t Special section: Seismic attributes

Spectral similarity fault enhancement

Dustin T. Dewett1 and Alissa A. Henza1
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Abstract
Fault interpretation in seismic data is a critical task that must be completed to thoroughly understand the
structural history of the subsurface. The development of similarity-based attributes has allowed geoscientists to
effectively filter a seismic data set to highlight discontinuities that are often associated with fault systems. Fur-
thermore, there are numerous workflows that provide, to varying degrees, the ability to enhance this seismic
attribute family. We have developed a new method, spectral similarity, to improve the similarity enhancement
by integrating spectral decomposition, swarm intelligence, magnitude filtering, and orientated smoothing. In
addition, the spectral similarity method has the ability to take any seismic attribute (e.g., similarity, curvature,
total energy, coherent energy gradient, reflector rotation, etc.), combine it with the benefits of spectral decom-
position, and create an accurate enhancement to similarity attributes. The final result is an increase in the qual-
ity of the similarity enhancement over previously used methods, and it can be computed entirely in commercial
software packages. Specifically, the spectral similarity method provides a more realistic fault dip, reduction of
noise, and removal of the discontinuous “stair-step” pattern common to similarity volumes.

Introduction similarity attribute would be a major and critical step

The application of seismic attributes to fault identi- (Randen et al., 2001). The idea of similarity enhance-
fication originates from work by Bahorich and Farmer ment directly led to the application of ant colony opti-
(1995) through the development of the coherence algo- mization to fault extraction causing a major step
rithm (crosscorrelation of adjacent seismic traces), forward for the industry (Pedersen et al., 2002). None-
which resulted in great efficiency gains by seismic in- theless, Aqrawi and Boe (2011) make the very specific
terpreters. However, the early coherence attribute per- point several years later, “[a]utomatic fault detection
formed poorly in high-noise data sets. The second-
and extraction is still considered to be a major chal-
generation similarity algorithm, based on multitrace
lenge for the industry.” In an attempt to address this
semblance, has less noise sensitivity (Marfurt et al.,
same problem, Dorn et al. (2012) develop the fault-en-
1998). A major drawback to these methods is the sen-
sitivity to amplitude discontinuities. Garsztenkorn and hanced attribute (also called AFE). Although indepen-
Marfurt (1999) propose a third-generation similarity al- dent of the swarm intelligence methods proposed by
gorithm computed through the calculation of the eigen- Pedersen et al. (2002), it was an equal step forward
values of a covariance matrix over a window of seismic in thinking for the industry.
data, which removed the amplitude sensitivity while in- One of the key tasks in seismic interpretation is the
creasing and localizing the fault response on the seismic identification of faults, and the efficiency gains that are
data. Improvements to dip estimation quickly followed possible through computer-based fault extraction are
through the development of the dip scan method, which enormous. Unfortunately, the similarity attributes de-
provides superior accuracy and precision to dip esti- scribed earlier are known for poor vertical resolution
mates in seismic data (Marfurt, 2006). Numerous edge and are more often used through horizontal slices or
detection algorithms followed including the introduc- horizon extraction based interpretation. This results
tion of the Sobel filter to seismic data by Aqrawi and in poor computer extractions. The major goal of simi-
Boe (2011). larity enhancement is to improve the vertical axis re-
Soon after the development of edge detection algo- sponse of the similarity volume and the computer-
rithms, came the concept of computer-based fault inter- based fault interpretations. We propose a new method
pretation. In 2001, similarity attributes were used for for enhancing the similarity volume, spectral similarity,
computer-based fault extraction, and it was quickly which produces a volume that increases the efficiency
identified that “conditioning” or enhancement of the of fault interpretation several times over traditional

1
BHP Billiton, Petroleum, Houston, Texas, USA. E-mail: [email protected]; [email protected].
Manuscript received by the Editor 16 July 2015; revised manuscript received 10 October 2015; published online 24 February 2016. This paper
appears in Interpretation, Vol. 4, No. 1 (February 2016); p. SB149–SB159, 12 FIGS.
http://dx.doi.org/10.1190/INT-2015-0114.1. © 2016 Society of Exploration Geophysicists and American Association of Petroleum Geologists. All rights reserved.

