GEOPHYSICS, VOL. 84, NO. 1 (JANUARY-FEBRUARY 2019); P. O25–O37, 17 FIGS.
10.1190/GEO2018-0369.1
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Image processing of seismic attributes for automatic fault extraction
Jie Qi1, Bin Lyu1, Abdulmohsen AlAli1, Gabriel Machado1, Ying Hu1, and Kurt Marfurt1
ABSTRACT
Along with horizon picking, fault identification and interpre-
tation is one of the key components for successful seismic data
interpretation. Significant effort has been invested in accelerat-
ing seismic fault interpretation over the past three decades. Seis-
mic amplitude data exhibiting good resolution and a high signal-
to-noise ratio are key to identifying structural discontinuities
using coherence or other edge-detection attributes, which in turn
serve as inputs for automatic fault extraction using image
processing or machine learning techniques. Because seismic
data exhibit not only structural reflectors but also seismic noise,
we have developed a fault attribute workflow that contains
footprint suppression, structure-oriented filtering, attribute
computation, “unconformity” suppression, and our new iterative
INTRODUCTION
Conventional interpretation of faults, channels, and horizons
from seismic amplitude data is time consuming. The limits of karst
and salt bodies can be particularly difficult to pick because of their
unclear boundary, discontinuous character, and steep edges (Qi
et al., 2014). Seismic attributes provide quantitative measures of
reflector amplitude, phase, frequency, continuity, and morphology
that can be subsequently used to identify patterns in the seismic data
using either 3D visualization or machines learning techniques. Seis-
mic attributes help interpreters, especially less experienced inter-
preters, in the identification of subtle features that may otherwise
be overlooked in seismic data.
Many geometric attributes have been developed to identify faults,
joints, and natural fractures. One of the first volumetric edge-detec-
tion attributes, coherence, is widely applied to map structural and
stratigraphic discontinuities in 3D seismic volumes. The first gen-
eration of coherence algorithms was proposed by Bahorich and
Manuscript received by the Editor 16 May 2018; revised manuscript received 31 July 2018; published ahead of production 06 September 2018; published
online 26 November 2018.
1
University of Oklahoma, ConocoPhillips School of Geology and Geophysics, Norman, Oklahoma, USA. E-mail: jie.qi@ou.edu (corresponding author);
bin.lyu@ou.edu; abdulmohsen.alali.1@ou.edu; gmachado@ou.edu; ying.hu-1@ou.edu; kmarfurt@ou.edu.
O25
energy-weighted directional Laplacian of a Gaussian (LoG)
operator. In general, tracking faults that exhibit a finite offset
through a suite of conformal reflectors is relatively easy. Instead,
we evaluate the effectiveness of this workflow by tracking faults
through an incoherent mass-transport deposit, where the low-
frequency contribution of multispectral coherence provides a
good fault image. Multispectral coherence also reduces the
“stair-step” fault artifacts seen on broadband data. Application
of statistical filtering can preserve the discontinuity’s bounda-
ries and reject incoherent backgrounds. Finally, iterative appli-
cation of an energy-weighted directional LoG operator provides
improved fault image by sharpening low-coherence anomalies
perpendicular and smoothing low-coherence anomalies parallel
to fault surfaces, while at the same time attenuating locally non-
planar anomalies.
Farmer (1995), which was based on crosscorrelation of seismic
traces. Semblance (Marfurt et al., 1998), eigenstructure (Gerszten-
korn and Marfurt, 1999), gradient structure tensor (Bakker, 2002),
and energy-ratio similarity (Chopra and Marfurt, 2007) are sub-
sequent coherence algorithms that operate on a spatial window of
five or more neighboring traces. Other variations of coherence al-
gorithms, including local structure entropy (Cohen and Coifman,
2002), variance (Van Bemmel and Pepper, 2000), gradient magni-
tude (Aqrawi and Boe, 2011), automated fault extraction (Dorn
et al., 2012), fault likelihood (Hale, 2013), predictive painting (Kar-
imi et al., 2015), directional structure-tensor coherence (Wu, 2017),
and generalized tensor-based coherence (Alaudah and AlRegib,
2017) provide similar results. Some of these algorithms are sensitive
to lateral change in the amplitude (such as the Sobel filter), whereas
other are sensitive to lateral change in the waveform (such as eigen-
structure and GST “chaos”) (Marfurt and Alves, 2015).
Several coherence algorithms are based on a covariance matrix;
others, such as semblance and variance, can be reformulated to do
 O26 Qi et al.

