time step size [s] 2e-5 2e-5 1e-4
Model type acoustic 2D acoustic 2D acoustic 2D
Boundary PML PML PML
Source type isotropic pressure isotropic pressure isotropic pressure
Wavelet Ricker Ricker Ricker
Central frequency [Hz] 50 50 50
No. of sources 97 97 49
Position of first source (x / z) [m] 20 / 21 20 / 21 200 / 201
Source increment [m] 10 10 200
No. of receivers 240 240 195
Position of first receiver (x / z) [m] 21 / 20.1 21 / 20.1 200 / 200.1
Receiver increment [m] 4 4 50
Output shot gather shot gather shot gather
Output attribute pressure pressure pressure

Table 4-1: List of input parameters for FDSIM modelling.

−8
x 10
0 1
First break pick
0.1
0.5
0.2
Amplitude
Time [s]

0.3
0

0.4

−0.5
0.5

0.6
−1
50 100 150 200
Receiver Number

Figure 4-2: Example of a typical shot gather generated with FDSIM. The first break picks are
marked green.
4-2 Model A 21

4-2 Model A

Model A served as a test model during the implementation of the fat ray tomography algo-
rithm. As stated above this model simulates the case of a sharp velocity contrast such as the
contact between soil and bedrock, the watertable or the base of weathering. The purpose of
this synthetic test was to establish whether the embedded anomalies and the transition to the
higher velocity (Figure 4-1a) can be faithfully recovered from this type of survey layout. Both,
ray-based and fat ray tomography were used and their performance is compared against each
other. As a measure for the performance of the tomographic inversions the root mean square
(RMS) value of the data residuals (i.e. difference between picked traveltimes and calculated
traveltimes) is used. In this noise-free synthetic data study the goal was to obtain an RMS
value as low as possible, i.e. within the numerical modelling errors (estimated to be < 0.5%).
From the 97 shot gathers, each having 240 traces (Table 4-1), a total number of 23’280 first
arrival traveltimes were ”picked”. However, for the tomographic inversion only 5’880 picked
traveltimes were actually used. It was observed that, with the reduced data set, models
having RMS values comparable to those from the inversion of the full data set were reached.
The benefit of using a reduced data set is the reduction of computational cost (run time and
memory usage).

4-2-1 Ray-based traveltime tomography

The classical ray-based tomography results for Model A are shown in Figure 4-3. The top
panel (Figure 4-3a) shows the initial (starting or first guess) velocity model. For the initial
model an averaged 1D gradient (see also Figure 4-5) was defined and applied over the whole
model dimensions of the true model (given in Figure 4-1a).
Figure 4-3b shows the result of the tomographic inversion for those parts constrained by the
rays. The black squares indicate the positions of the two block velocity anomalies, which
have not been satisfactorily recovered. Velocity values in the entire tomogram range from
750 m/s to about 4’500 m/s. A damping factor of 50 and a smoothing factor of 100 were used
to obtain this result after 7 iterations (Table 4-2). The RMS of the initial model (11.8 ms)
was reduced to a final RMS value of 0.7 ms. The inversion was terminated once the RMS
reduction was no longer considered significant (< 5% between iterations).
Figure 4-3c shows the difference between the inversion result (Figure 4-3b) and the initial
model (Figure 4-3a). The difference (velocity residual) is colour coded such that larger vp
values on the inverted section compared to the initial model are indicated in blue, and lower
vp values coloured in red. White colours indicate no changes from the initial model to the
inverted model. In the upper part of the model (20 m < z < 150 m), the inversion had to
adjust the P-wave velocity of the initial model downwards towards lower values, whereas for
the deeper part the opposite was the case. This is due to the choice of the 1D gradient of
the initial model, which is slightly offset from the two gradients used to make the background
part of the true model (see also Figure 4-5). In the area of the high velocity anomaly the
velocity residual reaches it’s positive maximum of about 1’500 m/s. At the location of the low
velocity anomaly (left square in Figure 4-3c) no pronounced change towards lower vp values is
observed. This is due to the fact that first arriving rays by-pass low velocity zones, meaning
that this part of the model is poorly constrained.
22 Synthetic data examples

In Figure 4-3d the difference between the inversion result (Figure 4-3b) and the true model
(Figure 4-1a) is displayed. Larger vp values on the inverted section compared to the true model
are indicated in blue, lower vp values in red. White colours show that the inversion recovered
the true P-wave velocity values. For large parts of the model above the velocity jump at
z = 150 m, the tomographic inversion properly reconstructed the true velocity distribution.
However, close to areas of strong velocity increase, the vp values are slightly overestimated.
This is due to the smoothing constraint included in the inversion: sharp vp transitions are
smeared out. This also explains the underestimation of the velocity right below the vp jump
at z = 150 m. Below this jump, the model is only constrained by a few rays. But vp values
are again closer to the true values away from the velocity transition. For the low velocity
anomaly, the vp values of the inverted section deviate significantly from the true values. For
the high velocity anomaly only a small area is well recovered. Most of the high velocity
block is underestimated (red colour). However, it can be observed from the inversion result
in Figure 4-3b that the whole area of the high velocity anomaly has slightly higher vp values
compared to it’s immediate surrounds (host rock).

To summarise, the ray-based tomography recovered the transition to high P-wave velocity
values. However, due to smoothing, the sharp boundary in the true model is blurred on
the inverted section. The high velocity anomaly was only partly recovered towards the true
velocity values. The low velocity anomaly is not visible on the tomogram in Figure 4-3b. Only
the missing ray coverage (grey patches in Figure 4-3b) in the area of the low velocity anomaly
give a clue (indirect evidence) to the presence of a low velocity zone. The penetration depth
of the rays is limited to about z = 200 m.

4-2-2 Fat ray tomography

The fat ray tomography was performed using the same data as for the ray-based tomography
(5’880 traveltime picks). The frequency was set to 50 Hz, which is the central frequency of
the Ricker wavelet used in the F DSIM simulation. The frequency value also corresponds to
roughly the centre of the frequency band in actual field seismic surveying. Both the damping
and smoothing factors were set to a value of 50 (Table 4-2). Starting from an RMS value of
11.7 ms for the initial model, an RMS of 0.8 ms was reached after 9 iterations (final RMS error
of ray-based tomography was 0.7 ms). However, in order to reach this RMS the maximum
value of vp allowed during the inversion had to be limited to 4’500 m/s, which is around the
maximum value in the true model. The starting model for the fat ray tomography was the
same as for the standard ray-based traveltime tomography (Figure 4-3a). Due to the greater
depth of penetration observed for the fat ray tomography, the initial model was extrapolated
to a depth of 400 m.

The result of the tomographic inversion using the fat ray algorithm is displayed in Figure 4-4a.
Figure 4-4b shows the difference between the tomogram in Figure 4-4a and the (extrapolated)
initial guess model of Figure 4-3a. The difference between the fat ray tomogram (Figure 4-4a)
and the true velocity model (Figure 4-1a) is shown in Figure 4-4c. White colours in Figures
4-4b and 4-4c indicate no difference between the tomogram and the initial or true model,
respectively. Higher vp values on the tomogram compared to either the initial or true model
are coloured in blue and lower velocity values are coloured in red.
4-2 Model A 23

(a) 0
50
100 4000
z [m]

150
200
250
3000

Velocity [m/s]
900 800 700 600 500 400 300 200 100 0
x [m]
(b) 0
50 2000
100
z [m]

150
200
1000
250

900 800 700 600 500 400 300 200 100 0


x [m]

(c) 0

Velocity residual [m/s]


50
1000
100
z [m]

150 0
200
−1000
250

900 800 700 600 500 400 300 200 100 0


x [m]

(d) 0
Velocity residual [m/s]

50 1000
100 500
z [m]

150 0
200 −500
250 −1000

900 800 700 600 500 400 300 200 100 0


x [m]

Figure 4-3: Ray-based traveltime tomography for Model A. The black squares mark the positions
of the embedded anomalies (see Figure 4-1). (a) Initial model. (b) Inversion result.
(c) Difference between inversion result and initial model. (d) Difference between
inversion result and true model (shown in Figure 4-1a).
24 Synthetic data examples

(a) 0 5000

100 4000

Velocity [m/s]
3000
z [m]

200

2000
300

1000

900 800 700 600 500 400 300 200 100 0


x [m]

(b) 0 1500

1000

Velocity residual [m/s]


100
500
z [m]

200 0

−500
300
−1000

−1500
900 800 700 600 500 400 300 200 100 0
x [m]

(c) 0

1000
Velocity residual [m/s]

100
500
z [m]

200 0

−500
300
−1000

900 800 700 600 500 400 300 200 100 0


x [m]

Figure 4-4: 50 Hz fat ray tomography for Model A. (a) Inversion result. The black squares mark
the positions of the embedded anomalies (see Figure 4-1). (b) Difference between
inversion result and initial model (Figure 4-3a). (c) Difference between inversion
result and true model in Figure 4-1a.
4-2 Model A 25

4-2-3 Comparison between ray-based traveltime tomography and fat ray tomog-
raphy

Table 4-2 lists the inversion parameters for both tomographic inversions. Each inversion
started with a comparable RMS value (11.8 ms for ray tomography and 11.7 ms for fat ray to-
mography). Also the final RMS value of 0.8 ms obtained by fat ray tomography is comparable
to the final value reached with classical ray tomography (0.7 ms). For the fat ray inversion
a smoothing factor half that used in the classical ray inversion was used but the maximum
velocity allowed during the inversion had to be set to 4’500 m/s. The damping factor was set
to a value of 50 for both inversion runs.

Ray Fat ray 50 Hz


No. of data 5’880 5’880
No. iterations 7 9
Damping 50 50
Smoothing 100 50
Max. vp [m/s] – 4’500
RMS initial model [ms] 11.8 11.7
RMS inversion result [ms] 0.7 0.8

Table 4-2: Inversion parameters for Model A.

Looking at the result of the fat ray tomography (Figure 4-4a), the most striking difference
compared to the result of the ray-based traveltime tomography (Figure 4-3b) is the smearing
of high velocity values in the lower part of the tomogram obtained using fat ray tomography
(z > 150 m in Figure 4-4a). Apart from that, similar observations can be made for the 50 Hz
fat ray tomography as for the ray-based tomography, viz., the low velocity anomaly is not
recovered, whereas the high velocity anomaly is.
From Figure 4-4b we find that the fat ray inversion yields similar results to the ray-based
tomography, with the reconstructed model vp values decreased in the upper depth range
compared to the initial model, and increased in the lower part of the section.
Furthermore, the difference between the fat ray tomogram and the true model is similar over
a large part of the model to that from the ray-based tomogram (Figure 4-3d). However, the
localised overestimation of vp in the lower part of the model (two blue areas below z = 150
in Figure 4-4c) are not visible or only very weakly indicated in the case of the ray-based
traveltime tomography (at z ≈ 170 m below the anomalies in Figure 4-3d).
Except for the low velocity block anomaly, both tomographic inversions recovered the main
features of the true velocity model (Figure 4-1a). Figure 4-5 gives a detailed comparison
between the ray tomography and the fat ray tomography results. The point by point difference
in reconstructed velocity values across the entire section is displayed in Figure 4-5a. Velocity-
depth profiles through both tomograms at the positions of the block anomalies i.e. at positions
x = 725 m and x = 275 m (location indicated by arrows in Figure 4-5a), are shown in Figures
4-5b and 4-5c. The velocity-depth curve for fat ray tomography is coloured red, whereas that
for classical ray tomography is coloured in blue. In addition, the true model (green line) and
the initial model (black line) are shown for reference.
26 Synthetic data examples

x=725m x=275m
(a) 0 1500

1000

Velocity residual [m/s]


100
500
z [m]

200 0

−500
300
−1000

−1500
900 800 700 600 500 400 300 200 100 0
x [m]

(b) 0 (c) 0

50 50

100 100

150 150
z [m]

z [m]
200 200

250 250

300 300

True model True model


350 Initial model Initial model 350
Ray tomography Ray tomography
Fat ray tomography Fat ray tomography
400 400
0 2000 4000 6000 0 2000 4000 6000
Velocity [m/s] Velocity [m/s]

Figure 4-5: Comparison between fat ray and ray-based tomography for Model A. (a) Difference
between the result of fat ray tomography (Figure 4-4a) and ray-based tomography
(Figure 4-3b). The black squares mark the position of the anomalous block inserts
(see Figure 4-1). (b) 1D velocity-depth profile through low velocity anomaly at x =
725 m. (c) 1D velocity-depth profile through high velocity anomaly at x = 275 m.
4-3 Model B 27

Figure 4-5a only shows the parts of the image constrained by both inversions. White colours
indicate no difference between the two tomograms, blue colours denote larger vp values on
the ray tomogram, and red colours denote larger values on the fat ray tomogram. We observe
that most of the upper part (z < 150 m) of the tomograms is similar. Differences exist below
the block anomalies and within the high velocity area (anomaly and z > 150 m). There is a
tendency towards higher vp values on the fat ray tomogram for the deeper part of the model.
This tendency is better revealed on the two velocity-depth profiles in Figures 4-5b and 4-5c.
On both tomograms (red and blue curves) the maximum velocity values are reached below
the velocity jump at z = 150 m. The red curve reaches its (limited) maximum of 4500 m/s
and stays at higher vp values compared to the blue curve. At z ≈ 240 m the vp curves tend
towards the values of the initial model, indicating little sensitivity and hence minimal updating
of the model here. Again the smoothness imposed by regularisation on the inversion can be
seen at the velocity jump at z = 150 m (smooth transition rather than a sharp boundary)
and at the location of the high velocity anomaly (Figure 4-5c). Ray tomography slightly
overestimates vp of the high velocity block and has a more localised peak compared to the
fat ray tomography result. Below the anomalous block the ray tomography better tackles
the lower velocities and reaches it’s maximum vp value after the velocity jump faster than
does fat ray tomography. For the low velocity anomaly (Figure 4-5b) the two tomographic
inversions recover very similar vp values.
To conclude, inverting the given data from Model A with the fat ray approach yields a slightly
larger RMS compared to standard ray tomography. It was possible to reduce the smoothing
factor but an additional hard constraint of limiting the maximum vp value allowed in the
inversion was applied. Except for the low velocity anomaly, fat ray as well as classical ray
tomography recovered the main features of the model. However, compared to the ray-based
tomography no improvement (e.g. resolving the low velocity anomaly) is apparent. The
fat rays cover a greater depth range down to (z ≈ 320 m) compared to the central rays
(z ≈ 200 m) but it is questionable how well these deeper parts of the tomogram are resolved.
This question is addressed in a second synthetic data example presented in the next section.

4-3 Model B

The image reconstruction capability of the fat ray algorithm is further investigated here by
means of synthetic data generated from acoustic simulations using Model B (Figure 4-1b)
as velocity input to FDSIM. The model is a checkerboard sitting on a gradual increase of
velocity with depth. A selection of relevant FDSIM input parameters is listed in Table 4-1.
The source and receiver layout is the same as for Model A. As mentioned earlier, Model B
has the dimensions x = 1’000 m and z = 500 m. The true velocity model consists of the un-
dulating velocity pattern shown in Figure 4-6a, superimposed on the two-layered background
gradient displayed in Figure 4-6b. The undulations have maximum amplitudes of ±10% of
the background velocity and wavelengths of 200 m in the horizontal direction and 100 m in
the vertical direction. In this synthetic data study the aim was to recover the undulating
velocity anomalies using classical ray tomography and fat ray tomography, the latter using
different frequencies of 100 Hz, 50 Hz and 25 Hz. Rather than extensively discussing the dif-
ferent tomograms, I focus in this section on the capability of fat ray tomography to recover
the undulating velocity filed.
28 Synthetic data examples

As for Model A, the number of data was reduced to 5’880 traveltime picks. The starting
model for the inversion is the 1D gradient shown in Figure 4-6b, but applied over the whole
model dimensions of the true model. Thus, velocity values of the starting model increase
from 750 m/s at z = 20 m to 2’900 m/s at the base of the model (z = 500 m). Table 4-3
lists the parameters for each tomographic inversion. Because the synthetic data is noise-free,
it was aimed for an RMS value as low as possible (within the numerical modelling errors,
estimated to be < 0.5%). The final RMS value is lowest for the ray tomographic inversion
(0.2 ms). For the fat ray tomographic inversion the RMS values increase with decreasing
frequency. Inverting the data set with 100 Hz, 50 Hz and 25 Hz fat rays yielded final RMS
values of 0.3 ms, 0.6 ms and 2.1 ms, respectively. The RMS error of the initial model is 6.4 ms
for the ray tomography and 6.3 ms for the fat ray tomography. A damping value of 50 was
used for all inversions. Compared to the smoothing value of 100 used for the ray tomographic
inversion, smoothing was reduced for the fat ray tomography (75 for 100 Hz and 50 Hz fat
rays, and 25 for the 25 Hz fat ray). Additionally, smoothing in the vertical direction was
reduced by a factor of 2.

Ray Fat ray 100 Hz Fat ray 50 Hz Fat ray 25 Hz


No. of data 5’880 5’880 5’880 5’880
No. iterations 8 9 9 5
Damping 50 50 50 50
Smoothing 100 75 75 25
RMS initial model [ms] 6.4 6.3 6.3 6.3
RMS inversion result [ms] 0.2 0.3 0.6 2.1

Table 4-3: Inversion parameters for Model B.