Interpretation / February 2016 SB149

 human-based fault interpretation and computer-based • Spectral similarity refers to the workflow de-
fault extraction techniques (Figure 1). scribed in this paper for similarity enhancement.

Definitions
Key terms for this paper include the following:
Motivation
• Similarity is a family of edge detection attributes Above, we mention two major steps forward in the
that include coherence, variance, the Sobel filter, enhancement of similarity volumes by Pedersen et al.
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

or similar algorithms. (2002) and Dorn et al. (2012). Both of these approaches
• Swarm intelligence is a family of algorithms that provide a unique look at fault enhancement, which pro-
use decentralized self-organization to perform a duce very different results. As we will discuss, spectral
task (examples include particle swarm optimiza- similarity draws from these ideas with the aim of im-
tion, ant colony optimization, or differential evo- proving upon them. Both of these methods perform well
lution). in many situations. However, we will focus upon their
• Machine learning is a subdiscipline of computer respective weaknesses because they perform poorly in
science that consists of algorithms that can learn similar situations and our goal is to improve upon those
from and make predictions on data (examples in- specific weaknesses.
clude artificial neural networks, self-organized The Pedersen et al. (2002) similarity enhancement
maps, and k-means clustering). uses the ant colony optimization technique (a swarm
intelligence algorithm) to connect discontinuous simi-
larity events and remove the common stair-step
anomaly seen in many similarity-based attributes. This
algorithm excels at fault connectivity and the retention
of appropriate fault dip. However, this technique is very
sensitive to noise and is not generally appropriate for
any but the largest faults (Figure 2a). Aqrawi and
Boe (2011) attempt to improve the results of this swarm
intelligence technique by improving the underlying sim-
ilarity attribute by changing from a semblance-based to
a Sobel filter-based attribute. Although this improved
the results significantly, the inherent sensitivity of ant
colony techniques (and likely all swarm intelligence al-
gorithms) to noise is significant.
The Dorn et al. (2012) AFE is a seismic attribute cen-
tered on user-driven filtering of the similarity magnitude
combined with an orientated smoothing parameter
(Figure 2b). This technique excels in the detection of
large faults, and, owing to the smoothing parameters,
provides excellent fault connectivity in those situations.
However, the technique suffers from poor performance
in the presence of small- to medium-sized faults and
lacks a method of interpolation to increase fault con-
nectivity.
The vast majority of similarity enhancement tech-
niques (including those mentioned above) commonly
suffer from three major classes of detection issues
(although not every method suffers from all classes).
The first is an abundance of near-vertical similarity re-
sponse — an effect likely related to either the algo-
rithm or the underlying similarity attribute (see the
solid rectangles in Figure 2). The second is the inherited
stair-step anomaly that is a common effect seen in the
underlying similarity volume (see the dotted rectangles
in Figure 2). Third, many faults are realistically ori-
ented, but they appear broken and discontinuous
Figure 1. Vertical section of the (a) spectral similarity coren-
(see the dashed rectangles in Figure 2). It is these three
dered with seismic amplitude and (b) faults extracted using classes of similarity enhancement problems that we are
computer-based fault interpretation derived from the spectral attempting to improve through our proposed spectral
similarity corendered with seismic amplitude. similarity attribute workflow.

SB150 Interpretation / February 2016

 Spectral similarity significant frequency contrasts between the shallow
As shown in Figure 2c, spectral similarity improves and deep section (e.g., subsalt Gulf of Mexico), one can
upon all three classes of problems discussed above. create a spectral similarity with frequency components
Spectral similarity results in fault dips that are in gen- for a shallow section that are drastically different than
eral agreement with the expectations by area experts for the deeper section. Another beneficial feature of
and structural geologists, the removal of stair-step er- spectral similarity is the ability to incorporate any
rors, and greatly improved fault connectivity. In addi- attribute type into the process (e.g., dip magnitude, cur-
tion, the range of values (represented by the change vature, reflector rotation, etc.). It is precisely this cus-
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