so (Lin et al., 2016). Marfurt et al. (1998) construct a covariance
matrix from a window of analytic traces (the original and its Hilbert
transform) along the structural dip. Smaller windows provide higher
lateral resolution and less stratigraphic mixing, but they may suffer
from random and coherent noise (Lin et al., 2016). Larger windows
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
simpler approach and sum the spectrally limited covariance matri-
ces into a single matrix, whereby the resulting multispectral coher-
ence image retains much of the value of the individual spectrally
limited coherence images.
Image-processing techniques attempt to mimic the human ability

results in smeared but less noisy images. Libak et al. (2017) com-
pute coherence with different sizes of analysis windows and find
that fault geometric attribute may be over- or underestimated if one
applies an inappropriate analysis window. In general, using a large
vertical window can stack steeply dipping fault discontinuities and
suppress stratigraphic discontinuities. However, if the vertical win-
dow is too large, temporally limited discontinuities such as channel
edges will be vertically smoothed. Large windows also exacerbate
stair-step artifacts of dipping faults. A good rule is to use a window
size that approximates the period of the highest frequency of interest.
Li and Lu (2014) and Li et al. (2015) demonstrate that coherence
computed from different spectral magnitude or spectral voice com-
ponents provides superior fault and channel images. Alaei (2012)
computes a fault attribute by integrating the results of spectral de-
composition and coherence estimation, combining the resulting
spectral images using RGB color blending. Dewett and Henza
(2016) build on this observation and combine multiple (greater than
three) spectrally limited coherence images using a self-organizing
map algorithm. Marfurt (2017) and Li et al. (2018) use a somewhat
to identify and extract larger patterns. Randen et al. (2001) and Ped-
ersen et al. (2002) apply a swarm-intelligence algorithm they called
ant-tracking to extract large faults. Their key innovation was the con-
trol of the fault dip azimuth and dip angle search directions by linking
the algorithm to a Schmidt diagram. Cohen et al. (2006) describe an
image-processing technique that processes the local-fault-extraction
volumes by directional filtering and threshold skeletonization along
horizons and vertical slices. Barnes (2006) generates a second mo-
ment tensor within an analysis window to reject anomalies parallel to
stratigraphy and to dilate the image to connect disjointed fault points.
Wu and Hale’s (2015) image-processing technique automatically
extracts intersecting fault planes from fault surfaces from three fault
attribute images of fault likelihood, fault strike, and dip (Hale, 2013).
Machado et al. (2016) and Qi et al. (2016a, 2017b) develop a fault
enhancement and skeletonization workflow based on the second mo-
ment tensor to improve the quality of faults in the coherence attribute.
Other image-processing filters include Hough transforms (AlBinHas-
san and Marfurt, 2003; Aarre and Wallet, 2011), Radon transforms
(Boe, 2012), and log-Gabor filters (Admasu et al., 2006). Zhang et al.