4-3-1 Comparison between ray-based traveltime tomography and fat ray tomog-
raphy at different frequencies

An overview of the image reconstruction capabilities of the different tomographic inversions


is shown in Figure 4-7. The images are to be compared against the true velocity undulations
of Figure 4-6a. Additionally, two 1D velocity-depth profiles are shown in Figure 4-8, for a
more detailed examination of the results.
The patterns shown in Figure 4-7 are calculated as follows:

mf inal − mini
mperc = × 100, (4-1)
mtrue

where mf inal is the inversion result, mini the initial model and mtrue the true velocity model.
The pattern, mperc , is the amplitude deviation (in percentage terms) of the inversion result
from the initial model, and essentially removes the background field. Setting mf inal = mtrue
in Eq. (4-1) would yield an mperc having maximum values of ±10%. Thus, the optimum
recovery capability of the inversion would be indicated by an ”amplitude percentage pat-
tern” with maximum aptitudes of ±10%, i.e. the velocity undulations shown in Figure 4-6a
(effectively being mtrue − mini ) would have been recovered perfectly.
4-3 Model B 29

x=550m x=450m
(a) 0 (b) 0
- + - + - + - + - +
100 + - + - + - + - + - 100
- + - + - + - + - +
200 + - + - + - + - + - 200
z [m]

z [m]
- + - + - + - + - +
300 + - + - + - + - + - 300
- + - + - + - + - +
400 + - + - + - + - + - 400
- + - + - + - + - +
500
800 600 400 200 0 0 2000
x [m] Velocity [m/s]

−200 −100 0 100 200


Velocity residual [m/s]

Figure 4-6: (a) Pattern of undulating velocity anomalies. The arrows indicate the positions of
the 1D velocity-depth profiles shown in Figure 4-8. The polarity of the anomalies is
indicated with + and − signs. (b) Velocity gradient of the initial and background
model.

Figure 4-7a shows the amplitude percentage pattern obtained by classical ray tomography,
and Figures 4-7b, 4-7c and 4-7d show the patterns yielded by 100 Hz, 50 Hz and 25 Hz fat ray
tomography, respectively. The ’+’ and ’−’ signs indicate the true polarity (see Figure 4-6) in
the respective field.
The ray tomographic inversion recovers the shallowest anomalies (20 m < z < 70 m, top row)
in Figure 4-7a rather well. The positive amplitudes are only slightly overestimated (> 10%).
Except for the two anomalies at the boundary of the model (x < 100 m and x > 900 m) the
second row of anomalies is also well recovered. However, amplitudes are better recovered in
the centre part of the model (300 m < x < 700 m). In the third row (120 m < z < 170 m) of
anomalies only the positive anomalies in the middle part of the model (300 m < x < 700 m)
are reliably reconstructed. The amplitudes for the negative anomalies in the third row are
underestimated. In the fourth row only the polarity of the anomalies is partly recovered. In
the deepest row (220 m < z < 270 m) which is constrained by ray tomography very little of
the velocity anomalies is recovered. From Figure 4-7a we observe that, for the given surface
seismic survey geometry, the imaging capability of ray tomography is more focussed towards
the centre of the model with increasing depth. Moreover, boundaries between anomalies of
the same polarity are diagonally smeared out with increasing distance from the centre.
Similar observations can be made for the fat ray tomography using a frequency of 100 Hz
(Figure 4-7b). The anomalies are also better recovered in the shallow and centre parts of
the model. Compared to the ray tomography result, a larger area is covered. However,
in most of the additionally covered areas the velocity anomalies are not recovered or only
poorly recovered (white colours at edges of the area covered by the fat rays). The ”diagonal
smearing” observed in Figure 4-7a is again visible. However, in the centre part of the model
the anomalies in the fourth row (170 m < z < 220 m) are clearly better recovered compared
30 Synthetic data examples

(a) 0 (b) 0
- + - + - + - + - + - + - + - + - + - +
100 + - + - + - + - + - 100 + - + - + - + - + -
+ - + - + - + - + - + - + - + - + - +
200 - + - + - + - + 200 - + - + - + - + -
z [m]

z [m]
+ - + - + + - + - + - + -
300 300 - + - + -

400 400

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

−10 −5 0 5 10 −10 −5 0 5 10
% %

(c) 0
- + - + - + - + - + (d) 0
- + - + - + - + - +
100 + - + - + - + - + - 100 + - + - + - + - + -
- + - + - + - + - + - + - + - + - + - +
200 + - + - + - + - + - 200 + - + - + - + - + -
z [m]

z [m]

+ - + - + - + - + + - + - + - + - +
300 + - + - + - + 300 - + - + - + - + -
+ - + - + - + - + - + -
400 400 - + - + -

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

−10 −5 0 5 10 −20 0 20
% %

Figure 4-7: ”Amplitude percentage patterns” calculated using Eq. (4-1). The undulating velocity
anomalies (Figure 4-6a) are perfectly recovered for values of ±10%. The ’+’ and
’−’ signs indicate the true polarity. (a) Pattern for ray tomography. (b) Pattern for
100 Hz fat ray tomography. (c) Pattern for 50 Hz fat ray tomography. (d) Pattern
for 25 Hz fat ray tomography.

to the ray tomography result. Even the negative anomaly in the next deeper row (220 m
< z < 270 m) between x = 500 m and x = 600 m is satisfactorily recovered. However, the
neighbouring positive anomalies are poorly recovered. In the deepest row (270 m < z < 320 m)
covered by the 100 Hz fat rays the velocity anomalies are not recovered at all.
For the fat rays having lower frequencies, large parts of the reconstructed velocity anomaly
patterns look erratic. With a 50 Hz fat ray (Figure 4-7c) only the top row and the centre
part of the second and third row of velocity undulations are acceptably recovered. In large
parts of the area covered by the 50 Hz fat rays the boundaries of the anomalies are smeared
and the shape and amplitudes of the anomalies are not reconstructed correctly. The 25 Hz
fat ray tomography performs worst of all (Figure 4-7c). Only the top row of anomalies is
reliably recovered. For large parts of the target field the inversion fails to recover the velocity
undulations. For some parts amplitudes are drastically overestimated (> 20%) and polarities
are wrong.
Figure 4-8 displays 1D velocity-depth profiles through the ray as well as the different fat ray
tomograms. In each of the tomograms, the 1D profiles were extracted at x = 450 m and
4-3 Model B 31

(a) 0 (b) 0

50 50

100 100

150 150

200 200

z [m]
z [m]

250 250

300 300

350 350
Initial model Initial model
Ray Ray
400 400
Fat ray 100Hz Fat ray 100Hz
Fat ray 50Hz Fat ray 50Hz
450 Fat ray 25Hz Fat ray 25Hz 450
True model True model
500 500
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Velocity [m/s] Velocity [m/s]

Figure 4-8: 1D velocity-depth profiles through the ray tomogram and the 100 Hz, 50 Hz and
25 Hz fat ray tomograms. (a) Velocity-depth profiles at x = 450 m. (b) Velocity-
depth profiles at x = 550 m.

x = 550 m (indicated by arrows in Figure 4-6). In Figure 4-8a the velocity curve for the 25 Hz
fat ray tomography (magenta) is the first to deviate significantly from the true model (green)
at approximately z = 120 m. The 50 Hz fat ray tomography (cyan curve) deviates from the
true model at z = 150 m. Thereafter, the cyan curve is undulating out of phase with the
velocity undulations of the true model. The ray (blue curve) and 100 Hz fat ray (red curve)
tomography exhibit similar results. Both curves deviate from the curve of the true model at
about z = 175 m. At that point the blue curve follows the black curve of the initial model.
The red curve of the 100 Hz fat ray tomography stays slightly closer to the true model and
even gives a hint for the positive undulation of the true model at z = 250 m, before following
the initial model again. For the velocity-depth profile displayed in Figure 4-8b the curve for
the 25 Hz fat ray tomography (magenta) also deviates first (z = 100 m) from the curve of the
true model. The cyan curve of the 50 Hz fat ray tomography deviates from the true model
at z ≈ 175 m and starts undulating out of phase with the true model. The blue curve (ray
tomography) deviates from the true model at z = 200 m and follows the black curve of the
initial model for z > 220 m. Clearly, the 100 Hz fat ray tomography performs best of all. The
deviation from the true model starts at z ≈ 225 m but the undulation of the true model is
still indicated by the red curve down to z = 250 m.
To summarise, in this synthetic data study additional information on the subsurface model
was gained using fat ray tomography. More of the undulating velocity pattern shown in
32 Synthetic data examples

Figure 4-6a was recovered by the 100 Hz fat ray tomography compared to the classical ray
tomography. However, the undulating velocity pattern was not well or even erroneously
reconstructed using fat ray tomography with lower frequencies (50 Hz and 25 Hz). The fact
that the velocity anomalies are better recovered in the centre part of the model is most likely
a consequence of the survey geometry and the (fat) ray coverage associated with it. In areas
where the model is less constrained, i.e. the edges of the covered area, the imaging capabilities
of all tomographic inversions were worse.

4-4 Model C

For the final synthetic data study Model C, which is shown Figure 4-1c, was used as veloc-
ity input to FDSIM. This model is of physical dimensions x = 10’000 m and z = 3’000 m
and consists of a checkerboard velocity pattern (shown in Figure 4-9a) superimposed on the
two-layered background velocity gradient displayed in Figure 4-9b. The sinusoidal velocity
undulations of the checkerboard have maximum amplitudes of ±10% of the background ve-
locity value and wavelengths of 2’000 m in the x-direction and 1’000 m in the z-direction. The
input parameters used in the acoustic FDSIM simulation are listed in Table 4-1. Based on the
survey geometry used for the acquisition of the real data example presented in Chapter 5, the
maximum offset was limited to 5’000 m. This was achieved by having only 99 of the total 195
channels active and simulating a roll-along of the receiver spread with the shot position. For
the first eight shots the first active receiver was at x = 200 m. From the ninth shot onwards,
the receiver spread rolled along with the shot position. For the last nine shots the first active
receiver was again held constant at x = 5’050 m.
As for Model B, the aim of this synthetic data study was to reconstruct the undulating velocity
anomalies depicted in Figure 4-9a, using standard ray tomography and fat ray tomography,
the latter at different frequencies of 75 Hz, 50 Hz and 25 Hz. From a total number of 49 shot
gathers, 4’851 first arrival traveltimes were picked and subsequently inverted. The 1D double-

x=7500m x=2500m
0 0
(a) - + - + - + - + - + (b) 500
1000 + - + - + - + - + - 1000
z [m]
z [m]

- + - + - + - + - + 1500
2000 + - + - + - + - + - 2000
- + - + - + - + - + 2500
3000
8000 6000 4000 2000 0 0 2000 4000 6000
x [m] Velocity [m/s]

−600 −400 −200 0 200 400 600


Velocity residual [m/s]

Figure 4-9: (a) Pattern of undulating velocity anomalies. The arrows indicate the positions of
the 1D velocity-depth profiles shown in Figure 4-11. The polarity of the anomalies
is indicated with + and − signs. (b) Velocity gradients of the initial model and
background field.
4-4 Model C 33

gradient background field shown in Figure 4-9b was applied over the entire true velocity model
to form the starting model for the inversion. The vp values of the starting model increase
from 500 m/s at the ground surface (z = 200 m) to a maximum velocity value of 6’200 m/s
at z = 3’000 m (base of the model). Table 4-4 lists the key parameters for each tomographic
inversion. For all inversions, the RMS error using the initial model is 40.2 ms. With classical
ray tomography and 75 Hz fat ray tomography a final RMS value of 1.4 ms was reached. Using
50 Hz fat ray tomography and 25 Hz fat ray tomography the final RMS values were 1.5 ms
and 2.1 ms, respectively. The damping factor was set to 200 for all inversions. As for Model
B, smoothing was reduced for the fat ray tomography compared to the value used for the
standard ray tomography (100). The smoothing factors used for the fat ray tomography were
75 for 75 Hz and 50 Hz fat rays and 50 for 25 Hz fat rays. Again, smoothing in the vertical
direction was reduced by a factor of 2 compared to the horizontal direction.

Ray Fat ray 75 Hz Fat ray 50 Hz Fat ray 25 Hz


No. of data 4’851 4’851 4’851 4’851
No. iterations 9 9 9 9
Damping 200 200 200 200
Smoothing 100 75 75 50
RMS initial model [ms] 40.2 40.2 40.2 40.2
RMS inversion result [ms] 1.4 1.4 1.5 2.1

Table 4-4: Inversion parameters for Model C.

4-4-1 Comparison between ray-based traveltime tomography and fat ray tomog-
raphy at different frequencies

The image reconstruction capabilities of the tomographic inversions are assessed in the same
way as they were for model Model B. In Figure 4-10 the ”amplitude percentage patterns” are
presented. Figure 4-11 shows two velocity-depth profiles for a more detailed examination of
the results.
As for Model B, the amplitude percentage patterns in Figure 4-10 are calculated according
to Eq. (4-1). The best possible reconstruction of the true checkerboard (Figure 4-9a) would
be indicated with patterns of ±10%. Again, the ’+’ and ’−’ signs denote the polarity of
the true velocity undulations. All tomographic inversions recover the shallowest anomalies
(200 m < z < 700 m) rather well. In the second row (700 m < z < 1’200 m) the anomalies
are also reconstructed quite well except for the anomalies at the edges of the model, i.e.,
x < 1’000 m and x > 9’000 m. However, for the fat ray tomography some velocity smearing,
predominantly for positive anomalies, is observed. This smearing is most prominent for the
75 Hz fat ray image (Figure 4-10b), where the two positive anomalies at 3’000 m < x < 4’000 m
and 5’000 m < x < 6’000 m in the second row are smeared into deeper parts of the model.
The smearing of these two anomalies decreases with decreasing frequency. The 25 Hz fat ray
tomography (Figure 4-10d) does not recover the shapes of the anomalies in the second row
as well as the other inversions do. But except for the anomaly at x < 1’000 m, amplitudes of
the anomalies are quite well recovered and their polarity is correct. Looking at the result of
classical ray tomography in Figure 4-10a, we observe that the ray coverage decreases in the
34 Synthetic data examples

(a) 0
500 - + - + - + - + - + 10
1000 + - + - + - + - + -
z [m]

- - - - -

%
1500 + + + + + 0
2000 - + - - + - +
−10
2500

9000 8000 7000 6000 5000 4000 3000 2000 1000 0


x [m]

0
(b)
500 - + - + - + - + - + 20

1000 + - + - + - + - + - 10
z [m]

- - - - -

%
1500 + + + + + 0
2000 + - + - - + - + - −10
2500 −20

9000 8000 7000 6000 5000 4000 3000 2000 1000 0


x [m]

0
(c) 20
500 - + - + - + - + - +
1000 + - + - + - + - + - 10
z [m]

- - - - -

%
1500 + + + + + 0
2000 + - + - + - + - + - −10
2500 - + -
−20
9000 8000 7000 6000 5000 4000 3000 2000 1000 0
x [m]

0
(d) 20
500 - + - + - + - + - +
1000 + - + - + - + - + - 10
z [m]

- - - - -
%

1500 + + + + + 0
2000 - + - + - + - + - −10
2500 - + -
−20

9000 8000 7000 6000 5000 4000 3000 2000 1000 0


x [m]