in fault colors from light gray to black) implies a con- tomizable feature that allows spectral similarity to excel
fidence or probability in the volume. In practice, this in every data set and basin in which it was applied (re-
attribute, when used with a computer-based fault inter- gardless of differences in geology).
pretation technique, can be further filtered by this con-
fidence to yield very realistic fault surfaces. Workflow description
As discussed below, we begin with spectral decom- The general form of our workflow is independent of
position; therefore, each attribute parameter and filter specific techniques or approaches (Figure 3). The spec-
is customized to a given frequency band. Our technique tral similarity workflow provides a great deal of custom-
is, therefore, highly adaptable to a range of data sets. ization based on individual preferences, data quality,
For example, when computing a similarity volume, one and time constraints. Therefore, it is not possible to de-
key input is the vertical window height, which ideally is fine the exact workflow for any given data set or
a function of the dominant wavelength of the interval of project, but the optimal customized algorithm is quickly
interest. This results in suboptimal parametrization in identified when constraints (data or time) are applied.
all areas with different dominant wavelengths (which We begin with a seismic data set that is filtered, as
can vary laterally and vertically). However, in a spectral needed, for attribute analysis. Then we follow with
volume, the optimal window height represents the dom- spectral decomposition (e.g., short-time Fourier trans-
inant wavelength of the entire volume. If a data set has form, continuous wavelet transform [CWT], matching

Figure 2. Composite vertical section image

of the same crossline using various edge en-
hancement calculations. (a) Swarm intelli-
gence method proposed by Pederson et al.
(2002). (b) Strike- and dip-based enhancement
proposed by Dorn et al. (2012). (c) Spectral
similarity proposed in this paper. Three differ-
ent classes of anomalies are highlighted based
on the characteristics of the Pederson et al.
and Dorn et al. style enhancements: (1) solid
rectangles indicate areas of poorly imaged
faults, (2) dotted rectangles indicate areas
in which are moderately well imaged, and
(3) dashed rectangles indicate well-imaged
faults improved in the spectral similarity
attribute.

Interpretation / February 2016 SB151

 pursuit, crosscorrelation, or constrained least-squares into edge-filtering and smoothing operations. We then
spectral analysis; Figure 3a). These peak-frequency combine each of these spectral fault-enhanced volumes
volumes (i.e., spectral voice or similar) are used to com- through addition, resulting in a similarity that has been
pute the seismic attributes (Figure 3b). Our experience enhanced through the inclusion of spectral decomposi-
indicates that higher frequencies (greater than 30 Hz) tion (Figure 3d). An alternative method to combine the
are more beneficial than are low frequencies (less than various volumes is through a machine-learning algo-
15 Hz), but this is data dependent. We commonly use rithm (in our tests, we used a self-organized map), a
geometric attributes calculated from these data; how- technique adapted from Basir et al. (2013). This adds
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ever, this workflow can be adopted to use other deriva- additional computation time, but it significantly reduces
tive volumes (e.g., the spectral phase) directly by the amount of intermediate data volumes created (Fig-
skipping the attribute generation step and applying the ure 4 shows the results of this optional approach).
swarm intelligence-based attribute to these volumes di-
rectly. Numerous frequency-based attributes are used Workflow customization
as input to swarm intelligence for lineament connection It is common for individuals to have preferences and
and interpolation between discontinuous events (Fig- biases to particular algorithms. This is why we present
ure 3c). In the final portion of the workflow, we use the spectral similarity algorithm in generic terms. The
each of these swam intelligence volumes as an input optimum spectral decomposition or similarity method

Figure 3. The spectral similarity workflow

and its four main stages of generation. The bi-
furcation in stage III allows for two different
techniques to combine the results.

SB152 Interpretation / February 2016

 will vary depending on the data specifics. Moreover, (Figure 5c) is a prominent feature along which many
individual biases or company policy may further con- faults nucleate and terminate, creating a fault trend that
strain algorithms (e.g., if corporate policy maintains is oblique to the regional paleostress field. In addition,
the use of only Sobel filter similarity). In addition, some early movement of the underlying salt created local var-
algorithms and implementations take considerably iations in the stress regime that affected later faulting
longer to compute than do others. Therefore, when time (Figure 5b and 5d). Owing to the local and regional pa-
constraints, personal/corporate bias, and data-depen- leostresses, we expect the fault patterns to be dominated
dent, goal-based constraints are applied, the ideal work- by normal faults with orientations that vary laterally. The
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