Figure 1. Workflow illustrating the steps used in our attribute preconditioning workflow. The first step is poststack seismic amplitude data
conditioning applying footprint suppression and structure-oriented filtering. This is followed by edge detection using broadband (conventional)
or multispectral coherence, and estimation of the energy about each voxel. Next, we suppress anomalies parallel to the reflector dip. Finally, we
apply an iterative energy-weighted LoG operator and directional skeletonization to the filtered coherence to generate enhanced fault images.
 Seismic attribute preconditioning O27

(2014) adapt biomedical finger-vein analysis tech-

nology to better extract fault anomalies. Wu and
Fomel (2018) propose an optimal surface-voting
method to enhance a fault attribute image and ex-
tract fault surfaces.
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Despite all these improvements, coherence al-

gorithms measure all discontinuities in the analy-
sis window, whether they are the target geologic
discontinuity of interest, a secondary but perhaps
stronger discontinuity or a discontinuity artifact
caused by seismic noise. Unlike a coherence algo-
rithm, human interpreters see discontinuities of
such faults at different scales, allowing them to Figure 2. A representative vertical slice through the seismic amplitude volume. Major
track a fault through a background of unrelated faults are visible above and below the mass-transport deposit, but they are difficult to
often smaller scale discontinuities. track through it.
In this paper, we present an attribute precon-
ditioning workflow to construct fault images that
are sufficiently smooth for subsequent fault au-
tomatic extraction (Figure 1). We begin with
the algorithm description. Then, we apply the
workflow to a data volume from the Taranaki Ba-
sin, where faults cut through a mass-transport de-
posit, starting with incoherent and coherent noise
suppression, followed by computing and sta-
tistically filtering the coherence attribute, and
ending with several iterations of an energy-
weighted directional Laplacian of a Gaussian
(LoG) filtering. We validate our fault attribute
workflow and a commercial swarm-intelligence
workflow using a traditional human-driven fault
interpretation as the ground truth.

Methodology
The second-order moment tensor
Machado et al.’s (2016) fault-enhancement
workflow is based on earlier work by Barnes
(2006), who constructs a second-order moment
tensor of an edge attribute to determine the
hypothesized fault anomalies’ orientation. The
second-order moment tensor is built from a coher-
ence attribute am within an M-voxel analysis win-
dow, where am has been modified so that coherent
portions of the survey have a value of zero (the
same algorithm also enhances edge anomalies
computed using Sobel filters and aberrancy).
We modify Machado et al.’s (2016) algorithm by
observing that all coherence anomalies are not
equally important. Specifically, we wanted to min-
imize fault stair-step artifacts in coherence images
of dipping faults. Lin et al. (2016) recognize that
seismic migration images the seismic wavelet
perpendicular to the reflector dip. Using an ana-
lytic seismic trace (the original trace and its Hil-
bert transform) as input, a reflector discontinuity
occurs where the instantaneous envelope is maxi- Figure 3. Vertical slices through (a) coherence computed from the original data, (b) co-
herence after footprint suppression and structure-oriented filtering, and (c) difference
mum. In the absence of nearby reflectors, away of coherence anomalies between (a and b). Note that coherence after noise rejection
from this point, wavelet sidelobes result in the dis- exhibits a better signal-to-noise ratio. Faults and other discontinuities are more easily
continuity continuing perpendicular to the reflec- interpretable within and through the mass-transport deposit zone.
 O28 Qi et al.

tor rather than along the true fault face. To partially address this prob- define the location of the center of the mass μ of coherence anomalies
lem, we assign a greater weight to discontinuities in which the wave- am within the analysis window to be
let envelope (or energy) is strongest. Additionally, we modified 8 PM
Machado et al.’s (2016) algorithm by noting that the second moment >
> μ1 ¼ P
W m am x1m
>
>
m¼1
;
tensor should be computed about the center of the mass rather than >
>
M
PM m¼1 W m am
<
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

about the center of the window, further improving the results. We

μ2 ¼ P
W a x
m m 2m
m¼1
;; (1)
>
> P
M
W m am
>
>
m¼1

>
M

: μ3 ¼ P
W m am x3m
> m¼1
M ;
W a
m¼1 m m

where xm is the vector distance of the mth voxel

from the center of the analysis window and W m ;is
a weight that depends on the local reflector
strength. The second moment tensor in the analy-
sis window can be written as
0 1
I 11 I 12 I 13
C ¼ @ I 12 I 22 I 23 A; (2)
I 13 I 23 I 33