Figure 4-10: ”Amplitude percentage patterns” calculated using Eq. (4-1). The undulating ve-
locity anomalies (Figure 4-9a) are perfectly recovered for values of ±10%. The
’+’ and ’−’ signs indicate the true polarity. (a) Pattern for ray tomography. (b)
Pattern for 75 Hz fat ray tomography. (c) Pattern for 50 Hz fat ray tomography.
(d) Pattern for 25 Hz fat ray tomography.
4-4 Model C 35

third row of anomalies (1’200 m < z < 1’700 m). However, those parts of the model which
are constrained by the rays indicate the true polarity of the anomalies. The 75 Hz and 50 Hz
fat ray tomography results shown in Figure 4-10b and 4-10c, respectively, do not recover the
anomalies in the third row. The 25 Hz fat ray tomography performs better than when using
the other two frequencies, with most of the polarities correctly recovered. However, the shape
and magnitudes of the anomalies do not match the true velocity model. In the fourth row of
anomalies (1’700 m < z < 2’200 m) the fat ray tomographic inversions perform in very similar
fashion. For the areas of the model well constrained by the fat rays, polarities are indicated
correctly. However, all tomographic inversions have difficulties in recovering the correct shape
of the anomalies. Differences exist in the recovered amplitudes. The 75 Hz and 50 Hz fat ray
tomography overestimate the amplitude of the positive anomaly at 7’000 m < x < 8’000 m (up
to values > 20% in Figure 4-10b and 4-10c). The 25 Hz fat ray tomography recovers the same
anomaly rather well but overestimates the amplitude of the negative anomaly at 2’000 m
< x < 3’000 m. The amplitude of the same negative anomaly recovered with the higher
frequency topographic inversions is closer to the true value. The amplitude of the positive
anomaly at 3’000 m < x < 4’000 m is overestimated for all fat ray tomographic inversions,
but least so when using 75 Hz fat ray tomography. Some parts of the model between 2’200 m
< z < 2’700 m are constrained by the fat rays but none of the tomographic inversions is able
to reconstruct an anomaly at this depth level.
Figures 4-11a and 4-11b show two 1D velocity-depth profiles through the classical ray to-
mogram and the 75 Hz, 50 Hz and 25 Hz fat ray tomograms at positions x = 2’500 m and
x = 7’500 m (indicated by arrows in Figure 4-9a). From Figure 4-11a we observe that the
curves for the different tomographic inversions are similar down to z = 1’500 m. At that
depth the magenta and cyan coloured curves for the 25 Hz and 50 Hz fat ray tomography,
respectively, match the true amplitude of the positive anomaly best but deviate from the
true model (green curve) for 1’500 m < z < 2’000 m. The negative anomaly with it’s max-
imum amplitude occurring at approximately z = 1’900 m is best recovered by the 75 Hz fat
ray tomography. The ray tomographic inversion does not reconstruct the true amplitudes of
these two anomalies. From the velocity-depth profile shown in Figure 4-11b we observe that
the 25 Hz fat ray tomography (magenta curve) performs best. The magenta coloured curve
closely follows the green curve representing the true model down to z ≈ 2’300 m. The other
curves already slightly deviate from the true model at z ≈ 1’300 m and underestimate the
amplitude of the negative anomaly at that position. The 75 Hz and 50 Hz fat ray tomography
overestimate the maximum amplitude of the next deeper positive anomaly (true position at
z = 2’000 m) and indicate it’s position at too shallow depth (z ≈ 1’900 m). The ray tomo-
graphic inversion (blue curve) gives a hint of this positive amplitude but follows the black
curve of the initial model beyond z = 2’000 m.
From this synthetic data example, similar conclusions can be drawn as for the other example
of Model B. Compared to classical ray tomography, additional information on the subsurface
model (Model C) was gained using fat ray tomography. The major difference compared to
Model B is the performance of the 25 Hz fat ray tomography. On this larger scale model, the
low frequency tomographic inversion is able to reconstruct the true model reasonably well,
even better than the other tomographic inversions. This was not the case for Model B where
the tomographic inversion with 25 Hz fat rays performed worst of all inversions, even though
for both simulations (Model B and C) a Ricker wavelet with a central frequency of 50 Hz was
used.
36 Synthetic data examples

Encouraged by the results of the different synthetic data examples, the fat ray algorithm was
applied to two real field data sets. The first data set, presented in Chapter 5, was recorded
as part of an exploration seismic survey. This survey covered a profile length of 14.6 km and
dictated the dimensions used in Model C. In Chapter 6 the fat ray algorithm is applied to a
data set recorded on a much shorter profile. The length of this profile is 830 m (comparable
to the profile length in Model B).

(a) 0 (b) 0
Initial model Initial model
Ray Ray
Fat ray 75Hz Fat ray 75Hz
500 Fat ray 50Hz Fat ray 50Hz 500
Fat ray 25Hz Fat ray 25Hz
True model True model

1000 1000

z [m]
z [m]

1500 1500

2000 2000

2500 2500

3000 3000
0 2000 4000 6000 8000 0 2000 4000 6000 8000
Velocity [m/s] Velocity [m/s]

Figure 4-11: 1D velocity-depth profiles through the ray tomogram and the 75 Hz, 50 Hz and
25 Hz fat ray tomograms. (a) Velocity-depth profiles at x = 2’500 m. (b) Velocity-
depth profiles at x = 7’500 m.
Chapter 5

Exploration-scale data example

In this chapter, the application of the fat ray tomography algorithm to an exploration-scale
real data example is examined. First the acquisition of the 2D seismic profile is briefly
described. Specific information such as the exact location of the profile is confidential and
cannot be given here. Before moving to the actual traveltime tomography results, the first
break picking necessary to obtain the traveltime data is discussed. The error in the data
was estimated using a traveltime reciprocity analysis. The picked traveltimes were inverted
using both classical ray tomography and fat ray tomography at different frequencies. The
results of the different inversions are discussed and a resolution assessment is made based on
checkerboard tests. At the end of the chapter, a brief interpretation is given, guided by a
sketch of a geological cross-section and a VSP survey close to the profile.

5-1 Data acquisition

The seismic data were acquired in a karstified region in Kurdistan. The profile crosses an
anticlinal structure and is oriented perpendicular to the strike of this structure. Carbonates
have been folded to form the anticline. At the surface the anticline is manifest as a topographic
high. Away from the anticline the lithology changes to dipping layers of sandstone and
siltstone. The geology is discussed in more detail in Section 5-4.
Table 5-1 lists the acquisition parameters of the seismic survey. The profile has a total length
of 14.6 km. Figure 5-1 shows the elevations of the receivers and shot points along the profile.
The surface topography varies by over 700 metres. The total number of shot points is 185, of
which 156 entailed a Vibroseis source (marked red in Figure 5-1) and the remaining 29 being
small buried explosions (marked in green). The dynamite shots were used at the flanks of
the anticline. The nominal spacing between the shot points is 50 m. At the dynamite shot
locations, charges of 2 × 0.75 kg were buried at a depth of 2.4 m. Four in-line Vibroseis units
formed a source array of 25 m length and each unit generated an 8 − 80 Hz upsweep for a total
duration of 8 s. A total number of 590 receivers were employed at a nominal spacing of 25 m.
The maximum number of active channels was 400.
38 Exploration-scale data example

Profile length [km] 14.6


No. Vibroseis shot points 156
No. dynamite shot points 29
Vibroseis source array 4 units
Vibroseis frequency band [Hz] 8-80
Vibroseis sweep length [s] 8
Dynamite source 2 × 0.75 kg at 2.4 m depth
Nominal source spacing [m] 50
No. receivers 590
Nominal receiver spacing [m] 25
No. of channels 400

Table 5-1: Acquisition parameters of the exploration-scale data.

The recorded data are of variable quality along different parts of the profile. In most of the
flat areas of the profile the data quality is relatively good. However, data recorded over the
anticline are of rather poor quality, i.e. characterised by a low signal-to-noise (S/N) ratio.
It is assumed that the observed S/N ratio at this location is a consequence of the rough
topography and the severe karstification of the limestone. Strong scattering and poor source
and/or receiver ground coupling are the principal culprits. Example source gathers are shown
in the next section.

SW NE
1400
Elevation [m asl.]

1200 Receiver
Source: Vibroseis
1000
Source: Dynamite
800
600
14000 12000 10000 8000 6000 4000 2000 0
x [m]

Figure 5-1: Source and receiver elevation along profile (vertically exaggerated by a factor 3).

5-2 First break picking

Two sample and representative common source gathers are shown in Figure 5-2. For the
source gather shown in Figure 5-2a the Vibroseis array was used, whereas the source gather
in Figure 5-2b displays data recorded for a dynamite shot. Individual trace amplitude scaling
has been applied. Note how the relative noise level increases with increasing offset. For
orientation purposes and to appreciate the effects of topography on the recorded traces, the
receiver elevation is plotted above each source gather (see also Figure 5-1). The first break
picks are marked in red. To enhance the first arrivals the data were slightly pre-processed
using a 6-12-60-120 zero-phase bandpass filter for the Vibroseis source gathers and a 7-14-60-
120 minimum-phase bandpass filter for the dynamite shot gathers.
5-2 First break picking 39

Variable data quality along different parts of the profile is apparent on both source gathers.
In particular, the data recorded over the anticline are of poor S/N ratio. The low S/N ratio
hampered first break picking to a large degree in those parts of the profile. For example,
on the source gather displayed in Figure 5-2b data quality decreased for distance positions
< −1’000 m (towards the anticline) and it was only possible to pick first breaks (red dots in
Figures 5-2a 5-2b) for distance positions > −1’000 m. But also localised traces of bad quality
are present in the data, for example in Figure 5-2a the traces around distance position of
−3’200 m, where no first break picks could be made.
Picking was generally easier for near-source traces having absolute offset values < 3’000 m.
For these near-offset traces the S/N ratio was acceptable and allowed picking with an accuracy
roughly estimated to be ±5 ms. However, there are also source gathers for which the S/N
ratio was low for absolute offsets below 3’000 m, e.g. the source gather in Figure 5-2b. The
low S/N ratio and the broadening of the first-arrival peaks caused a decrease in the picking
accuracy with increasing offset (roughly ±10 ms to ±15 ms for offsets > 3’000 m). In a first
round of picking, a total number of 51’600 first breaks were picked from the total of 185 shot
gathers. An initial ray tomographic inversion of the traveltime picks yielded a relatively high
RMS value of 14 ms. This value is comparable to the estimated average data error of about
13 ms, which was derived using a traveltime reciprocity analysis (see Section 5-2-1).
To further investigate the inversion accuracy (and hence the RMS error), traveltime residual
maps were computed and analysed (Figure 5-3). The traveltime residual is the difference
between the observed (picked) data and the data predicted for the inverted model. In Figure 5-
3, shot number is plotted along the horizontal axis and receiver number along the vertical
axis. Each residual is colour coded according to the colour bar shown. Blue colours in
Figure 5-3 indicate larger observed traveltimes compared to the synthetic traveltime data
(positive traveltime residual), whereas red colours depict negative traveltime residuals (larger
calculated traveltimes). Residual maps help to detect systematic problems such as noisy
geophones which appear as horizontal stripes. Likewise, poorly coupled shots are identified
by vertical stripping. Also systematic deviations related to different offsets are detectable as
regions of high or low residuals. The traveltime residual map for the first test inversion which
yielded an RMS value of 14 ms is shown in Figure 5-3a. Beside the large traveltime residuals
for the near-offset region (blue stripe running diagonally in Figure 5-3a), two regions of high
residuals can be recognised (indicated by arrows). One zone of larger traveltime residuals
is located between shot numbers 40 and 80 and between receiver numbers 400 and 550; the
second zone of predominantly blue colours in Figure 5-3a lies between shot numbers 120 and
170 and receiver numbers 200 and 350. The relatively high RMS value and the observations
of systematic misfits on the traveltime residual map (Figure 5-3a) necessitated a review of
the first break picks.
I decided to remove the picks associated with the two zones of high traveltime residuals
described above (see Figure 5-3a). In these two zones the original traveltime picks most likely
did not correspond to the first arriving phase, which was probably masked by the considerable
level of noise. After discarding the unreliable picks, a total number of 42’771 traveltime picks
were left as possible input to the tomography algorithm. But for the reasons given in Section
5-3, the number of picks was further significantly reduced to 5’338 for purposes of tomographic
inversion. The traveltime residual map obtained using ray tomography to invert the reduced
data set is shown in Figure 5-3b. The two zones of high traveltime residuals were removed and
when compared to the first test inversion, the maximum residual was reduced from 110 ms to
40 Exploration-scale data example

(a)
Elevation (m)

SW NE

Elevation (m)
1200 1200
1000 1000
800 800
Offset (m)
-4062 -3062 -2062 -1062 95 1088 2088 3088 4088

100 100
200 200
300 300
400 400
500 500
600 600
Time (ms)

Time (ms)
700 700
800 800
900 900
1000 1000
1100 1100
1200 1200
1300 1300
1400 1400
1500 1500
1600 1600

(b)
Elevation (m)

Elevation (m)
SW NE
1200 1200
1000 1000
800 800
Offset (m)
-4062 -3062 -2062 -1062 -73 938 1938 2938

100 100
200 200
300 300
400 400
500 500
600 600
Time (ms)

Time (ms)

700 700
800 800
900 900
1000 1000
1100 1100
1200 1200
1300 1300
1400 1400
1500 1500
1600 1600

Figure 5-2: Sample source gathers. The upper panel in each figure shows the receiver elevation.
Red: First break picks. (a) Source: Vibroseis. (b) Source: Dynamite.
5-2 First break picking 41

(a)
500
Receiver number

400

300

200

100

20 40 60 80 100 120 140 160 180


Shot number

−100 −80 −60 −40 −20 0 20 40 60 80 100


Traveltime residual [ms]

(b)
500
Receiver number

400

300

200

100

20 40 60 80 100 120 140 160 180


Shot number

−80 −60 −40 −20 0 20 40 60 80


Traveltime residual [ms]

Figure 5-3: Traveltime residual maps. The traveltime residual is defined as the difference be-
tween the picked traveltime data and the calculated traveltime data for a given
source-receiver pair. (a) Traveltime residuals after a first test inversion using 51’600
traveltime picks. Note the regions of systematic misfits marked by the two arrows.
(b) Traveltime residuals of the result shown in Figure 5-5a (using a reduced data
set of 5’338 first break picks).
42 Exploration-scale data example

90 ms (see colour bars in Figure 5-3). However, the large traveltime residuals for the near-
offset range are still present. The reduction of the average (RMS) error in the data was not
striking. A traveltime reciprocity analysis for the reduced data set yielded an average error of
12 ms (see next section for the reciprocity analysis). The final model of the ray tomographic
inversion from which the traveltime residual map in Figure 5-3b was calculated yielded an
RMS error of 11.4 ms (see also Section 5-3-1).

5-2-1 Traveltime reciprocity analysis for data error estimation

For the estimation of the overall (average) error or uncertainty in the data a traveltime reci-
procity analysis was carried out (e.g. Zelt, 1999; Schmelzbach et al., 2008). The general idea is
to extract reciprocal source-receiver pairs and calculate the difference between the reciprocal
traveltimes, i.e., interchanging source and receiver positions, and comparing traveltimes in
each case. Ideally, traveltimes from two source-receiver pairs, having source and receiver at
the same locations but interchanged on the second occasion, should yield the same travel-
times following reciprocity principles of common raypaths. Any differences can be attributed
to changed ground conditions (e.g. a disturbed zone around a former shot location, causing
travel time delays on the reciprocal measurement) or would indicate an error in the picked
traveltimes. In practice, source and receiver locations rarely coincide. Thus, a maximum
distance allowed between a source and a reciprocal receiver position (at the nominal station
location) is specified when searching for coincident source and receiver positions. All source-
receiver pairs having a separation less than the maximum distance allowed are selected and
their reciprocal traveltime differences are calculated.
For the reciprocity analysis of the reduced traveltime data set (5’338 picks, see next section)
the maximum distance allowed between a former source position and a new receiver position
was set to 60 m, yielding 1’141 reciprocal differences. In other words, we allowed a maximum
discrepancy in reciprocal station positions of 60 m. Assuming a normal distribution of the
reciprocal differences, the standard deviation was estimated to be 12 ms. This value, inter-
preted as a (conservative) data uncertainty estimate, was used as the target RMS error for
the tomographic inversions described in the next section.

5-3 Traveltime tomography

The picked first arrival traveltimes were inverted using classical ray tomography and fat ray
tomography at different frequencies. As for some of the synthetic data examples (Chapter 4),
the amount of traveltime picks was reduced for the tomographic inversions. By selecting only
picks from every second shot and every fourth receiver in the data set (effectively reducing the
nominal source and receiver spacing to 100 m), the amount of data was reduced from originally
42’771 to 5’338 picks. As already stated in Chapter 4 the reduction of data decreased the
computational cost considerably, while the models obtained from the inversions of the reduced
data set have comparable RMS values and were essentially identical to those models obtained
by inverting the full data set.
As the input model to all tomographic inversions, a 2D variable velocity field (Figure 5-4b)
based on a 1D Evjen function (Figure 5-4a) as described in Greenhalgh and King (1981)
5-3 Traveltime tomography 43

was chosen. The parameters of the Evjen function were found by trial an error. Paralleling
the surface topography in the model, the variable gradient velocity model was applied over
the whole model dimensions, to yield the initial model shown in Figure 5-4b. To assess the
dependency of the inversion result on the initial model, also alternative initial models were
tested as input to the inversions. These models are shown in Appendix B, together with the
associated inversion results. All results are rather similar and exhibit the same features as
the result described in the next section.

(a) 0 (b) 0

1000

z [m]
1000 2000
z [m]

2000 14000 12000 10000 8000 6000 4000 2000 0


x [m]

3000 2500 3000 3500 4000 4500 5000 5500 6000 6500
2000 4000 6000
Velocity [m/s] Velocity [m/s]

Figure 5-4: Initial model for the ray-based and fat ray tomographic inversions. (a) 1D variable
gradient velocity model calculated according to the Evjen function. (b) 2D initial
model.

5-3-1 Ray-based traveltime tomography

Table 5-2 lists the parameters of the ray tomographic inversion. The RMS error of the initial
model is 74.4 ms. This value was reduced within 4 iterations to the final RMS value of
11.4 ms, which is close to the target RMS value (∼ 12 ms). The damping and smoothing
factors were set to values of 100 and 4’000, respectively. Additionally, the smoothing in the
vertical direction was reduced by a factor 2. Damping and smoothing values were found
through test runs of the inversion using various numbers for damping and smoothing. For a
first test run of the inversion, damping and smoothing were both set to high values (e.g. a
damping factor of 6’400 and a smoothing factor of 8’000). For the subsequent inversion run the
damping factor was halved, while the smoothing was kept constant. This was continued until
the smallest damping factor was found such that the RMS value for the corresponding model
reached the target RMS error (set by the estimated data uncertainty). Having determined the

No. of data 5’338


No. iterations 4
Damping 100
Smoothing 4’000
Max. vp [m/s] 5’800
RMS initial model [ms] 74.4
RMS inversion result [ms] 11.4

Table 5-2: Inversion parameters for classical ray tomography.