flow for a given project is easily identified. data set is of excellent quality with minimal noise. The
notable exception to this is the relatively low signal-to-
Case study no. 1: High signal to noise noise area (Figure 5a).
The first data set chosen to illustrate our method is
from central Texas, USA. In general, this area is an ex- Customized workflow
tensional regime, with normal faults striking approxi- We began with a crosscorrelation-based spectral de-
mately perpendicular to the extension direction. How- composition (Gao, 2013). This type of spectral decom-
ever, the underlying structure of the region and vertical position is acceptable for geometric interpretation, and
stratigraphic variations has influenced the deformation the computation time is shorter than for other methods.
patterns in this package, resulting in multiple fault ori- We used three volumes at approximately 20, 37, and
entation trends and detachment levels. A paleoreef 43 Hz to compute a modified eigenstructure similarity

Figure 4. A time slice of a machine learning

algorithm (self-organized maps) being used to
combine three swarm-intelligence volumes
into one prior to the filter and smoothing step.

Figure 5. Time slice through the seismic am-

plitude volume corendered with the (a) fault-
enhanced similarity and (b) the spectral sim-
ilarity that illustrates the data quality and lat-
eral variation of seismic character in the data
set. The rectangles highlight areas of interest,
which include (A) a low signal-to-noise zone,
(B) lateral changes to geology, (C) the re-
sponse to stratigraphic thinning, and (D) a
high signal-to-noise zone with large faults. A
histogram for each similarity enhancement
is shown highlighting the range of values of
those volumes.

Interpretation / February 2016 SB153

 for each spectral volume (Garsztenkorn and Marfurt, volume. The response was a preference among all geo-
1999). When calculating dip, we used a dip-scan method scientists for the spectral similarity algorithm (approx-
with a maximum of 30° and an increment of 2° (Marfurt, imately 96% favorable).
2006). We followed with swarm intelligence described Figure 6 illustrates two fault enhancements to the
by Pedersen et al. (2002), and a magnitude-based filter- similarity volume, the fault-enhanced volume (Dorn
ing and smoothing operation (Dorn et al., 2012) on each et al., 2012) and spectral similarity discussed in this
frequency volume. Finally, we added the three fre- paper. The quality of the enhancement varies laterally,
quency-based attribute volumes together. The total size owing to zones of relatively poor signal to noise and
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

of the original seismic is approximately 17 GB, and the changes in peak frequency. The spectral similarity
total size of all intermediate volumes, parameter tests, results in more distinct fault patterns and interactions
software projects, duplicated data, and SEGY exports is than the fault-enhanced volume. Specifically, the spec-
969 GB. Most of these data were intermediate scratch tral similarity volume (Figure 6b) performs very well in
data that were not retained. area A, where a large fault zone significantly reduces
the signal-to-noise ratio (S/N). Areas B and D have
Results significantly improved fault connectivity in the spectral
To evaluate the quality of the resultant spectral sim- similarity volume and provide an excellent guide for
ilarity, we asked two dozen structural and geophysical fault interpretation and refinement of existing fault
experts (with an average of 17 years of experience, surfaces. In addition, the increased range of values of
knowledge of multiple basins, and several Ph.D. hold- the data yields an implied level of confidence directly
ers) to compare our results to a traditional similarity from the data volume. The darker faults are more pro-

Figure 6. Time slice of the (a) fault-enhanced volume and (b) spectral similarity volume derived from the amplitude shown in
Figure 5. In area A, the noise from a major fault zone makes any fault interpretation from the fault-enhanced volume difficult,
whereas spectral similarity can easily interpret the major faults. The faults in the fault-enhanced volume in area B (which cuts into
the overlying formation) have been filtered out. The faults in this same area in the spectral similarity are present and clear, owing to
the multiple volumes that comprise the spectral similarity attribute. Similarly, areas C and D in the fault-enhanced volume lack the
fault connectivity and clarity that is present in the spectral similarity.