where the components of the second moment

tensor are
X
M
I jk ¼ W m ðxjm −μj Þðxkm −μk Þam : (3)
m¼1

Eigendecomposition of the energy weighted

second-order moment tensor results in three ei-
genvalues λj and three eigenvectors vj . The values
of λ3 and v3 are key to the subsequent analysis. If
the three eigenvalues are ordered as λ1 ≥ λ2 ≫ λ3 ,
we have a planar coherence anomaly, where the
first and second eigenvectors v1 and v2 represent
Figure 4. Vertical slices through (a) spectral voice components at 15, 25, 40, 55 Hz, as directions parallel to the planar anomaly. In con-
well as the broadband seismic amplitude volumes, and (b) corresponding slices through trast, the third eigenvector v3 represents the direc-
coherence using five traces, 10 ms analysis windows. The vertical resolution is propor- tion perpendicular to the planar anomaly in the
tional to that of the CWT wavelet used to compute each spectral voice. At lower (10 and spherical analysis window. The eigenvectors v1 ,
25 Hz) frequency seismic amplitude voices, the mass-transport deposit zone appears less
chaotic than at higher (40 and 55 Hz) frequency seismic amplitude voices. Coherence v2 , and v3 have three components:
computed from each spectral voice shows additional discontinuities, but it exhibits a lower 8
signal-to-noise ratio than broadband coherence. Faults through the mass-transport deposits c1 v11 þ X
< v1 ¼ X c2 v12 þ X
c3 v13 ;
are more continuous on coherence computed from low-frequency spectral voices. c c c3 v23 ; ;
v ¼ X1 v21 þ X2 v22 þ X (4)
: 2
c1 v31 þ X
v3 ¼ X c2 v32 þ X
c3 v33 ;

c1 , X
where the unit vectors X c2 , and X
c3 are oriented
to north, east, and down. Machado et al. (2016)
and Qi et al. (2017) show the fault dip azimuth
attribute to be ATAN2(v31 , v32 ) and the fault dip
magnitude to be ACOS(v33 ).

Iterative directional smoothing and sharping

Laplacian and Gaussian operators are com-
monly used in filtering photographic images.
The LoG operator smooths short-wavelength arti-
facts of images by the Gaussian operator prior to
Figure 5. A vertical slice through multispectral coherence. Note that faults are more sharpening the images by the Laplacian operator.
continuous and exhibit fewer stair-step artifacts than seen on the broadband coherence Taking into account the hypothesized fault orien-
shown in Figure 3b. tation (the eigenvectors v1 , v2 , and v3 ), the LoG
 Seismic attribute preconditioning O29

operator can directionally smooth parallel to fault surfaces and
sharpen perpendicular to fault surfaces. After the first process, the
output fault probability will be input to the energy-weighted LoG
filtering iteratively until the fault image is sufficiently smoothed and
sharpened. The filter usually stabilizes after three iterations. Because
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
The Taranaki Basin is divided into two main structural com-
ponents, where the Western Platform contains relatively undeformed
sedimentary blocks and the Eastern Mobile Belt is relatively faulted
compared with the Western Platform (Palmer, 1985; King and
Thrasher, 1996). The seismic data lie in the eastern mobile belt
and cover approximately 500 km2 with 50 × 50 m bins. Our chal-