44 Exploration-scale data example

optimal damping factor, the smoothing was then progressively reduced until a smooth model
consistent with the target RMS error was obtained. The high smoothing value was needed
to ensure a stable inversion. Employing lower values yielded a partly blocky model. The
maximum velocity value allowed in the inversion was capped at 5’800 m/s. In other words,
if the inversion update yielded a value larger than this then it was fixed at 5800 m/s. Such
hard constraints are important and preclude obtaining unrealistic (geologically implausible)
models.
Using the inversion parameters described above, the ray tomographic inversion yielded the
image displayed in colour-coded form in Figure 5-5a. The ray coverage associated with the
final inverted model is shown in Figure 5-5b. In Figure 5-5c the difference between the
inversion result (Figure 5-5a) and the initial model (Figure 5-4b) is depicted. Figure 5-5d
is a plot of the traveltime residual distribution for the initial and final models versus the
absolute source-receiver offset. It is clearly apparent that the traveltime residuals have been
significantly reduced as a result of the inversion.
We observe in the inverted subsurface P-wave velocity model of Figure 5-5a ’low’ velocity
zones on either side of the topographic high. These zones, centred on positions x = 12’000 m
and x = 2’000 m, exhibit velocity values of around 2’500 m/s (blue colours) close to the
ground surface and about 4’500 m/s (yellow colours) at depth. Both zones thin out towards
the topographic high. Below these zones, wavespeeds in excess 5’000 m/s and as high as
5’800 m/s (velocity cap) are visible. The lowest velocity values (< 2’500 m/s) are observed at
the crest of the topographic high (between x = 4’000 m and x = 6’000 m). The velocity values
of approximately < 3’700 m/s for z > 300 m and 4’400 m < x < 5’200 m are considerably lower
than the velocity values reached in the neighbouring high velocity areas (maximum values
of 5’800 m/s). The subsurface velocity distribution, and especially these ’anomalously’ low
velocities, will be further discussed in Section 5-4 where I offer a geological interpretation.
From the ray coverage shown in Figure 5-5b we see that only in the part of the profile
x > 10’000 m the model is less well constrained by the rays. In this region there are only
receivers and no shots (see Figure 5-1). The maximum depth of penetration for the rays is
approximately 1’500 m. But for most parts of the model the penetration depth is limited to
a depth range of 500 m to1’000 m
The change in velocity values from the initial model to the final inverted model is illustrated
in Figure 5-5c. The velocity residuals displayed are calculated as the difference between the
wavespeeds of the final result (Figure 5-5a) and the initial model (Figure 5-4b). Thus, negative
velocity residuals (red colours) indicate higher velocity values on the initial model whereas
positive residuals (blue colours) depict higher velocities on the final model. The maximum
velocity change is approximately 2’000 m/s, which is considerable. Generally speaking, the
areas having high velocity values (> 5’000 m/s) in Figure 5-5a show positive residuals, whereas
those areas having lower velocities (< 4’500 m/s) show negative traveltime residuals. Only
small parts of the final model are unchanged from the initial model (shown by the white
colour Figure 5-5c).
In addition to the RMS values given in Table 5-2, the actual distribution of traveltime residuals
is also presented in Figure 5-5d to provide a measure of the improved prediction of the observed
data through inversion (Schmelzbach et al., 2008). The traveltime residual is calculated as
the difference between the picked (observed) and the calculated (synthetic) traveltimes. The
large traveltime residuals of the initial model (shown by the light grey dots) at greater offsets
5-3 Traveltime tomography 45

(a) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

(b) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

2500 3000 3500 4000 4500 5000 5500


Velocity [m/s]
(c) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−2000 −1500 −1000 −500 0 500 1000 1500 2000


Velocity residual [m/s]
(d)
Traveltime residual [ms]

200

Initial model
−200
Inversion result

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Offset [m]

Figure 5-5: Classical ray tomography. (a) Result of ray tomographic inversion. The colour
coding of the velocity values is given in Figure 5-5b. (b) Ray coverage. (c) Difference
between the inversion result shown in Figure 5-5a and the initial model shown in
Figure 5-4b. (d) Traveltime residuals (picked - calculated traveltimes) versus offset
for the initial model (grey) and the inversion result (blue). The black lines mark the
mean traveltime residuals of the inversion result within 200 m offset bins. The red
line indicates zero residual.
46 Exploration-scale data example

are a consequence of the chosen initial model. From Figure 5-1 we see that no shots occur at
positions x > 10’000 m. Thus, only rays with relatively large source-receiver distance (offset)
will sample the model in those parts. Since the initial model has too high velocity values
for x > 10’000 m, the calculated traveltime will be too low. This results in large traveltime
residuals. The large RMS value of the initial model (74.4 ms) is the result of the wide scatter
of the traveltime residuals for the initial model. The blue dots in Figure 5-5d depict the
traveltime residuals of the final model. The tight concentration of the blue dots around the
zero residual (red) line explains the dramatic reduction of the RMS error to 11.4 ms. However,
in the near-offset range (offsets < 600 m) the traveltime residuals are skewed towards positive
values, indicating a slightly worse data fit for the near offsets compared to offsets > 600 m.

5-3-2 Fat ray tomography

Fat ray tomography was performed using frequencies of 75 Hz, 50 Hz and 25 Hz. In order
to choose appropriate frequencies for the fat ray tomography, a spectral analysis of the first
arriving phase was performed. The sample amplitude spectrum shown in Figure 5-6 was
calculated over 12 traces in a time window of 120 ms around the first breaks. Prior to the fast
Fourier Transform (FFT), tapering was done using a smooth cosine bell over the beginning
and ending 10% part of each trace. From the amplitude spectrum in Figure 5-6, we observe
that the central frequency is at 50 Hz. Thus the primary focus of the fat ray tomography was
directed at the inversion using 50 Hz.
Figures 5-7a, 5-7b and 5-7c display the tomograms resulting from fat ray tomography at
frequencies of 25 Hz, 50 Hz and 75 Hz, respectively. The corresponding velocity residual plot
and the traveltime residual distribution vs. offset plot for the 50 Hz fat ray tomographic
inversion are shown in Figures 5-8a and 5-8b, respectively. The same type plots for the other
two frequencies can be found in Appendix B.
Table 5-3 lists the parameters used for the fat ray tomographic inversions. The same number
of traveltime picks (5’338) were used as for the classical ray tomographic inversion. The
damping factor was set to a value of 100. Compared to the ray tomography, it was possible
to reduce the smoothing factor from 4000 to 3000. Again, smoothing in the vertical direction
was reduced by a factor 2 compared to the horizontal direction. The RMS error of the initial

Frequency [Hz]
50 100 150 200

-5

-10
dBPower

-15

-20

-25

-30

Figure 5-6: Amplitude spectrum of first arriving phase.


5-3 Traveltime tomography 47

75 Hz 50 Hz 25 Hz
No. of data 5’338 5’338 5’338
No. iterations 5 5 5
Damping 100 100 100
Smoothing 3’000 3’000 3’000
Max. vp [m/s] 5’800 5’800 5’800
RMS initial model [ms] 74.4 74.4 74.4
RMS inversion result [ms] 11.3 11.3 11.7

Table 5-3: Inversion parameters for fat ray tomography at 75 Hz, 50 Hz and 25 Hz.

model was 74.4 ms. This value was reduced to 11.3 ms during the inversion runs using 75 Hz
and 50 Hz fat ray tomography. For the tomographic inversion using 25 Hz fat rays, a somewhat
larger final RMS value of 11.7 ms was reached. In common with the standard ray tomography,
the maximum velocity allowed during the inversion was limited to 5’800 m/s.
The three tomograms (for 25 Hz, 50 Hz and 75 Hz) shown in Figure 5-7 look rather similar.
Moreover, the results of fat ray tomography closely resemble the velocity image obtained using
classical ray-based inversion (Figure 5-5a). All tomograms exhibit the same main features.
At the crest of the topographic high, all inversions yielded low velocity values (< 2’500 m/s)
down to z ≈ 300 m. Also the region between x = 4’400 m and x = 5’200 m having lower
velocity values compared to it’s surroundings is present on all tomograms. On either side
of the topographic high, a dipping layer of high velocity values (5’000 m/s to 5’800 m/s) is
apparent. Above this high velocity zone (indicated by the red colours), the velocity decreases
as we move upward from ∼ 4’500 m/s at the base to velocity values < 3’000 m/s at the ground
surface. The depth range covered by the fat rays increases with decreasing frequency. The
25 Hz fat rays reach a maximum depth of approximately 2’000 m below the ground surface at
x = 12’000 m (Figure 5-7a). The 50 Hz tomogram will be analysed in more detail in Section
5-4.
The velocity residual, i.e. the difference between the tomogram resulting from the 50 Hz fat
ray tomography (Figure 5-7b) and the initial model (Figure 5-4b), is displayed in Figure 5-
8a. Most parts of the actual velocity tomogram having velocity values > 5’000 m/s (i.e., red
colours on the image of Figure 5-7b) show positive velocity residuals in Figure 5-8a (blue
colour). A positive velocity residual means that the inversion increased the velocity from the
initial model value. The opposite case (a change towards lower velocity values) is indicated by
negative velocity residuals in Figure 5-8a. Most parts of the final tomogram having velocity
values smaller than 4’500 m/s show negative velocity residuals in Figure 5-8a. An exception
is the deepest part in the model constrained by fat rays. For x > 10’000 m and z > 2’000 m
the velocity values are > 5’000 m/s (see Figure 5-7b) but the residual is negative.
The traveltime residuals versus offset distribution shown in Figure 5-8b is very similar to that
for classical ray tomography given in Figure 5-5d. The widespread scatter of the initial model
(light grey dots) is drastically reduced towards the zero residual line (shown in red) as was
also observed for classical ray tomography in section 5-3-1. Again, the data fit for offsets
< 600 m is worse (traveltime residuals skewed to positive values) than for the larger offsets
(scattering around zero line).
48 Exploration-scale data example

(a) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

(b) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

0
(c)
1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

2500 3000 3500 4000 4500 5000 5500


Velocity [m/s]

Figure 5-7: Inversion results of fat ray tomography at frequencies of (a) 25 Hz, (b) 50 Hz, (c)
75 Hz.
5-3 Traveltime tomography 49

(a) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−2000 −1500 −1000 −500 0 500 1000 1500 2000


Velocity residual [m/s]
(b)
Traveltime residual [ms]

200

Initial model
−200
Inversion result

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Offset [m]

Figure 5-8: (a) Difference between the inversion result of the 50 Hz fat ray tomography shown
in Figure 5-7b and the initial model shown in Figure 5-4b. (b) Traveltime residuals
(picked - calculated traveltimes) versus offset for the initial model and the inversion
result. The black lines mark the mean traveltime residuals within 200m offset bins.
The red line indicates zero residual.
50 Exploration-scale data example

5-3-3 Checkerboard tests

For a meaningful interpretation of an inversion result the analysis of solution robustness and
model resolution are critical steps (Rawlinson and Sambridge, 2003). Synthetic reconstruc-
tions, such as the so-called checkerboard tests, are one way of analysing the reliability of the
solution. Checkerboard tests were performed for the results of the 50 Hz fat ray tomography
and the classical ray tomography. I used a checkerboard-testing scheme adapted from Zelt
(1998).
For the checkerboard test, the preferred final model of the tomographic inversion is selected.
The preferred final models for the 50 Hz fat ray tomography and the classical ray tomography
are shown in Figure 5-7b and Figure 5-5a, respectively. Similar to the synthetic tests (Chapter
4), sinusoidal velocity variations (anomalies) having maximum amplitudes of ±10% of the
velocity field in the preferred final model were computed and added to the preferred final
model. This yields a checkerboard model which is used to compute a synthetic data set.
Gaussian noise, having a standard deviation equal to the RMS error of the observed data
(i.e. 12 ms), was added to the synthetic data. Employing the preferred inversion result as the
starting model, the synthetic data are then inverted using the same free parameters (damping
and smoothing). The inversion of the synthetic data aims at recovering the checkerboard.
Figure 5-9a displays the checkerboard which was added to the final model of the 50 Hz fat
ray tomography. Because there is no visible difference between the checkerboards calculated
based on the two background fields given in Figure 5-7b and Figure 5-5a, only that for the
50 Hz fat ray tomography is shown. The wavelength of the sinusoidal anomalies is 6’000 m
in the x-direction and 1’200 m in the z-direction. In keeping with the synthetic data tests
described in Chapter 4, the recovered checkerboards are visualised as amplitude deviations (in
percentage terms) of the inversion result from the initial model. The amplitude deviation is
calculated according to Eq. (4-1). Here, the inversion result is the image obtained by inverting
the synthetic checkerboard data and the initial model is the preferred inversion result of the
classical ray or 50 Hz fat ray tomography. The recovered checkerboards using 50 Hz fat ray
and classical ray tomography are shown in Figures 5-9b and 5-9c, respectively. Since the
maximum amplitude of the velocity values in the true checkerboard is ±10% of the value of
the respective background fields, the optimum recovery of the true checkerboard would be
indicated by sinusoidal anomalies having maximum amplitudes of ±10% in Figures 5-9b and
5-9c.
Comparing the true checkerboard in Figure 5-9a against the reconstructed velocity fields of
Figures 5-9b and 5-9c, we observe that for x < 12’000 m polarities of the velocity undulations
are correct. However, the shapes of the anomalies are not fully recovered and some of the
anomalies smear into neighbouring fields. In both checkerboard tests the amplitude of the
negative anomaly (red) at the crest of the topographic high is overestimated (values of up to
−15%). Similarly, both inversions overestimate the maximum amplitude (+15%) of the pos-
itive anomaly at x < 3’000 m and 1’200 m < z < 1’800 m. For the positive anomaly between
x = 9’000 m and x = 12’000 m and for 1’800 m < z < 2’400 m, which is partly constrained
by the fat ray coverage, the amplitude is recovered quite well using fat ray tomography (Fig-
ure 5-9b). The amplitude for the same anomaly seems to be slightly underestimated in those
parts constrained by the ray tomography (Figure 5-9c).
To summarise, for x > 12’000 m the anomalies are not recovered correctly. This might be
a consequence of the limited coverage of fat rays and pencil (infinite frequency) rays (see
5-3 Traveltime tomography 51

0
(a)
1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−500 −400 −300 −200 −100 0 100 200 300 400 500
Velocity residual [m/s]

0
(b)
1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−15 −10 −5 0 5 10 15
%
0
(c)
1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−15 −10 −5 0 5 10 15
%

Figure 5-9: Checkerboard test for 50 Hz fat ray and classical ray tomography. (a) Checkerboard
added to preferred inversion result of the 50 Hz fat ray tomography (Figure 5-7b).
Checkerboard recovered (in percentage terms) using (b) 50 Hz fat ray tomography
and (c) classical ray tomography. The checkerboard would be perfectly recovered
for values of ±10%.
52 Exploration-scale data example

Figure 5-5b) in that area. For the rest to the profile (i.e., x < 12’000 m) both tomographic
inversions recovered the true polarity and the approximate size of the anomalies from the true
checkerboard. However, amplitudes and shapes are not everywhere correct.