SB154 Interpretation / February 2016

 nounced in the spectral similarity method than in the how an interpreter can use both products for manual
fault-enhanced method, whereas the lighter colors or computer-based interpretation. Preexisting faults
may be smaller faults or even an artifact (i.e., the typical can also be refined by using spectral similarity for qual-
double-fault expression commonly seen in time-migrated ity control. The noticeable increase in the sharpness of
data). Differences between the fault-enhanced and the the fault response combined with the range of values of
spectral similarity method are also visible in the vertical the data volume imply that computer-based fault ex-
section (Figure 7a; the full supplementary figure can be traction techniques would perform very well with the
accessed through the following link: s1.mpg). The fault- spectral similarity attribute.
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

enhanced method commonly results in fault dips that are To investigate the validity of the features identified
nearly vertical, whereas the spectral similarity method by spectral similarity, we extracted faults over a short
results in faults that are dipping at moderate angles time interval using a data range cutoff and enforcing a
(Figure 7b). strict lower limit to the number of points clustered (via
The ultimate goal of any similarity product is to aid in k-means clustering) to 10,000. We extracted 332 total
interpretation. Figure 8 shows the spectral similarity fault planes, which were then checked for quality by
corendered with the seismic amplitude, illustrating a Ph.D. structural geologist with experience in this

Figure 7. Corendered vertical section com-

paring the results from the (a) fault-enhanced
similarity enhancement and (b) spectral sim-
ilarity. The red rectangle highlights an area
where the spectral similarity shows more ac-
curate fault dip, and the yellow arrows indi-
cate areas where the spectral similarity has
increased the connectivity of the fault re-
sponse.

Interpretation / February 2016 SB155

 basin. It was determined that 33% of the faults (108) re- the computer, but their cluster sizes were below our
quired no editing, 63% of the faults (209) required only 10,000-point lower bound. In the authors’ experience,
basic editing, and the remaining 4% (15) required edit- the faults extracted from previous similarity attributes
ing of the 3D mesh (the software incorrectly connected (and enhanced versions) require more significant edit-
the points). Moreover, “basic editing” consisted of split- ing, and the spectral similarity attribute greatly increases
ting (dominantly in the vertical direction) the correctly the efficiency of fault interpretation over traditional com-
placed points that were clustered together to make puter-based fault extraction workflows. In fault extrac-
larger fault planes. These results are shown in Figure 9. tion comparisons, conducted by the same Ph.D.
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Our initial constraints, specifically the number of points structural geologist, similar numbers of faults were inter-
required per cluster, were overly conservative, as indi- preted by the computer, but the extracted surfaces were
cated from the number of readily identifiable faults left overly vertical and required significantly more complex
uninterpreted. These faults were, in fact, interpreted by and time-consuming point editing. Using an average

Figure 8. Time slice from the spectral simi-

larity rendered with a crossline of seismic am-
plitude illustrating results of semimanual fault
interpretation using commercial software. The
yellow, blue, and purple points are interpreted
fault planes.

Figure 9. Computer-based fault interpretation using a minimum point population of 10,000 points per cluster rendered with the
spectral similarity volume. The yellow faults require no edits, the blue faults require minor edits, and the red faults require in-
terpolated mesh edits. The result was an increase in productivity of 6×–8× computer-based fault interpretations on other attribute
volumes.

Figure 10. Corendered time slice of peak frequency and peak magnitude from CWT spectral decomposition with the spectral
similarity volume. Peak frequency/magnitude provides an indication of lateral stratigraphic variation, and the spectral similarity
provides structural information.

SB156 Interpretation / February 2016

 elapsed time over 20 fault edits from faults extracted on area created local variations from the regional stress
the spectral similarity attribute and commercially avail- regime that influenced fault orientations (Figure 11b,
able similarity enhancements, we estimate an increase rectangle).
in human productivity of 6×‒ 8× (the computation time
is not accounted for). Customized workflow
In addition to fault surface interpretation, the spec- In contrast to case study no. 1, this data set is
tral similarity attribute can provide a clear way to heavily contaminated by noise (Figure 11, left side; the
understand and communicate geologic complexity full supplementary figure can be accessed through the
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

through multiattribute analysis and display. By combin-

ing this structural attribute with the peak frequency and
peak magnitude from spectral decomposition, we can
display the structural grain from the spectral similarity
and the lateral lithologic variation as interpreted from
the peak frequency and peak magnitude volumes (Fig-
ure 10). In this case, the low signal-to-noise zone in the
lower center of the image is easily identifiable, while
still displaying the major fault trends through this area.
These faults were later confirmed through quality con-
trol and manual interpretation of the amplitude volume
by basin experts. Many faults act as boundaries to the
peak frequency, whereas other similar faults do not.
This type of information may be beneficial to well plan-
ning and well placement in various reservoirs.