we will wish to sharpen perpendicular to the hypothesized fault

(along eigenvector v3 ) and smooth parallel to the fault (along eigen- lenge is to trace discontinuities through siltstone, silty mudstone, and
vectors v1 and v2 ), we first rotate our natural (east, north, vertical) interbedded sandstone formations through more chaotic data zones,
x-coordinate system, to a ξ-coordinate system aligned with the eigen- including angular unconformities, turbidites, and mass-transport de-
vectors v1 , v2 , and v3 axes: posits, each of which generates its own coherence anomalies that over-
print the fault anomalies of interest. Figure 2 shows a vertical slice
ξ ¼ Rx; (5) through the seismic amplitude volume showing faults that cut the
layered sediments as well as a 500 m thick mass-transport deposit
where R is the rotation matrix given by
zone. This relatively chaotic mass-transport deposit gives rise to
0 1 0 10 1 low-coherence anomalies that mask the through-going faults (Fig-
ξ1 v11 v12 v13 x1 ure 3a). The acquisition footprint, low fold, noise bursts, inappropriate
@ ξ2 A ¼ @ v21 v22 v23 A@ x2 A; (6) migration velocities, processing artifacts, and algorithmic limitation
ξ3 v31 v32 v33 x3 can give rise to additional coherence anomalies that may mask struc-
tural or stratigraphic features of interest (Marfurt and Alves, 2015).
where vij are the directional components in equation 4. In our im- Our workflow starts with footprint suppression. We implement
plementation, we want to smooth more along and less perpendicular Alali et al.’s (2018) workflow, which uses a 2D continuous wavelet
to the faults, so we set transform applied to an edge-detection attribute to design a mask
that is subsequently applied to the amplitude slices to suppress the
1 1 footprint noise. Next, we apply a principal-component structure-ori-
σ 23 < σ 21 ¼ σ 22 ; (7)
3 3 ented filter, which improves the signal-to-noise ratio of the seismic
data and sharpens discontinuities along the structure dip. The com-
where σ 3 is the larger of the two bin dimensions. Finally, we wish to parison of coherence before and after the footprint suppression and
modify the LoG operator to be directional: sharpening along the structure-oriented filtering is shown in Figure 3. We use a 3 × 3 × 3
direction perpendicular to the planar discontinuity
(along the ξ3 axis)
  X M  
d2 1 ξ23 m
G a ¼ − þ
dξ23 m¼1
σ 23 σ 43
  2 
ξ ξ2 ξ2
× exp − 1 m2 þ 2 m2 þ 3 m2 am ; (8)
2σ 1 2σ 2 2σ 3

where G represents the Gaussian operator to be

elongated along the planar axes. The terms σ 1 , σ 2 ,
and σ 3 are standard deviation of the Gaussian
operator. After the first iteration, the output fault
probability will be used as input to the next iter-
ation of energy-weighted LoG filtering until the
fault image is sufficiently smoothed and sharped.
It usually takes three to five iterations to obtain a
reasonable result. As with structure-oriented filter-
ing to improve the signal-to-noise ratio of the am-
plitude data (e.g., Fehmers and Höcher, 2003),
iterative application of smaller windows provides
a computationally more efficient algorithm and
also one that adapts to curved surfaces.