5-3-4 Comparison between ray-based tomography and fat ray tomography

First of all, we observe that the tomograms obtained using the different tomographic inversions
are rather similar (see Figure 5-5a and Figure 5-7). The RMS errors reached for the final
iteration of the inversion all lie between 11.3 ms and 11.7 ms which is close to the target RMS
value of 12 ms (the rather conservative estimate of the average error in the data).
As already observed from the results of the synthetic data studies, the tomograms obtained
using fat ray tomography cover a larger depth range compared to those obtained using stan-
dard ray tomography. In Figure 5-10a the covered depth ranges of the fat ray tomographic
inversions at different frequencies are illustrated as contours which are superimposed on the
result of the ray tomographic inversion. The greatest depths are covered when 25 Hz fat ray
tomography is used (black line in Figure 5-10a). The depth ranges covered by the 50 Hz
and 75 Hz fat rays are depicted as grey and white contours, respectively. For example, at
x = 5000 m the maximum depth reached by pencil rays is at z ≈ 660 m where as the 25 Hz
fat rays reach a maximum depth of penetration at z ≈ 1270 m. However, the additionally
constrained parts of the model using fat ray tomography, only provide meaningful information
if these parts of the model are resolved.
As already seen in the previous section, for the well constrained part of the tomogram (x <
12’000 m) the gross features (i.e., amplitudes and size of the anomalies) of the checkerboard
were recovered in most parts of the model (see Figure 5-9b). This is especially true for the
positive anomaly between x = 9’000 m and x = 12’000 m and 1’800 m < z < 2’400 m, which
is is reasonably well recovered using 50 Hz fat ray tomography. Thus, in this area, the model
is constrained by the 50 Hz fat rays and the checkerboard test demonstrated that features are
resolved such that additional information on the subsurface model, which was not provided
by the classical ray tomography, is gained.
To further compare the classical ray against the fat ray tomography, the difference between
the tomograms obtained using classical ray and 50 Hz fat ray tomography is shown in Figure 5-
10b. Blue colours indicate larger velocity values on the classical ray tomogram, whereas red
colours indicate larger velocities on the 50 Hz fat ray tomogram. The two tomograms show
the least difference (white colours in Figure 5-10b) where wavespeeds are below 4’500 m/s (see
Figure 5-10a) for x < 3’000 m and x between 8’000 m and 11’000 m. The low velocities in the
region of the topographic high have slightly larger values on the tomogram obtained using
ray tomography (blue colours in Figure 5-10b). Except for x > 12’000 m, the colour coded
velocity difference shows features that are small compared to the resolution of the tomography
(assessed with the checkerboard test). However, there are no systematic differences such as
higher velocity values on one of the tomograms for the true higher velocity areas in the
subsurface model. Similarly, no systematic differences between the result of classical ray
tomography and the results of fat ray tomography at frequencies of 25 Hz and 75 Hz as well
as differences between the results of the different fat ray tomographic inversions were observed.
For completeness these difference plots are shown in Appendix B.
5-3 Traveltime tomography 53

(a) 0

1000
z [m]

2000 75 Hz fat ray tomography


50 Hz fat ray tomography
25 Hz fat ray tomography

14000 12000 10000 8000 6000 4000 2000 0


x [m]

2500 3000 3500 4000 4500 5000 5500


Velocity [m/s]

(b) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−600 −400 −200 0 200 400 600


Velocity residual [m/s]

Figure 5-10: Comparison between ray tomography and 50 Hz fat ray tomography. (a) Contours
indicating depth range of the model covered by fat rays superimposed on the result
of the ray tomographic inversion. Black: 25 Hz fat ray; Grey: 50 Hz fat ray; White:
75 Hz fat ray. (b) Difference between final model of classical ray and 50 Hz fat ray
tomographic inversions (Figure 5-5a and Figure 5-7b, respectively).
54 Exploration-scale data example

5-4 Interpretation

In this section I give a brief geological interpretation of the tomogram obtained using 50 Hz
fat ray tomography (Figure 5-11a). Auxiliary information for the interpretation includes a
geological cross-section supplied by OMV (Figure 5-11b) and interval velocities obtained using
vertical seismic profiling (VSP) in a borehole close to the seismic line (Figure 5-11c).

The VSP was performed in a well, located ∼ 1.5 km away from the seismic line, on the crest
of the topographic high. In order to compare the interval velocities against the tomogram
velocities, the location of the well was projected onto the seismic line. The vertical black
bar in Figure 5-11a indicates the projected well location on the profile and the depth covered
by the VSP experiment. The recording interval of the VSP occupies the depth range from
200 m to 2’000 m. The interval velocities (red line in Figure 5-11c) range from 2’000 m/s at
200 m depth to a maximum of 6’600 m/s at 1’500 m depth. Even for unfractured carbonates,
this maximum velocity value is rather high and is at the top end of wavespeeds for limestone
and dolomite given in the literature (e.g. 2’500 m/s − 6’500 m/s for dolomite given in Keary
et al. (2002); 2’000 m/s − 6’250 m/s for limestone given in Knödel et al. (2007)). Note that
velocity values towards the lower limits of these ranges correspond to wavespeeds in fractured
carbonates. Velocity values obtained using 50 Hz fat ray tomography are shown in blue, and
the green curve represents the velocity values of the initial model used in the inversion. At the
projected well location, the final model obtained from the tomographic inversion is constrained
to a depth of ∼ 700 m (deepest point of the blue curve in Figure 5-11c). Except for depths
above 200 m, the velocities on the tomogram are significantly smaller compared to the values
of the interval velocities. At a depth of ∼ 700 m, where the blue curve ends, the velocities
obtained from the tomographic inversion tend towards the values of the interval velocity
(∼ 4’500 m/s at 700 m). The occurrence of the rather low velocities on the tomogram below
the crest (teal colours for z > 300 m in Figure 5-11a) is not fully understood. Geologically, the
lower velocities could be a consequence of severe karstification and possible faulting within
the core of the anticline. However, the interval velocities indicate higher values (almost up
to 6’000 m/s) in this area. Thus, it is more likely that these low velocities are an artefact
from the tomographic inversion (e.g. compensation of the neighbouring high velocities) or
the result of systematic first break picking errors, such as consistently picking a wrong phase.

The cross-section shown in Figure 5-11b (provided by OMV) is the only geological information
available. The horizontal extension of the cross-section is unknown and the depth is only
approximately known (∼ 4’000 m). The cross-section is vertically exaggerated by a factor
of almost 2 and it is not of the same depth scale as the tomogram shown in Figure 5-11a.
However, looking at the geological cross-section, I interpret velocity values > 5’000 m/s (red
colours in Figure 5-11a) as folded carbonates (i.e. limestone and dolomite, see Figure 5-11b).
The low velocities at the crest of the anticline are an indication for strong karstification. The
shale layer (brown colour in Figure 5-11b) was eroded in the area of the topographic high, but
the traveltime tomography has recovered the shale layer as lower velocity values (< 4’500 m/s)
on either side of the anticline. Based on the given information it is not possible to state with
certainty whether the correct dip angle of the carbonate and shale layers has been recovered.
Furthermore, the tomographic inversion was not able to recover small scale features such as
the anhydrite layer (depicted in light red in Figure 5-11b) or features associated with the
small scale velocity fluctuations of the interval velocity in Figure 5-11c.
5-4 Interpretation 55

(a) SW NE
0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

2500 3000 3500 4000 4500 5000 5500


Velocity [m/s]

0
(b) (c)

SW NE
500
0

1000
z [m]

Depth [m]
1500

~ 4000 2000

Legend
Limestone Shale
Interval velocity (VSP) 2500
Dolomite Sandstone
Velocity from inversion
Anhydrite Claystone Initial model

3000
2000 3000 4000 5000 6000 7000
Velocity [m/s]

Figure 5-11: (a) Tomogram obtained using 50 Hz fat ray tomography. The black bar indicates
the depth of the well (used for the VSP) and it’s projected location on the profile.
The true well location is ∼ 1.5 km away from the seismic line. (b) Geological
cross-section over the anticline (provided by OMV). The depth information is only
approximately known, the length of the cross-section is unknown. (c) Interval
velocity from VSP (red). Blue: Velocity from the tomographic inversion at the
position indicated with the black bar in Figure 5-11a. Green: Initial model.
56 Exploration-scale data example

5-5 Summary

The application of fat ray tomography to this exploration-scale real data example worked well.
The tomograms were obtained by inverting 5’338 traveltime picks and RMS errors between
11.3 ms and 11.7 ms were reached on the final inversion. The target RMS value was assessed
using a traveltime reciprocity analysis which yielded a value of 12 ms. The results of the fat
ray tomographic inversions compare well to the result of the classical ray tomography.
For the classical ray tomography and the 50 Hz fat ray tomography a resolution assessment
based on checkerboard tests was performed. The polarities and the anomaly sizes of the
checkerboard were recovered by both tomographic inversions, but the shapes and amplitudes
of the anomalies were not everywhere correctly reconstructed. Compared to the classical
ray tomography, fat ray tomography covered a greater depth range. The checkerboard test
showed that most parts of the subsurface constrained by fat rays but not by classical pencil
rays were resolved in the context of the checkerboard test.
Based on a geological cross-section of the area, the velocities on the 50 Hz fat ray tomogram
were geologically interpreted. The high velocities on the tomogram were interpreted as a
folded carbonate layer forming an anticline. The low velocities at the crest of the anticline
are an indication of karstified carbonates. The cause of a low velocity zone in the core of
the anticline is unclear. Possible explanations include picking errors and hence an inversion
artefact, such as a compensation by the inversion for high velocities either side, or a real
geological feature. The interval velocities obtained from a nearby VSP experiment (∼ 1.5 km
offset from the profile) were compared against the tomogram velocities. Assuming that the
geology does not vary significantly between the profile and the well location, the VSP revealed
that the wavespeeds should be higher for the area on the tomogram, where the low velocity
zone is located. The lower velocities on either side of the anticline match a shale layer seen
on the geological cross-section. Small scale features are, as expected, not resolved by the
traveltime tomographic inversion.
Chapter 6

Small-scale high-resolution data


example

This chapter is concerned with the application of the fat ray tomography algorithm to a
second real data set example. Compared to the data set in Chapter 5, the distance scale
of the survey data examined here (investigation depth, line length and source and receiver
spacing) is considerably smaller by a factor of approximately 10. The data were acquired as
part of a multi-method geophysical field campaign that was aimed at imaging the internal
structure and depth of a Quaternary valley in northern Switzerland (Schmelzbach et al.,
2014).
First arrival traveltime data were inverted using the fat ray tomography approach at different
frequencies and the results were compared against an existing subsurface image obtained
using classical ray tomography (Reiser et al., 2014). Beside the initial model used in Reiser
et al. (2014), tests were carried out of using a second initial model which was derived using a
Herglotz-Wiechert inversion. One reason for applying fat ray tomography to this data set was
to test whether tomography using fat rays can more reliably delineate the contact between the
Quaternary sediments and the underlying bedrock. Classical ray tomography only vaguely
indicated the sediment to bedrock interface. A borehole in the survey area provides ground-
truth and is of value for interpreting the data. Again, checkerboard tests are used to assess
the resolution of the tomography results.

6-1 Traveltime tomography

The key acquisition parameters of the seismic survey are listed in Table 6-1. The seismic
profile in question is of length 830 m. Figure 6-1 shows the elevations of the geophone and
shot locations along the profile. The surface elevation varies from 505 m asl. in the northern
part of the profile to 470 m asl. in the south. In total, 78 shots (buried explosive charges
∼ 50 g) were fired at a nominal interval of 6 m. In total, 284 receivers at a spacing of 3 m
were employed.
58 Small-scale high-resolution data example

Due to a low S/N ratio, the identification of the first arrivals at far offsets was not possible
on all shot gathers. A total number of 7’196 first arrival travel times were picked from the
possible 22’152 traces. The maximum source-receiver offset of most traveltime picks is limited
to ∼ 250 m. Based on the picking accuracy and a traveltime reciprocity analysis, the data
uncertainty (error) was estimated to be ∼ 3 ms (Schmelzbach et al., 2014).

Profile length [m] 830


No. shot points 78
Source Dynamite (∼50 g)
Nominal source spacing [m] 6
No. receivers 284
Nominal receiver spacing [m] 3

Table 6-1: Acquisition parameters of the small-scale high-resolution data .

N S
510
Elevation [m asl.]

500 Receiver
490 Shot
480
470
800 700 600 500 400 300 200 100 0
x [m]

Figure 6-1: Source and receiver elevation along profile (vertically exaggerated by a factor 7).

As for the data in Chapter 5, the frequency content of the first arriving phase was assessed
through a spectral analysis. Figure 6-2 shows a sample amplitude spectrum (for frequencies
between 0 and 250 Hz) which was calculated over 10 traces in a time window of 120 ms. Ta-
pering of the rectangular window was done prior to FFT using a smooth cosine bell over the
beginning and ending 10% part of each trace. The spectrum peaks at 50 Hz and is of approx-
imate bandwidth (−30 dB points) 150 Hz. For the fat ray tomographic inversions applied to
this data set, I chose 25 Hz, 50 Hz and 100 Hz as input frequencies to the tomography.

Frequency [Hz]
50 100 150 200

-10
dBPower

-20
-30
-40
-50

Figure 6-2: Sample amplitude spectrum of first arriving phase (frequency range: 0 - 250 Hz).
6-1 Traveltime tomography 59

6-1-1 Variable velocity gradient initial model

Figure 6-3 displays the variable velocity gradient initial model used for the inversions (Reiser
et al., 2014). The 1D velocity depth profile depicted in Figure 6-3a was extracted at a
horizontal distance of x = 600 m from the 2D model in Figure 6-3b. Within the first 20 m
below the ground surface the velocity increases from 250 m/s to 1’500 m/s and stays at that
value over an interval of 30 m. Thereafter, the velocity increases to 3’700 m/s over the next
65 m depth interval. At a depth of 115 m below the ground surface, the velocity gradient
is gradually reduced to 7 s−1 . The velocity depth profile shown in Figure 6-3a reaches a
maximum velocity of 4’800 m/s at the base of the model.

(a) 0 (b) 0

50
100

z [m]
100
200
z [m]

150
300
200 800 600 400 200 0
x [m]
250

300
0 2000 4000 6000 1000 2000 3000 4000
Velocity [m/s] Velocity [m/s]

Figure 6-3: Initial velocity model used by Reiser et al. (2014). (a) 1D velocity function at
position x = 600 m. (b) 2D initial velocity model.

6-1-2 Results of ray-based and fat ray tomography at different frequencies using
the variable gradient initial model

The key inversion parameters used for the classic ray-based and fat ray tomography are listed
in Table 6-2. Reiser et al. (2014) used damping and smoothing factors of 200 for the classical
ray tomography. The RMS error of their initial model was 10 ms and the inversion reduced
this value within 20 iterations to a final RMS error of 3.8 ms. In order to reach an RMS
traveltime misfit close to the estimated data uncertainty of 3 ms, the damping and smoothing
factors were reduced to 100 for the 100 Hz and 50 Hz fat ray tomography, yielding RMS errors
of the final inversion runs of 3.6 ms and 3.9 ms, respectively. For the 25 Hz fat ray tomography
the smoothing was further reduced to a value of 50 yielding a final RMS traveltime misfit of
4.2 ms.
The final models obtained using the different tomographic inversions are shown in Figure 6-4.
As a consequence of the limited number picks in the far-offset range (offsets > 250 m), the
model is considered to be reliable (constrained by the deepest penetrating rays) to only shallow
depth along the first ∼ 300 m of the profile. For ray tomography, this means that only the
upper ∼ 20 m of the subsurface model is constrained in this part of the profile. For the fat ray
tomographic inversions at 100 Hz , 50 Hz and 25 Hz the model is constrained in the same area
down to depths of ∼ 40 m, ∼ 50 m and ∼ 60 m, respectively. On all tomograms we observe
similar features. At shallow depths (∼ 10 m below the ground surface), lenses of localised
high velocities (values between 2’500 ms and 3’500 ms) are apparent in all tomograms. A
60 Small-scale high-resolution data example

(a) 0 (b) 0

100 100
z [m]

z [m]
200 200

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

1000 2000 3000 4000 5000 6000 1000 2000 3000 4000 5000 6000
Velocity [m/s] Velocity [m/s]

(c) 0 (d) 0

100 100

z [m]
z [m]

200 200

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

1000 2000 3000 4000 5000 6000 1000 2000 3000 4000 5000 6000
Velocity [m/s] Velocity [m/s]

Figure 6-4: Inversion results using the velocity field in Figure 6-3b as initial model. (a) Classical
ray-based tomography. (b) 100 Hz fat ray tomography. (c) 50 Hz fat ray tomogra-
phy. (d) 25 Hz fat ray tomography.

Ray Fat ray 100 Hz Fat ray 50 Hz Fat ray 25 Hz


No. of data 7’196 7’196 7’196 7’196
No. iterations 20 6 5 5
Damping 200 100 100 100
Smoothing 200 100 100 50
RMS initial model [ms] 10 10 10 10
RMS inversion result [ms] 3.8 3.6 3.9 4.2

Table 6-2: Inversion parameters (variable gradient initial model).

elongated localised high velocity zone is located at shallow depth between 180 m and 380 m.
Another distinct zone of high velocities at shallow depths is centred around x = 590 m. The
most eye-catching feature of the plots is the high velocity zone centred at 580 m at a depth of
∼ 150 m. On the tomogram obtained using classical ray tomography (Figure 6-4a) maximum
velocity values of ∼ 5’000 m/s are reached in this high velocity zone. Velocities are even higher
on the tomograms obtained using fat ray tomography. A maximum velocity of 6’000 m/s is
obtained using 25 Hz fat ray tomography.

The differences between the final and the initial models for both classical ray tomography
and 50 Hz fat ray tomography are shown in Figures 6-5a and 6-5b, respectively. Compared
to the initial model, the velocity values in the high velocity zones described above are larger
on the tomograms. Thus, positive velocity residuals (blue colours) are observed in Figures
6-1 Traveltime tomography 61

6-5a and 6-5b) for these high velocity areas. The maximum positive velocity residuals are
1’200 m/s Figure 6-5a and 1’800 m/s in Figure 6-5b. We also observe that large parts of the
two subsurface models are unchanged by the inversions (white colours).

Figure 6-5c shows the distribution of the traveltime residuals (i.e., difference between picked
and calculated traveltimes) for the initial model (light grey) and the results of the classical
(blue) and 50 Hz fat ray (green) tomography. The wide scatter of the traveltime residuals
calculated for the initial model (RMS value of 10 ms) is reduced to a similar amount (∼ 4 ms)
by both tomographic inversions (blue and green dots concentrated around the red zero residual
line). Looking at the density of the green and blue dots in Figure 6-5c, the limited number
of traveltime picks for offsets > 250 m becomes apparent. For offsets < 25 m, the traveltime
residuals for both inversions are biased towards positive values, indicating a worse data fit
compared to greater offsets. Differences between the traveltime residual distributions for the
classical ray and the 50 Hz fat ray tomography are very subtle. Compared to ray tomography,
the data fit obtained using 50 Hz fat ray tomography is slightly worse for offsets > 250 m.