Case study no. 2: Low signal to noise

Similar to case study no. 1, the area for case study
no. 2 is in an extensional regime, with normal faults
striking approximately perpendicular to the extensional
direction. Basement structures and salt movement have
influenced the deposition and deformation patterns of
the study area. Movement on a basement fault was ac-
commodated by a series of faults oblique to the regional
extension direction (Figures 11 and 12 rectangle). In ad-
dition, salt movement in the northeast area of the study

Figure 12. Time slice at the level of Figure 11 from (a) the
spectral similarity volume and (b) a modified eigenstructure
Figure 11. Time slice comparison of data prior to filtering similarity. The fault zone highlighted by the rectangle is poorly
and data after structure-orientated mean-/median-based filter- resolved on the eigenstructure similarity, but it resolves into a
ing. The rectangle highlights a major fault zone in the data. trend of parallel faults on the spectral similarity.

Interpretation / February 2016 SB157

 following link: s2.mpg). Therefore, as a first step, we swarm intelligence, and orientated filtering into our
applied a series of progressively larger windowed, workflow, which comes at a cost of computation time
structurally orientated, and median/mean combination and scratch storage space. We provide for the use of
filters (Figure 11, right side). Small-scale fault identi- curvature, total energy, or similarity style attributes
fication is limited owing to heavy noise contamination, combined into one fault detection volume. This enables
and noise reduction was paramount. Our filters began the interpreter to identify which attributes highlight
in a 3 × 3 window and were followed by 6 × 6 and 9 × 9 faults (or other linear features) optimally in their data
window filters. As in case study no. 1, we then used set. Our spectral similarity workflow reduces or elimi-
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

crosscorrelation-based spectral decomposition (Gao, nates many algorithmic anomalies present in similarity
2013). Based on the visual inspection, we chose the 67 attributes by estimating a more realistic fault dip, reduc-
and 40 Hz volumes for our analysis. We also used the ing noise, and removing stair-step discontinuities.
full-stack modified eigenstructure similarity (Garsz-
tenkorn and Marfurt, 1999) for quality control. On each
Acknowledgments
peak frequency volume, we computed a Sobel filter-
We would like to thank BHP Billiton and Global Geo-
based similarity (Al-Dossary and Al-Garni, 2013), which
physical Services Inc., for encouraging this work and
was followed by Pedersen et al. (2002)-style swarm in-
for permission to present these results. We would also
telligence and fault enhancement (Dorn et al., 2012).
like to thank our colleagues at BHP Billiton who pro-
When calculating dip, we used a dip-scan method with
vided valuable suggestions and edits. We specifically
a maximum of 60° and an increment of 2° (Marfurt,
would like to thank C. Docherty, who provided valuable
2006). We then volumetrically summed these two vol-
assistance and several images used in this study.
umes. The size of the original seismic is approximately
25 GB, and the total size of all intermediate products,
parameter tests, software projects, duplicated data, and References
SEGY exports was 2.6 TB (63 GB of final products). The Al-Dossary, S., and K. Al-Garni, 2013, Fault detection and
total time required for computation was approximately characterization using a 3D multidirectional Sobel filter:
29 h, and required intermittent human interaction (to Presented at the Saudi Arabia Section Technical Sym-
initiate a process). posium and Exhibition, SPE, SPE-168061-MS.
Aqrawi, A., and T. Boe, 2011, Improved fault segmentation
Results using a dip guided and modified 3D Sobel filter: 81st An-
The data set used in case study no. 2 is both conta- nual International Meeting, SEG, Expanded Abstracts,
minated by noise and heavily affected by poor acquis- 999–1003.
ition coverage. This data set is of a lower fold and has Bahorich, M. S., and S. L. Farmer, 1995, 3‐D seismic discon-
an overall lower S/N than does case study no. 1. None- tinuity for faults and stratigraphic features: The coher-
theless, spectral similarity greatly outperforms full- ence cube: 65th Annual International Meeting, SEG,
stack coherence and provides a clear indication of the Expanded Abstracts, 93–96, doi: 10.1190/1.1887523.
complex fault systems present in the data (Figure 12; Basir, H., A. Javaherian, and M. Yaraki, 2013, Multi-attrib-
the full supplementary figure can be accessed through ute ant-tracking and neural network for fault detection:
the following link: s3.mpg). Specifically, the rectangle A case study of an Iranian oilfield: Journal of Geophys-
from Figure 12 highlights a complex fault system ics and Engineering, 10, 015009, doi: 10.1088/1742-2132/
whose existence is hinted at in the full-stack similarity,
10/1/015009.
yet it is not fully resolved. Spectral similarity allows
Dorn, G., B. Kadlec, and M. Patty, 2012, Imaging faults in 3D
the interpreter to identify and interpret these systems
seismic volumes: 82nd Annual International Meeting,
more effectively.
SEG, Expanded Abstracts, doi: 10.1190/segam2012-1538.1.
Gao, D., 2013, Wavelet spectral probe for seismic structure
Conclusions
interpretation and fracture characterization: A work-
Fault interpretation remains a time-consuming and
flow with case studies: Geophysics, 78, no. 5, O57–O67,
tedious task that is punctuated by moments of complex-
doi: 10.1190/geo2012-0427.1.
ity and difficulty. Edge-detection attributes are a critical
Garsztenkorn, A., and K. J. Marfurt, 1999, Eigenstructure-
tool in the interpreter’s toolbox to assist with and in-
crease efficiency of fault interpretation workflows. based coherence computations as an aid to 3-D struc-
We have demonstrated that our new method of similar- tural and stratigraphic mapping: Geophysics, 64,
ity enhancement, spectral similarity, greatly increases 1468–1479, doi: 10.1190/1.1444651.
the vertical and horizontal response of discontinuities. Marfurt, K. J., 2006, Robust estimates of reflector dip and
Moreover, we believe that spectral similarity lends itself azimuth: Geophysics, 71, no. 4, P29–P40, doi: 10.1190/1
quite readily to computer-based fault extraction tech- .2213049.
niques, and we have shown the potential for a dramatic Marfurt, K. J., R. Kirlin, S. Farmer, and M. Bahorich, 1998,
increase in productivity when using this technique as a 3-D seismic attributes using a semblance-based coher-
basis for such extractions. These improvements are a ency algorithm: Geophysics, 63, 1150–1165, doi: 10.1190/
direct result of the inclusion of spectral decomposition, 1.1444415.