Application to the Taranaki Basin data,

New Zealand
The limitations of coherence in fault definition
The seismic data are located offshore
southeastern Taranaki Basin, which is a Cretaceous
rift basin on the western coast of New Zealand.
Figure 6. Vertical slices through (a) multispectral coherence after a structure-oriented
median filter that suppresses anomalies that cut reflector and retains anomalies parallel
to reflectors and (b) fault enhancement obtained by subtracting the data shown in
(a) from those shown in Figure 5.
 O30 Qi et al.
(inline sample × crossline sample × time sample) rectangular analy-
sis window to compute coherence. After the structure-oriented fil-
tering, lateral and vertical discontinuities are preserved, random and
coherent noise is suppressed, and the signal-to-noise ratio of the
discontinuities is increased. Figure 3c shows a vertical slice through
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
the difference between the original coherence and the coherence
computed after noise reduction. Major faults and stratigraphic dis-
continuities within turbidites exhibit a better signal-to-noise ratio
and higher lateral and vertical resolution. Small faults and subtle
discontinuities in the mass-transport deposits are highlighted after
coherent noise suppression, becoming more easily interpretable.
The filtered seismic data are also used to compute the energy weight
that will be used later.
Figure 7. Vertical slices through LoG fault probability computed from the filtered attrib-
ute shown in Figure 6b, (a) with the energy weight, (b) without the energy weight, and
(c) LoG fault probability computed from the unfiltered attribute. Note that the fault prob-
ability computed from the filtered coherence exhibits fewer isolated noise. Use of the
energy weight results in faults exhibits fewer stair-step artifacts and smoother fault
surfaces.
Coherence of spectral voices and multispectral coherence
It is difficult to trace the major faults through the mass-transport
deposit in the seismic amplitude volume (Figure 2). Coherence
computed from the original noise-reduced data (Figure 3b) show
faults masked by the discontinuities within the mass-transport
deposit zone. Figure 4a shows four spectral voice components
computed using a continuous wavelet transform for 10, 25, 40, and
55 Hz, whereas Figure 4b shows the corresponding coherence re-
sponses. Note that the coherence computed from different spectral
voices exhibits different geologic details. Coherence of the low-fre-
quency spectral voice (10 Hz) only shows the larger discontinuities,
whereas coherence of the high-frequency spectral voice (55 Hz)
exhibits smaller geologic details as well as inco-
herent noise. We observe that faults cutting
through the mass-transport deposit zone are more
continuous and sharper in the low-frequency
spectral voices 10 and 25 Hz. The multispectral
covariance matrix contains geologic details that
are represented by sample vectors. Summing
these sample vectors provide a means of sum-
ming the geologic details. Geologic anomalies
consistent across each spectral voice are rein-
forced in the multispectral covariance matrix.
In contrast, features such as high-frequency noise
presented in only one spectral voice will be
suppressed by this process. Figure 5 shows the
multispectral coherence. The computation of
multispectral coherence involves eight spectral
covariance matrices, which range from 5 to
85 Hz at 10 Hz increments. Because we just used
a 3 × 3 × 3 analysis window in the computation
of multispectral coherence, the highest frequency
containing the coherence image is approximately
83ð1∕3  0.004 sÞ Hz. Note that compared with
the broadband coherence computed from the
structure-oriented filtered amplitude volume
(Figure 3b), the multispectral coherence (Fig-
ure 5) exhibits clearer and more continuous
faults. Incoherent noise along dipping reflectors
and stair-step artifacts are suppressed. Subtle
geologic discontinuities are strengthened and
become more interpretable on the vertical
section.
Attribute image filtering
We apply a 3D structure-oriented median filter
to the multispectral coherence to reject narrow
steeply dipping faults but retain stratigraphic
anomalies (Figure 6a). The size of the analysis
window is 5 × 5 × 0 (inline sample × crossline
sample × time sample), such that the filter is
applied only in the lateral dimension. We then
subtract these values from the coherence image
shown in Figure 5 to obtain an improved fault
image in Figure 6b, where stratigraphic anoma-
lies are suppressed, and vertical variations such
as fault anomalies are preserved.
 Seismic attribute preconditioning O31

Iterative energy-weighted LoG filtered coherence
We then filter the filtered coherence iteratively using our energy-
weighted directional LoG operator. We set σ 1 and σ 2 are equal to
three times the bin size, which are equal to 150 m, and σ 3 is the bin
size, which is equal to 50 m. Figure 7a shows a vertical slice through
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
parallel to the stratigraphy are suppressed. Figure 7b shows the re-
sult after the directional LoG filtering but without an energy weight,
which is close to the original directional LoG filtering proposed
by Machado et al. (2016). We compute these two fault probability
images with an identical input and window size (σ 1 ¼ σ 2 ¼ 150 m;
σ 3 ¼ 50 m). The iteratively energy-weighted LoG filtered fault

the filtered coherence after three iterations of the energy-weighted probability shows better fault resolution and a higher signal-
LoG filtering. Note that faults have greater contrast and are more to-noise ratio. For a zero-phase wavelet, the energy weight favors
continuous, with fewer stair-step artifacts compared with the origi- discontinuities aligned with the reflector. Figure 7c shows the
nal broadband (as well as the multispectral) coherence. Anomalies coherence after the iterative energy-weighted LoG filtering, but