The rather high wavespeeds obtained with fat ray tomography (see Figures 6-4b, 6-4c and
6-4d) was the motivation to test a different initial model, which will be described in the next
section.

(a) 0 0
(b)
100 100
z [m]

z [m]

200 200

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

−1000 −500 0 500 1000 −1500 −1000 −500 0 500 1000 1500
Velocity residual [m/s] Velocity residual [m/s]

(c)
30
Traveltime residual [ms]

20
10
0
−10
Initial model
−20 Ray tomography
−30 50Hz fat ray tomography

0 50 100 150 200 250 300 350 400


Offset [m]

Figure 6-5: Velocity residuals (inversion result - initial model) for (a) classical ray tomogra-
phy and (b) 50 Hz fat ray tomography. (c) Traveltime residual (picked - calculated
traveltimes) shown for the initial model (light grey), final model from classical ray to-
mography (blue) and final model from 50 Hz ray tomography. The red line indicates
zero residual.
62 Small-scale high-resolution data example

6-1-3 Herglotz-Wiechert initial model

It is possible to derive the velocity versus depth curve directly from the observed traveltime
data when the Earth conforms to a 1D velocity distribution using the famous Herglotz-
Wiechert (H-W) inversion formula, first published in the early 1900s. The H-W method is
described in detail in most textbooks on theoretical seismology (e.g Aki and Richards, 2002).
In Appendix D I summarise the essential aspects. A numerical implementation of the integra-
tion formula, based on mathematical spline function fitting of the averaged traveltime versus
offset (T − X) data, was used to derive an initial velocity model for subsequent tomography.
The deduced velocity-depth profile is displayed in Figure 6-6a. It was calculated down to a
depth z = 210 m. The raw T − X data from the first arrivals for all shots and the corre-
sponding slowness versus distance plot for the best fit curve to the traveltimes are given in
Appendix D.
The velocity-depth values obtained using the H-W method show an approximately linear
velocity function (constant gradient) from the surface (z = 25 m) to z = 210 m (see Figure 6-
6a). Within this depth interval the velocity values increase from 1’000 m/s to 4’200 m/s. In
order to extend the modelling domain beyond the maximum depth obtained from the H-W
method, velocity values were extrapolated with a very low gradient to a maximum depth of
z = 300 m.
The 2D initial model in Figure 6-6b is the result of applying the 1D velocity-depth profile
obtained from the H-W inversion at all points along the entire profile, but taking the initial
velocity at the elevation of the surface topography.

(a) 0 (b) 0

50
100

z [m]
100
200
z [m]

150

200 800 600 400 200 0


x [m]
250

300
0 2000 4000 6000 1000 1500 2000 2500 3000 3500 4000 4500
Velocity [m/s] Velocity [m/s]

Figure 6-6: Initial velocity model calculated using the Herglotz-Wiechert (H-W) method. (a) 1D
velocity depth profile. Note the maximum depth obtained from the H-W inversion
is at z = 210 m. (b) 2D H-W derived initial velocity model.

6-1-4 Results of ray-based and fat ray tomography using the Herglotz-Wiechert
initial model

Figures 6-7a and 6-7b show the final models which were obtained using classical ray tomog-
raphy and 100 Hz fat ray tomography and employing the H-W starting model (Figure 6-6b).
The fat ray tomographic inversion using 50 Hz and 100 Hz fat rays yielded RMS values of the
final models of 4.3 ms and 3.9 ms, respectively. Here, the preferred result of the 100 Hz fat
6-1 Traveltime tomography 63

ray tomography is shown. The tomogram obtained using 50 Hz fat ray tomography is shown
in Appendix C.
The key inversion parameters are listed in Table 6-3. The initial RMS error of 10.2 ms was
reduced by both tomographic inversions to the same final RMS value of 3.9 ms (the RMS
values of the final models using ray and 100 Hz fat ray tomography and employing the variable
gradient initial model of Reiser et al. (2014) were 3.8 ms and 3.6 ms, respectively). Smoothing
factors of 100 and 25 were used for the ray tomography and the 100 Hz fat ray tomography,
respectively. Additionally, smoothing in the vertical direction was reduced by a factor of 2.
For both inversions, the damping factor was set to 100 and the maximum wavespeed allowed
during the inversion was capped at 5’000 m/s. The hard constraint of velocity limitation
was employed to keep the estimated velocity values within the expected velocity range for
the bedrock lithology and velocities observed in a borehole check-shot experiment along the
profile (see Section 6-3).

Ray Fat ray 100 Hz


No. of data 7’196 7’196
No. iterations 8 11
Damping 100 100
Smoothing 100 25
Max. vp [m/s] 5000 5000
RMS initial model [ms] 10.2 10.2
RMS inversion result [ms] 3.9 3.9

Table 6-3: Inversion parameters (Herglotz-Wiechert initial model).

(a) 0 (b) 0

100 100
z [m]

z [m]

200 200

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

1000 2000 3000 4000 5000 1000 2000 3000 4000 5000
Velocity [m/s] Velocity [m/s]

Figure 6-7: Inversion results using the Herglotz-Wiechert initial model. (a) Classical ray tomog-
raphy. (b) 100 Hz fat ray tomography.

Overall, the main features already observed on the images in Figure 6-4, viz., zones of localised
high velocity at shallow depth (∼ 10 m below the surface) and the deeper high velocity zone,
are similarly indicated by the tomographic inversions using the alternative H-W initial model
shown Figure 6-7. For the classical ray-based inversion the localised shallow high velocity
lens centred at x = 200 m is only very vaguely indicated on the tomogram in Figure 6-7a
due to the limited ray coverage. Similarly, only a few rays penetrate deeper than z = 150 m
where high velocity values are observed. In contrast, the deep high velocity zone centred at
x = 600 m is a prominent feature on the tomogram obtained using 100 Hz fat ray tomography
64 Small-scale high-resolution data example

(Figure 6-7b). Note the effect of limiting the maximum velocity allowed during the inversion
to 5’000 m/s.

The plots of the difference between the inversion results and the initial model (Figure 6-6b) are
shown in Figures 6-8a (ray tomography) and 6-8b (100 Hz fat ray tomography). Again, similar
observations as for the case where the variable gradient initial model of Reiser et al. (2014)
was used (Section 6-1-2) can be made. Namely, in areas identified as high velocity zones on
the tomograms (Figure 6-7) the velocity residual is positive (blue colour). Maximum positive
velocity residuals range up to 1’600 m/s and 2’000 m/s for the ray tomographic inversion
(Figure 6-8a) and the 100 Hz fat ray tomographic inversion (Figure 6-8b), respectively. For
the latter, the largest velocity difference is observed for the shallow elongated high velocity
zone centred at x = 200 m. In this area the velocities on the tomogram (Figure 6-7b) reach
values as high as 3’000 m/s. The traveltime residual versus offset distribution is shown in
Figure 6-8c. The data fit for offsets < 50 m was not improved using the Herglotz-Wiechert
initial model (compare with Figure 6-5c). The traveltime residuals calculated for both final
models (green and blue dots) are skewed towards positive residuals in the near-offset range.
For larger offsets, the wide scatter of the initial model (light grey dots) was reduced by the
inversions (concentration of blue and green dots around zero residual line). The results of the
different tomographic inversions are further discussed and compared against interval velocities
obtained from a check-shot experiment in Section 6-3.

(a) 0 (b) 0

100 100
z [m]

z [m]

200 200

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

−1500 −1000 −500 0 500 1000 1500 −1500−1000 −500 0 500 1000 1500
Velocity residual [m/s] Velocity residual [m/s]
(c)
30 Initial model
Traveltime residual [ms]

20 Ray tomography
100Hz fat ray tomography
10
0
−10
−20
−30
0 50 100 150 200 250 300 350 400
Offset [m]

Figure 6-8: Velocity residuals (the inversion result minus the H-W initial model) for (a) classical
ray tomography and (b) 100 Hz fat ray tomography. (c) Traveltime residual (picked
minus calculated traveltimes) shown for the initial model (light grey), final model
from classical ray tomography (blue) and final model from 50 Hz ray tomography.
The red line indicates zero residual.
6-2 Resolution assessment: Checkerboard tests and column sum of Jacobian
matrix 65

6-2 Resolution assessment: Checkerboard tests and column sum


of Jacobian matrix

Checkerboard tests, following the same scheme as described in Section 5-3-3, were performed
for the results of the classical ray tomography (Figure 6-4a) and the 50 Hz fat ray tomography
(Figure 6-4c) using the variable gradient velocity field as initial model, as well as for the result
of the 100 Hz fat ray tomography (Figure 6-7b) using the H-W initial model. In addition to
the checkerboard tests, a second resolution analysis based on the column sum of the Jacobian
matrix is presented.

6-2-1 Checkerboard tests

Figure 6-9a displays the checkerboard added to the preferred final model of the 50 Hz fat ray
tomography (Figure 6-4c). The checkerboards added to the final models of the classical ray
and 100 Hz fat ray tomographic inversion look very similar to the pattern in Figure 6-9a and
so are not shown here. The sinusoidal velocity undulations (anomalies) of the checkerboards
have maximum amplitudes of ±10% of the background velocity field and the wavelengths in
the x-direction and z-direction are 400 m and 80 m, respectively.

(a) 0 (b) 0

100 100
z [m]

z [m]

200 200

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

−400 −200 0 200 400 −20 −10 0 10 20


Velocity residual [m/s] %

(c) 0 (d) 0

100 100
z [m]

z [m]

200 200

800 600 400 200 0 800 600 400 200 0


x [m] x [m]

−15 −10 −5 0 5 10 15 −20 −10 0 10 20


% %

Figure 6-9: Checkerboard tests. (a) Checkerboard added to the preferred result of the 50 Hz
fat ray tomography (Figure 6-4c). Checkerboard recovered (in percentage terms)
using (b) classical ray tomography (variable gradient initial model), (c) 100 Hz fat
ray tomography (H-W initial model) and (d) 50 Hz fat ray tomography (variable
gradient initial model).
66 Small-scale high-resolution data example

The recovered checkerboard is illustrated as amplitude deviation (in percentage terms, cal-
culated according to Eq. (4-1)) of the inversion result from the initial model. Here, the
inversion result is the image obtained by inverting the synthetic checkerboard data and the
initial model is the preferred final model of the respective tomographic inversion (see also Sec-
tion 5-3-3). The recovered checkerboards for the classical ray tomography and the 50 Hz fat
ray tomography using the variable gradient initial model are shown in Figures 6-9b and 6-9d.
The recovered checkerboard for the 100 Hz fat ray tomography using the Herglotz-Wiechert
initial model is shown in Figure 6-9c. On all images in Figures 6-9b through 6-9d we observe
that for the shallowest velocity undulations of the checkerboard the polarities are correctly
recovered. However, shapes and amplitudes are not very well recovered. The ray tomography
(Figure 6-9b) and the 50 Hz fat ray tomography indicate the true polarities of the anomalies
at 120 m < z < 160 m and x values between 400 m and 800 m. For the anomaly at 600 m
< x < 800 m and 40 m < z < 80 m the 100 Hz fat ray tomography indicates the true polarity
(Figure 6-9c). However, large areas of all the checkerboards are not recovered.

6-2-2 Column sum of Jacobian matrix

The rather poorly reconstructed checkerboards motivated me to consider a further resolution


measure, the column sum of the Jacobian matrix which contains the sensitivities. Ideally,
one could use the formal model resolution matrix (see Menke, 1989) to assess the goodness
of a final inversion result, but its computation is expensive, involving the inversion of a large
pseudo-Hessian matrix. As shown by Meles et al. (2012), in the context of GPR inversion,
a useful and economical proxy is the column sum of absolute values of the Jacobian matrix.
The sensitivities associated with the jth inversion cell in the model are contained in the jth
column of the Jacobian matrix. The summation of the sensitivities in the jth column gives
an estimate of the overall sensitivity associated with the jth inversion cell. To a large degree,
the column sum is influenced by the number of fat rays covering the specific inversion cell
and the velocity value within that cell. The column sum can be interpreted as an indication
for the fat ray coverage.

Plotting each column sum as colour coded pixel at the position of the respective inversion cell
in the subsurface model yields the images displayed in Figures 6-10a (50 Hz fat ray tomogra-
phy) and 6-10b (100 Hz fat ray tomography). The largest column sums are located at shallow
depths between 5 m and ∼ 40 m below the ground surface. In those parts the anomalies of
the checkerboard were better recovered than in most other parts of the subsurface model (see
Figures 6-9c and 6-9d). For large parts of the models, the column sums are < 400.

The images in Figures 6-10c and 6-10d are visualisations of the column sum interpreted as
fat ray coverage on the tomograms in Figures 6-4c (50 Hz fat ray tomography) and 6-7b
(100 Hz fat ray tomography). The parts of the subsurface model having column sums > 400
are completely visible. For column sums < 400, the image is progressively faded out with
decreasing column sum. Based on the (somewhat arbitrary) column sum threshold of 400,
only the shallow parts of the models (down to ∼ 40 m below the surface) and to some extent
the area of the deeper high velocity zone are adequately constrained by the fat rays.
6-3 Interpretation 67

(a) (b)

(c) (d)

Figure 6-10: Column sum of Jacobian matrix for (a) 50 Hz fat ray tomography (variable gradient
starting model) and (b) 100 Hz fat ray tomography (H-W starting model). (c)
Faded-out 50 Hz fat ray tomogram (Figure 6-4c; variable gradient starting model)
(d) Faded-out 100 Hz fat ray tomogram (Figure 6-7b; H-W starting model)

6-3 Interpretation

The geological interpretation follows the interpretation given in Schmelzbach et al. (2014).
In Figure 6-11a the tomogram obtained using 100 Hz fat ray tomography and the Herglotz-
Wiechert initial model is shown again. Figure 6-11b displays the tomogram obtained using
classical ray tomography and the variable gradient initial model. The black bar indicates
the location and depth of a borehole located along the profile. This borehole was used for
a check-shot seismic experiment and a geological analysis of the drilling cuttings. Note that
the geological interpretation is predominantly based on the information obtained from the
borehole (provided by Nagra, the Swiss National Cooperative for the Disposal of Radioactive
Waste). Due to the fact that the model is only weakly constrained for depths > 50 m below
the subsurface (see Section 6-2-1 and Section 6-2-2), any interpretation based on any of
the tomograms is problematic and should be made with care. In Figure 6-11c the interval
velocities derived from the check-shot survey (red line) are shown together with the velocity-
depth profiles obtained from different tomographic inversions at the location of the borehole.
The left panel in Figure 6-11c displays the velocities from the ray tomographic inversion (black
curve) and the 50 Hz fat ray tomography (blue curve) using the variable gradient initial model
(green curve). The right panel shows the velocities from the ray tomography (black curve)
and the 100 Hz fat ray tomography (blue curve) using the H-W initial model (green curve).
Between the two panels the interpreted stratigraphic column is shown (from Schmelzbach
et al., 2014).
68 Small-scale high-resolution data example

(a) 500 6000

450
Elevation [m asl.]

400 5000
350

300
4000
250

Velocity [m/s]
800 700 600 500 400 300 200 100 0
x [m] 3000
(b) 500

450
Elevation [m asl.]

400 2000

350

300
1000
250

800 700 600 500 400 300 200 100 0


x [m]
(c) 500 500

480 480
Moraine
460 460

440 440
Fluvial
Elevation [m asl.]

Elevation [m asl.]
420 Sediments 420

400 400

380
Lacustrine 380
Sediments
360 360

340 340
Limestone
320 320

0 2000 4000 6000 0 2000 4000 6000


Velocity [m/s] Velocity [m/s]
50 Hz fat ray tomography 100 Hz fat ray tomography
Ray tomography Ray tomography
Interval velocity Interval velocity
Initial model Initial model

Figure 6-11: (a) Tomogram obtained using 100 Hz fat ray tomography (Herglotz-Wiechert initial
model). (b) Tomogram obtained using ray tomography (from Reiser et al., 2014).
The black bar indicates the location and depth of a borehole. (c) Interval velocity
from check-shot experiment in the borehole and velocity-depth profile extracted
from tomograms obtained using the variable gradient initial model (left panel),
and the H-W initial model (right panel). The middle panel displays an interpreted
stratigraphic column (from Schmelzbach et al., 2014).
6-3 Interpretation 69

Within the first ∼ 25 m below the surface (from elevation 485 m asl. to 460 m asl.) the
velocities increase from 500 m/s to 2’000 m/s (see Figure 6-11c). Based on the information
from the borehole, this interval is interpreted as glacial moraine deposits (variable amounts
of sand, silt and gravel).

Below an elevation of 460 m asl., and over a depth interval of roughly 5 m, the interval ve-
locity plot shows localised high values (∼ 3’000 m/s, red curve). All tomographic inversions
indicate the localised high velocity values. However, compared to the values of the interval
velocities, the velocities obtained from ray tomography (black curves) are lower (∼ 2’000 m/s
at 460 m asl.). The 50 Hz and the 100 Hz fat ray tomographic inversions yield velocity values
of ∼ 2’200 m/s and ∼ 2’800 m/s, respectively. However, compared to the interval velocities,
the localised high velocities are indicated at somewhat greater depth (∼ 450 m asl.).