SB158 Interpretation / February 2016

 Pedersen, S. I., T. Randen, L. Sønneland, and O. Steen, of Oklahoma. His research interests include the applica-
2002, Automatic 3D fault interpretation by artificial ants: tion of multiattribute analysis to structural and strati-
72nd Annual International Meeting, SEG, Expanded graphic features and leveraging seismic attributes to
Abstracts, 512–515. increase the efficiency of the seismic interpreter.
Randen, T., S. Pedersen, and L. Sønneland, 2001, Automatic
extraction of fault surfaces from three‐dimensional seis-
Alissa A. Henza received a Ph.D
mic data: 71st Annual International Meeting, SEG, Ex-
(2009) from Rutgers University, study-
Downloaded 08/09/16 to 4.35.206.178. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

panded Abstracts, 551–554.

ing the fault patterns produced by
multiple phases of noncoaxial exten-
sion. Before joining BHP Billiton,
Dustin T. Dewett received an M.S. she co-lead the AAPG/SEPM short
(2013) in geophysics from the Univer- course “Seismic Expression of Struc-
sity of Oklahoma, studying the seismic tural Styles: A Modeling Perspective”
character of Middle-Eastern carbon- at the 2008 AAPG Annual Convention
ates near hydrocarbon reservoirs. and Exhibition. She also assisted with research projects
After joining BHP Billiton in 2013, funded by Husky Energy to investigate the development
he became involved in quantitative of fault patterns during multiple phases of extension in
seismic interpretation in conventional the Jeanne d’Arc Basin, Newfoundland, Canada. Her re-
and unconventional reservoirs. He is a search interests include the effects of pre-existing fault
veteran of the U.S. Navy Submarine Force, and he is cur- patterns on newly forming faults and fractures, and devel-
rently pursuing a Ph.D. in geophysics from the University oping techniques to aid in fault and fracture interpretation.

Interpretation / February 2016 SB159