The sediments in the interval from 460 m asl. to 400 m asl. are interpreted to be of fluvial
origin. On average, the interval velocities are slightly below 2’000 m/s, but fluctuations are
observed (red curve in Figure 6-11c). While drilling through the fluvial deposits (mainly
gravel with variable amounts of sand), many boulders were encountered, which are probably
the cause of the velocity fluctuations. From 440 m asl. to 400 m asl. the check-shot velocity-
depth profiles shows an interval with decreased velocities compared to the overlying section. In
this region the velocity-depth profiles extracted from the different tomograms follow the initial
models (green curves). This might be an additional indication that the subsurface model is
not well constrained at these depths, which was already evident from the checkerboard tests
(inadequate recovery of the checkerboard, see Figure 6-9) and from the analysis of the column
sum of the Jacobian matrix (see Figure 6-10).

A zone of lacustrine sediments (fine, silty sand and gravel) is indicated by the relatively
constant interval velocities of ∼ 2’200 m/s between 360 m asl. to 400 m asl.. For this depth
interval, the velocities from the tomographic inversions using the variable gradient initial
model (black and blue curves in the left panel in Figure 6-11c) differ significantly from the
interval velocities (red). The difference between the interval velocity and the velocities from
the tomographic inversions is smaller in the case of the Herglotz-Wiechert initial model (green
curve in right panel in Figure 6-11c). However, this is due to the fact that the H-W initial
model is closer to the interval velocities and the velocities from the tomographic inversions
follow the initial values, especially at depth where there is low sensitivity and hence inability
to update values during the inversion.

At the contact between the valley infill and the limestone bedrock (360 m asl.), the interval
velocity in the check-shot log jumps from ∼ 2’200 m/s to ∼ 3’900 m/s. Also the velocity-
depth profiles extracted from the different tomograms show an increase in wavespeed around
the depth of the contact. However, both fat ray tomographic inversions overestimate the
limestone velocity. In Figure 6-11c the velocities from the 50 Hz fat ray tomography (blue
curve, left panel) peaks at ∼ 5’800 m/s at 360 m asl. and the velocities from the 100 Hz fat
ray tomography (blue curve, right panel) reach their (capped) maximum velocity value at
350 m asl.. The ray tomographic inversion using the Herglotz-Wiechert initial model (black
curve, right panel) is closer to the interval velocity at that depth and peaks slightly above
360 m asl. at ∼ 3’900 m/s. The maximum velocity value of the ray tomographic inversion
shown in the left panel in Figure 6-11c (black) is reached at approximately 10 m above the
bedrock contact.
70 Small-scale high-resolution data example

6-4 Summary

In this chapter the application of fat ray tomography to a small-scale seismic experiment was
demonstrated. Compared to the exploration-scale survey presented in Chapter 5 the spatial
extent of the small-scale experiment is smaller by a factor of ∼ 10.
For the fat ray tomographic inversions, two different initial models were tested. First the
variable gradient initial model used in Reiser et al. (2014) was employed; second, a starting
model derived using the Herglotz-Wiechert method was used. Fat ray tomography using the
variable gradient initial model yielded unrealistic high velocity values in the deeper parts of
the tomograms and, thus, the maximum velocity allowed during the inversion was limited
for the tomographic inversions using the H-W initial model. Generally speaking, fat ray
tomography at 100 Hz yielded the most plausible results (also having the RMS values closest
to the target RMS error estimated to be 3 ms) but the difference to the fat ray tomography
at 50 Hz was not large in the case of the variable gradient initial model.
Checkerboard tests were performed for the results of ray tomography and 50 Hz fat ray tomog-
raphy using the variable gradient initial model, as well as for the 100 Hz fat ray tomography
using the Herglotz-Wiechert initial model. From all checkerboard tests it became evident that
only the shallowest ∼ 40 m in the tomograms are (partly) resolved. For these resolved parts
of the tomogram the largest column sums of the Jacobian matrix were observed, providing
confirmation that the image is most reliable in the shallower parts of the section but not at
depth.
The geological interpretation followed the interpretation given in Schmelzbach et al. (2014)
and was based on the information obtained from a borehole located on the profile (interval
velocities and the analysis of drilling cuttings). A selection of 1D velocity-depth profiles
extracted from different tomograms at the position of the borehole were compared against
the interval velocities from the check-shot experiment. A shallow high velocity excursion
of the interval velocities as well as the contact between the sediments and the limestone
bedrock are indicated by all tomographic inversions. The fat ray tomographic inversions
overestimate the limestone velocities and the shallow high velocity zone is indicated at larger
depths compared to the interval velocity. However, I personally prefer the tomogram obtained
using 100 Hz fat ray tomography and the H-W initial model because in my opinion, the high
velocities indicated by the check-shot log at 460 m asl. are better estimated compared to other
tomographic inversions.
Chapter 7

Conclusions and Outlook

In this thesis I developed and applied a fat ray (Fresnel volume) tomography algorithm for
improved imaging of the subsurface using conventional surface-recorded seismic data (first-
break times). The imaging capabilities of the fat ray tomography algorithm were investigated
on three synthetic (two small scale and one large scale model) and two real data (small scale
and large/exploration scale) examples, and were compared to standard ray-based tomography.

7-1 Conclusions

In general, using fat ray tomography it was possible to gain modest, additional informa-
tion about the subsurface model compared to classical ray tomography. This was shown by
the small scale and large scale synthetic checkerboard model inversions as well as for the
exploration-scale field data set, where a checkerboard test indicated resolved velocity anoma-
lies which were not reconstructed using classical ray tomography. However, for a second
field example at a much smaller scale, this was not the case. The inverted model is poorly
constrained by inadequate ray coverage in its deeper parts and so it was not possible to
demonstrate superior performance of the fat ray algorithm.
For the checkerboard synthetic data example using the small scale model (Model B; source
wavelet with 50 Hz dominant frequency) the highest frequency, i.e., 100 Hz, used for the fat
ray tomography performed best in recovering the velocity fluctuations. On the other hand,
for the large scale synthetic model (Model C; source wavelet with 50 Hz dominant frequency),
the fat ray tomography operating at the lowest test frequency (25Hz) performed best. This
leads me to the conclusion that, within the limitations of the available frequency band of
the seismic survey, the choice of the frequency for the fat ray tomography should not only
be based on the frequency characteristics of the data but should also be partly based on the
distance/depth scale of the experiment.
With respect to the usual regularisation applied to stabilise and constrain tomographic inver-
sions, two observations can be made. First, compared to the classical ray tomography, it was
generally possible to reduce the smoothing for the fat ray tomographic inversions. Second, in
72 Conclusions and Outlook

the presence of a discontinuous jump in velocity (such as the high velocity step in synthetic
Model A), the fat ray tomography was more prone to overshooting artefacts than the classical
ray tomography. Therefore, the maximum velocity value allowed during the inversion had to
be restricted to some maximum value for the fat ray tomography but not for the classical
ray tomography. This was also the case for the small-scale field data set, where the velocity
increases abruptly at the sediment - bedrock interface.
Using fat ray tomography, it was hoped to achieve better imaging of low velocity structures
which are notoriously difficult to image with conventional traveltime tomography because the
infinitesimal thin first-arrival rays by-pass such structures. By contrast, we expected Fresnel
volume rays to be influenced by nearby low velocity zones. However, it was not possible
to adequately reconstruct the low velocity block insert of synthetic Model A with fat ray
tomography, while an equal-size high-velocity anomaly was adequately resolved.
In the case of the small scale field data set, checkerboard tests and a companion pseudo-
resolution/sensitivity analysis of the inverted model based on the column sum of the Jacobian
matrix showed that the subsurface model is only very weakly constrained by rays or fat rays
at depths in excess of 50 m below the surface. A major reason for this is the limited amount of
far-offset travel time picks (due to low signal-to-noise ratio) and thus, only a relatively small
number of deeply penetrating (fat) rays. As a consequence, the imaging capabilities of both
classical ray and fat ray tomography are very limited for depths > 50 m. Any interpretation
of deeper geological structures based on the tomograms is problematic and should be made
with care.

7-2 Outlook

Further synthetic modelling studies are required to:

• Investigate the relationship between frequency of the fat ray tomography, the corre-
sponding wavelength and scale of the survey and the size of subsurface target
• Examine other models incorporating low velocity blocks of differing, size, depth and
contrast

In addition, future work could profitably be spent researching other more technical aspects of
fat ray tomography. One major drawback of fat ray tomography is that it is computationally
more demanding than classical ray tomography due to the altered sensitivity calculation.
Further improvements of the algorithm and possible programming in a low level computing
language (e.g. C or Fortran) could significantly reduce the computational cost of fat ray
tomography.
Throughout this thesis project, a linear weighting scheme was applied across the Fresnel vol-
ume for calculating the sensitivities. It might be worthwhile testing other weighting functions
such as a bell-shaped (Gaussian) profile. Moreover, a comparison of the fat ray approach
used in this thesis to other frequency-dependent traveltime tomography approaches would
be interesting. In particular a comparison with the more sophisticated (and computationally
more expensive) approach of calculating the sensitivity kernels using full forward modelling
of the wavefield (e.g. Liu et al., 2009) would be revealing.
7-2 Outlook 73

The occurrence and cause of the low velocity zone observed in the core of the anticline on
the tomograms of the exploration-scale field data is not fully understood. A further review
of the first break picks and the inclusion of the VSP data in the tomographic inversion could
be beneficial and lead to a better understanding of this feature.
Finally, the Austrian company OMV, who provided the exploration-scale field data set, have
shown interest in using the tomograms produced within this thesis, for a recalculation and
re-assessment of the static-corrections to be applied in processing of the seismic reflection
data.
74 Conclusions and Outlook
Bibliography

Aki, K. and Richards, P. (2002). Quantitative Seismology. University Science Books.

Bai, C.-Y., Li, X.-W., Huang, G.-J., and Greenhalgh, S. (2013). Simultaneous inversion for
velocity and reflector geometry using multi-phase Fresnel volume rays. Pure and Applied
Geophysics, doi: 10.1007/s00024-013-0686-6.

Cervenỳ, V. and Soares, J. E. P. (1992). Fresnel volume ray tracing. Geophysics, 57(7):902–
915.

Chen, J., Jaiswal, P., and Zelt, C. A. (2013). A case history: Application of frequency-
dependent traveltime tomography and full waveform inversion to a known near-surface
target. In SEG Technical Program Expanded Abstracts 2013, chapter 340, pages 1743–
1748. Society of Exploration Geophysicists.

Constable, S. C., Parker, R. L., and Constable, C. G. (1987). Occam’s inversion: A practical
algorithm for generating smooth models from electromagnetic sounding data. Geophysics,
52(3):289–300.

Gance, J., Grandjean, G., Samyn, K., and Malet, J.-P. (2012). Quasi-Newton inversion of
seismic first arrivals using source finite bandwidth assumption: Application to subsurface
characterization of landslides. Journal of Applied Geophysics, 87:94–106.

Greenhalgh, S. and King, D. (1981). Curved raypath interpretation of seismic refraction data.
Geophysical Prospecting, 29(6):853–882.

Husen, S. and Kissling, E. (2001). Local earthquake tomography between rays and waves: fat
ray tomography. Physics of the Earth and Planetary Interiors, 123(2):127–147.

Jocker, J., Spetzler, J., Smeulders, D., and Trampert, J. (2006). Validation of first-order
diffraction theory for the traveltimes and amplitudes of propagating waves. Geophysics,
71(6):T167–T177.

Keary, P., Brooks, M., and Hill, I. (2002). An Introduction to Geophysical Exploration.
Blackwell Science.
76 Bibliography

Knödel, K., Lange, G., and Voigt, H. (2007). Environmental Geology: Handbook of Field
Methods and Case Studies. Springer.

Lanz, E., Maurer, H., and Green, A. G. (1998). Refraction tomography over a buried waste
disposal site. Geophysics, 63(4):1414–1433.

Liu, Y., Dong, L., Wang, Y., Zhu, J., and Ma, Z. (2009). Sensitivity kernels for seismic Fresnel
volume tomography. Geophysics, 74(5):U35–U46.

Marquardt, D. W. (1970). Generalized inverses, ridge regression, biased linear estimation,


and nonlinear estimation. Technometrics, 12(3):591–612.

Meles, G. A., Greenhalgh, S. A., Green, A. G., Maurer, H., and Van der Kruk, J. (2012).
GPR full-waveform sensitivity and resolution analysis using an FDTD adjoint method.
IEEE Transactions on Geoscience and Remote Sensing, 50(5):1881–1896.

Menke, W. (1989). Geophysical Data Analysis: Discrete Inverse Theory. Academic press,
San Diego.

Oldenburg, D. W. and Li, Y. (2005). Inversion for applied geophysics: A tutorial. In Near-
Surface Geophysics, chapter 5, pages 89–150. Society of Exploration Geophysicists.

Paige, C. C. and Saunders, M. A. (1982). LSQR: An algorithm for sparse linear equations and
sparse least squares. ACM Transactions on Mathematical Software (TOMS), 8(1):43–71.

Podvin, P. and Lecomte, I. (1991). Finite difference computation of traveltimes in very con-
trasted velocity models: a massively parallel approach and its associated tools. Geophysical
Journal International, 105(1):271–284.

Pratt, R. G. and Goulty, N. R. (1991). Combining wave-equation imaging with traveltime


tomography to form high-resolution images from crosshole data. Geophysics, 56(2):208–224.

Rawlinson, N. and Sambridge, M. (2003). Seismic traveltime tomography of the crust and
lithosphere. Advances in Geophysics, 46:81–199.

Reiser, F., Schmelzbach, C., Horstmeyer, H., Sollberger, D., Rabenstein, L., Maurer, H., and
Robertsson, J. (2014). Characterization of the Neuhauserwald Quaternary valley, north-
ern Switzerland, using high-resolution seismic-reflection and seismic-refraction imaging. In
Geophysical Research Abstracts, EGU General Assembly 2014, volume 16, EGU2014-4057.

Schmelzbach, C., Reiser, F., Sollberger, D., Rabenstein, L., Horstmeyer, H., Sutter, E., Mau-
rer, H., Robertsson, J., and Greenhalgh, S. (2014). Multi-method geophysical imaging of a
Quaternary valley in northern Switzerland. In 84th International annual meeting, Society
of Exploration Geophysicists, accepted.

Schmelzbach, C., Zelt, C., Juhlin, C., and Carbonell, R. (2008). P-and SV-velocity structure
of the South Portuguese Zone fold-and-thrust belt, SW Iberia, from traveltime tomography.
Geophysical Journal International, 175(2):689–712.

Schuster, G. T. and Quintus-Bosz, A. (1993). Wavepath eikonal traveltime inversion: Theory.


Geophysics, 58(9):1314–1323.
Bibliography 77

Sheng, J. and Schuster, G. T. (2003). Finite-frequency resolution limits of wave path travel-
time tomography for smoothly varying velocity models. Geophysical Journal International,
152(3):669–676.

Slaney, M., Kak, A. C., and Larsen, L. E. (1984). Limitations of imaging with first-
order diffraction tomography. IEEE Transactions on Microwave Theory and Techniques,
32(8):860–874.

Snieder, R. and Lomax, A. (1996). Wavefield smoothing and the effect of rough velocity per-
turbations on arrival times and amplitudes. Geophysical Journal International, 125(3):796–
812.

Spetzler, J. and Snieder, R. (2004). The Fresnel volume and transmitted waves. Geophysics,
69(3):653–663.

Tarantola, A. (1987). Inverse Problem Theory: Methods for Data Fitting and Model Param-
eter Estimation. Elsevier Science Pub. Co. Inc., New York.

Vasco, D. W., Peterson Jr, J. E., and Majer, E. L. (1995). Beyond ray tomography: Wavepaths
and Fresnel volumes. Geophysics, 60(6):1790–1804.

Vidale, J. (1988). Finite-difference calculation of travel times. Bulletin of the Seismological


Society of America, 78(6):2062–2076.

Virieux, J. (1986). P-sv wave propagation in heterogeneous media: Velocity-stress finite-


difference method. Geophysics, 51(4):889–901.

Watanabe, T., Matsuoka, T., Ashida, Y., et al. (1999). Seismic traveltime tomography using
Fresnel volume approach. In SEG, Expanded Abstracts, volume 18, pages 1402–1405.

Woodward, M. J. (1992). Wave-equation tomography. Geophysics, 57(1):15–26.

Wu, R.-S. and Toksöz, M. N. (1987). Diffraction tomography and multisource holography
applied to seismic imaging. Geophysics, 52(1):11–25.

Zelt, C. A. (1998). Lateral velocity resolution from three-dimensional seismic refraction data.
Geophysical Journal International, 135(3):1101–1112.

Zelt, C. A. (1999). Modelling strategies and model assessment for wide-angle seismic travel-
time data. Geophysical Journal International, 139(1):183–204.
78 Bibliography
Appendix A

Fat ray tomography algorithm

A-1 Matlab code of forward solver


function all = cj_Fat_SolveForward ( all ) 1
%
tic
wbb = waitbar (0 , ’ Solve forward problem ... ’) ;
% Initialize some variables
slow = all . vpm . h ./ all . vpm . vp ; 6
s l o w = r e s h a p e ( slow , all . vpm . nxz , 1 ) ;
tt = zeros ( size ( slow ) ) ;
all . data . tpcalc = zeros ( all . data . n ,1) ;
all . data . raylen = zeros ( all . data . n ,1) ;
all . data . tpw_var = ones ( all . data . n ,1) ; 11
if ( all . inv . b h i n v _ m o d e > 0)
all . data . alpha_s = nan ∗ ones ( all . data . n ,1) ;
all . data . alpha_r = nan ∗ ones ( all . data . n ,1) ;
end ;
x = ( 0 : a l l . v p m . n x −1) ; 16
z = ( 0 : a l l . v p m . n z −1) ;
[ xx , z z ] = m e s h g r i d ( x , z ) ;

% Initialize index arrays for G matrix


% ### - - - - - - - - - CJ - - - - - - - - - ### 21
i f ( a l l . i n v _ m o d e == 1 )
Gi = [ ] ;
Gj = [ ] ;
Gv = [ ] ;
end ; 26
% ### - - - - - - - - - - - - - - - - - - - - - - ###
% Open output file for ray segements
if ( all . r a y o u t > 0)
nrays = 0;
f i d _ r a y = fopen ( [ all . fname . w o r k i n g _ d i r e c t o r y ’ r a y _ t m p . dat ’ ] , ’ wb ’ ) ; 31
fwrite ( fid_ray , nrays , ’ int ’ ) ;
end ;
% ### - - - - - - - - - CJ - - - - - - - - - ###
e x=i s f i e l d ( a l l . i n v , ’ M ’ ) ;
i f e x ==0; % S e t u p a l l . i n v . M f o r i t e r a t i o n n u m b e r 0 ( v p i n i o n i n v e r s i o n grid ) 36
v m o d e l = r e s h a p e ( a l l . v p m . vp , a l l . v p m . nxz , 1 ) ;
f o r b b =1: a l l . i n v . n v p
k k=a l l . v p m . i n v i d x==b b ;
a l l . i n v . M ( b b , 1 )=m e a n ( 1 . / v m o d e l ( k k ) ) ; % S L O W N E S S
end 41
end
% ### - - - - - - - - - - - - - - - - - - - - - - ###
% Loop over all sources
nerr1 = 0;
nerr2 = 0; 46
for a = 1: all . data . nshot
waitbar ( a/ all . data . nshot )

% C o n v e r t to grid coordinates
sx = all . geo . x ( all . data . shots ( a ) )/ all . vpm . h ; 51
sz = all . geo . z ( all . data . shots ( a ) )/ all . vpm . h ;
ii = find ( all . data . s n o == a l l . d a t a . s h o t s ( a ) ) ;
80 Fat ray tomography algorithm

rx = all . geo . x ( all . data . rno ( ii ) ) / all . vpm . h ;


rz = all . geo . z ( all . data . rno ( ii ) ) / all . vpm . h ;
nrec = length ( rx ) ; 56

% Check if s o u r c e s and r e c e i v e r s are within the grid


i f ( ( s x < 1 ) | ( s x > ( a l l . v p m . n x −1) ) | ( s z < 1 ) | ( s z > ( a l l . v p m . n z −1) ) )
error ( ’ Source % d (% f % f ) is o u t s i d e of the grid ’ , all . data . shots ( a ) , sx ∗ all . vpm . h , sz ∗ all . vpm . h
);
end ; 61
for b = 1: nrec
i f ( ( r x ( b ) < 1 ) | ( r x ( b ) > ( a l l . v p m . n x −1) ) | ( r z ( b ) < 1 ) | ( r z ( b ) > ( a l l . v p m . n z −1) ) )
error ( ’ R e c e i v e r % d (% f % f ) is o u t s i d e of the grid ’ , all . data . rno ( ii (1) ) , rx ( b )∗ all . vpm . h ,
rz ( b )∗ all . vpm . h ) ;
end ;
end ; 66

% C o m p u t e travel times ( source tt field , cj )


t t = t i m e _ 2 d ( a l l . v p m . nx , a l l . v p m . nz , s l o w , sx , s z ) ;
a l l . d a t a . t p c a l c ( i i ) = i n t e r p 2 ( xx , zz , r e s h a p e ( tt , a l l . v p m . nz , a l l . v p m . n x ) , rx , r z ) ;
all . data . tpcalc ( ii ) = all . data . tpcalc ( ii ) .∗ all . data . g e o c o r r ( ii ) ; 71

% To c o r r e c t i o n
i f ( a l l . i n v . t o m o d e == 1 ) , a l l . d a t a . t p c a l c ( i i ) = a l l . d a t a . t p c a l c ( i i ) + a l l . d a t a . t o ; e n d ;
i f ( a l l . i n v . t o m o d e == 2 ) , a l l . d a t a . t p c a l c ( i i ) = a l l . d a t a . t p c a l c ( i i ) + a l l . d a t a . t o ( a ) ; e n d ;
76
% ### - - - - - - - - - CJ - - - - - - - - - ###
i f ( a l l . i n v _ m o d e == 1 )
% C o m p u t e fatray and set up G matrix
f o r b =1: l e n g t h ( i i )
% R e c e i v e r tt field 81
t t r=t i m e _ 2 d ( a l l . v p m . n x , a l l . v p m . n z , s l o w , r x ( b ) , r z ( b ) ) ;
% sum of src and rec tt field
t t s u m=t t+t t r ;
d t=t t s u m −a l l . d a t a . t p c a l c ( i i ( b ) ) ;
% Setup sensitivity kernel (w: weight ) after Watanabe (1999) 86
w = 1 − 2∗ a l l . i n v . f r e q ∗ d t ; % L i n e a r w e i g h t
w ( w < 0) = 0 ;
% w (w >0) =1; % No weight
% sum of all cells c o v e r e d by fat ray ( A )
s u m _ w = s u m ( w ( w >0) ) ; % 91

% sum within each i n v e r s i o n cell ( ak )


G R O W _ i n i = s p a r s e ( ( a l l . v p m . i n v i d x +1) , ( a l l . v p m . i n v i d x ) ∗0+1 , w ) ; % e d i t e d cs
% SENSITIVITY calculation ( row wise )
GROW = G R O W _ i n i ( 2 : end ) ./ sum_w .∗ all . data . tpcalc ( ii ( b ) ) ./ all . inv . M ; 96

% Write rows to G matrix


Gj2 = find ( GROW ) ;
Gi2 = ii ( b )∗ ones ( length ( Gj2 ) ,1) ;
Gv2 = GROW ( Gj2 ) ; 101
Gi = [ Gi ; Gi2 ] ;
Gj = [ Gj ; Gj2 ] ;
Gv = [ Gv ; Gv2 ] ;
end
% ### - - - - - - - - - - - - - - - - - - - - - - ### 106

end ; % if ( all . i n v _ m o d e == 1)
end ; % for a 1: all . data . nshot

111
% Create G matrix and compute residuals and RMS
i f ( a l l . i n v _ m o d e == 1 )
a l l . d a t a . G = s p a r s e ( Gi , Gj , Gv , a l l . d a t a . n , a l l . i n v . n v p ) ; % #### CJ #####
end ;
116
all . inv . res = all . data . tp − all . data . tpcalc ;
all . inv . resw = all . inv . res .∗ all . data . tpw .∗ all . data . tpw_var ;
a l l . i n v . r m s ( a l l . i n v . i t n o +1) = s q r t ( s u m ( a l l . i n v . r e s w . ˆ 2 ) / s u m ( a l l . d a t a . t p w .∗ all . data . tpw_var ) ) ;

i n v 2 d m _ m e s s a g e ( s p r i n t f ( ’ RMS i t e r a t i o n % d : % f [ ms ] ’ , . . . 121
a l l . i n v . i t n o , 1 0 0 0 . 0 ∗ a l l . i n v . r m s ( a l l . i n v . i t n o +1) ) ) ;
if ( ( a l l . i n v . i t n o > 0 ) & ( a l l . i n v . i t n o <= a l l . i n v . n i t ) )
i n v 2 d m _ m e s s a g e ( s p r i n t f ( ’ RMS r e d u c t i o n of i t e r a t i o n % d : % f [ ms ] (% f %%) ’ , . . .
all . inv . itno , . . .
−( a l l . i n v . r m s ( a l l . i n v . i t n o +1) − a l l . i n v . r m s ( a l l . i n v . i t n o ) ) ∗ 1 0 0 0 , . . . 126
( a l l . i n v . r m s ( a l l . i n v . i t n o +1) − a l l . i n v . r m s ( a l l . i n v . i t n o ) ) / . . .
all . inv . rms ( all . inv . i t n o ) ∗ 100) ) ;
end ;

% Save tpcalc 131


tpcalc = all . data . tpcalc ;
save ( [ all . fname . working_directory sprintf ( ’ tpcalc_ %03 d . mat ’ , all . inv . itno ) ] , ’ tpcalc ’) ;

% CJ : Save G matrix
G=a l l . d a t a . G ; 136
save ( [ all . fname . working_directory s p r i n t f ( ’ G_ %03 d . mat ’ , all . inv . itno ) ] , ’G ’ ) ;

toc
close ( wbb ) ;
Appendix B

Exploration-scale data: Supplementary


information

B-1 Alternative initial models

Anticlinal model

0 0

z [m]
500 1000
z [m]

1000 2000
1500 14000 12000 10000 8000 6000 4000 2000 0
x [m]
2000
2000 4000 6000
3000 3500 4000 4500 5000 5500
Velocity [m/s]
Velocity [m/s]

Figure B-1: 1D velocity-depth profile of initial model (left). 2D initial model (right).

No. of data 42’771


No. iterations 5
Damping 200
Smoothing 8’000
Max. vp [m/s] 5’800
RMS initial model [ms] 66.8
RMS inversion result [ms] 9.6

Table B-1: Inversion parameters for classical ray tomography (anticlinal initial model).
82 Exploration-scale data: Supplementary information

Inversion result

Figure B-2 shows the result of classical ray tomography using the velocity field shown in
Figure B-1 as initial model for the tomographic inversion.

(a) 0
z [m]

1000

2000
14000 12000 10000 8000 6000 4000 2000 0
x [m]

2500 3000 3500 4000 4500 5000 5500


Velocity [m/s]

(b) 0
z [m]

1000

2000
14000 12000 10000 8000 6000 4000 2000 0
x [m]

−2000 −1500 −1000 −500 0 500 1000 1500 2000


Velocity residual [m/s]

Figure B-2: (a) Inversion result. (b) Difference between the inversion result (Figure B-2a) and
the initial model (Figure B-1)

Multi gradient model

0 0
z [m]

500 1000
z [m]

1000
2000
1500 14000 12000 10000 8000 6000 4000 2000 0
x [m]
2000
0 5000 10000
2000 2500 3000 3500 4000 4500 5000 5500 6000
Velocity [m/s]
Velocity [m/s]

Figure B-3: 1D multi-gradient of initial model (left). 2D initial model (right).

Inversion result

Figure B-4 shows the result of classical ray tomography using the velocity field shown in
Figure B-3 as initial model for the tomographic inversion.
B-1 Alternative initial models 83

No. of data 42’779


No. iterations 6
Damping 200
Smoothing 8’000
Max. vp [m/s] 5’800
RMS initial model [ms] 75.5
RMS inversion result [ms] 10.9

Table B-2: Inversion parameters for classical ray tomography (multi gradient initial model).

(a) 0
z [m]

1000

2000
14000 12000 10000 8000 6000 4000 2000 0
x [m]

2000 2500 3000 3500 4000 4500 5000 5500


Velocity [m/s]

(b) 0
z [m]

1000

2000
14000 12000 10000 8000 6000 4000 2000 0
x [m]

−2000 −1500 −1000 −500 0 500 1000 1500 2000


Velocity residual [m/s]

Figure B-4: (a) Inversion result. (b) Difference between the inversion result (Figure B-4a) and
the initial model (Figure B-3)
84 Exploration-scale data: Supplementary information

B-2 25 Hz and 75 Hz fat ray tomography

Velocity residual [m/s]


(a) 0

1000
1000
z [m]

0
2000
−1000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

Velocity residual [m/s]


(b) 0 2000

1000
z [m]

0
2000

−2000
14000 12000 10000 8000 6000 4000 2000 0
(c) x [m]
Traveltime residual [ms]

200

Initial model
−200
Inversion result

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
(d) Offset [m]
Traveltime residual [ms]

200

Initial model
−200
Inversion result

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Offset [m]

Figure B-5: (a) Difference between the result of the 25 Hz fat ray tomography (Figure 5-7a) and
the initial model (Figure 5-4b). (b) Difference between the result of the 75 Hz fat
ray tomography (Figure 5-7c) and the initial model (Figure 5-4b). (c) Traveltime
residuals vs. offset for 25 Hz fat ray tomography. (d) Traveltime residuals vs. offset
for 75 Hz fat ray tomography..
B-3 Differences between results of different tomographic inversions 85

B-3 Differences between results of different tomographic inver-


sions

0
(a)
1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−600 −400 −200 0 200 400 600


Velocity residual [m/s]

0
(b)
1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−500 −400 −300 −200 −100 0 100 200 300 400 500
Velocity residual [m/s]

Figure B-6: (a) Difference between the results of classical ray and 25 Hz fat ray tomography.
(b) Difference between the results of classical ray and 75 Hz fat ray tomography.
86 Exploration-scale data: Supplementary information

(a) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−250 −200 −150 −100 −50 0 50 100 150 200 250


Velocity residual [m/s]

(b) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−400 −300 −200 −100 0 100 200 300 400


Velocity residual [m/s]

(c) 0

1000
z [m]

2000

14000 12000 10000 8000 6000 4000 2000 0


x [m]

−250 −200 −150 −100 −50 0 50 100 150 200 250


Velocity residual [m/s]

Figure B-7: (a) Difference between the results of 50 Hz and 25 Hz fat ray tomography. (b)
Difference between the results of 75 Hz and 25 Hz fat ray tomography.(c) Difference
between the results of 75 Hzand 50 Hz fat ray tomography.
Appendix C

Small-scale high-resolution data:


Supplementary information

C-1 Herglotz-Wiechert initial model: 50 Hz fat ray tomography

Fat ray 50 Hz
No. of data 7’196
No. iterations 11
Damping 50
Smoothing 50
Max. vp [m/s] 5’000
RMS initial model [ms] 10.2
RMS inversion result [ms] 4.3

Table C-1: Inversion parameters for 50 Hz fat ray tomography (H-W initial model).
88 Small-scale high-resolution data: Supplementary information

50

100
z [m]

150

200

250

800 700 600 500 400 300 200 100 0


x [m]

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 6000
Velocity [m/s]

Figure C-1: Final model obtained using 50 Hz fat ray tomography and the H-W initial model.
Appendix D

Herglotz-Wiechert method

Here I give only the briefest of descriptions of the Herglotz-Wiechert method. For a detailed
mathematical treatment, the reader is referred to Aki and Richards (2002). Following the
derivation in Aki and Richards (2002), the depth of penetration zj (cj ) for a given ray having
velocity cj at its deepest point and recorded at a source-receiver offset distance Xj is obtained
from:

ZXj
1
zj (cj ) = cosh−1 (pcj ) dX, (D-1)
π
0

Here p is the ray parameter, which can be determined from the traveltime data T (X) as the
slope of the curve at any point X:

dT sin i sin i0 1
p= = = = . (D-2)
dX c(z) c0 c(zmax )

In practice, dT /dX is determined by fitting a spline function to the T − X data (which need
not be equi-spaced), which is the smoothest curve through the given points in the sense that
the area under the curve is minimal. In fact, the same spline routine can be used to calculate
the integral of Eq. (D-1). Note that for the ray in question, which arrives at distance Xj :

1 dT
pj = =
(D-3)
cj dX X=Xj

To find the depth of penetration we must integrate over all previous ray parameters (for rays
having larger departure angles (from the vertical) at the source and penetrating to shallower
depths).
Figure D-1 diagrammatically illustrates the various terms and rays involved in Eq. (D-1).
90 Herglotz-Wiechert method

The raw T − X data on which the H-W inversion was performed for the example of Chapter
6, is shown in Figure D-2a. This involves all first arrivals from all shots. The line of best fit,
which constituted the input to H-W inversion, is shown in red. The corresponding p − X plot
is given in Figure D-2b. The p values (dT /dX values) were obtained from the mathematical
spline sub-routine.

dX
cj =
dT Xj
T
dT
p=
dX

X
0 Xj(cj)

i0 zj

i cj

Figure D-1: Sketch of a ray path and T − X plot.

−4
x 10
(a) 0.25 (b) 12

0.2 10

0.15 8
T [s]

0.1 6

0.05 4
Data
Fit
0 2
0 100 200 300 400 500 0 100 200 300 400 500
X [m] X [m]

Figure D-2: (a) Raw T − X data (blue). Red: line of best fit. (b) Ray parameter p (dT /dX
values) versus X.

You might also like