Figure 8. Vertical slices through (a) the seismic amplitude volume corendered with fault probability, and (b) the fault attribute corendered with the
fault dip magnitude and fault dip azimuth. Note that the fault attribute highlights fault planes and subtle discontinuities in the mass-transport
deposit zone on the seismic amplitude volume. The yellow color in (b) indicates the faults dipping to the south, and the blue color indicates the
faults dipping to the north. The red fault dips toward the southeast cuts this north–south vertical slice at 45°, and it thus appears “thicker.”
 O32 Qi et al.

without structure-oriented median filtering. Faults in the mass-
transport deposit zone are less continuous than those in Figure 7a.
We corender our fault probability volume with the seismic ampli-
tude volume in Figure 8a. Figure 8b shows the corendering of the
fault azimuth (against a cyclic color bar), fault dip magnitude
(against monochrome gray and transparency), and fault probability
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
(Figures 9b and 10b). Figures 9d and 10d show time slices through
the corendered images of fault probability, fault dip magnitude, and
fault dip azimuth. Faults and other discontinuities are easily sepa-
rated, and fault planes are also easily tracked by the corendered
images.

(against monochrome black and transparency). In this image, we
see two predominant fault sets — faults dipping to the southeast
(appearing as orange) and faults dipping to the northwest (appearing
as cyan). Next, we examine our results through time slices at
t ¼ 0.8 s (Figure 9) and t ¼ 1.08 s (Figure 10). Red arrows indicate
discontinuities within the mass-transport deposits. The small-scale
discontinuities within the mass-transport deposits are suppressed,
random noise is rejected, whereas larger, through-going faults
are enhanced, allowing the fault planes to be easily tracked
The preconditioned attribute for automatic fault extraction
We validate our fault probability for automatic fault extraction us-
ing a commercial software package. Figure 11a shows a time slice
through the “swarm-intelligence” result with the input of a variance
attribute. The variance attribute is computed using the same 3 × 3 × 3
analysis window as the previous coherence images, whereas the sub-
sequent swarm-intelligence algorithm was run using 5 voxels as
the initial distance, 3 voxels as the increment of each step of a swarm

Figure 9. Time slices at t ¼ 0.8 s through the (a) original coherence, (b) variance, (c) fault probability, and (d) fault probability corendered
with the fault dip magnitude and fault dip azimuth. The workflow shown in Figure 1 suppresses coherence anomalies subparallel to strati-
graphic, and this results in smoother, more continuous faults that cut the stratigraphy. Note that the coherence before our workflow and variance
exhibits a strong acquisition footprint. Red arrows indicate the edges of rotated fault blocks in the mass-transport deposits, which appear as
parallel discontinuities. Blue arrows indicate an acquisition footprint that cannot be suppressed.

Figure 10. Time slices at t ¼ 1.08 s through the (a) original coherence, (b) variance, (c) fault probability, and (d) fault probability corendered
with the fault dip magnitude and fault dip azimuth. Note that the mass-transport deposits mask the through going faults, which are difficult to
interpret on coherence and variance, but are more easily interpreted on the fault probability. Red arrows indicate fault blocks within the mass-
transport deposits.
 Seismic attribute preconditioning
agent, and the threshold of steps of each agent is 10 voxels. Fig-
ure 11b shows the fault patches resulting from automatic fault extrac-
tion of the swarm-intelligence volume. The fault patches computed
from the unfiltered variance volume exhibit strong footprints that are
misinterpreted as a suite of near-vertical parallel planar surfaces. Fig-
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
ure 12a shows the time slice through the fault probability volume,
whereas Figure 13b shows the corresponding fault patches. Major
faults in the fault probability attribute provide superior input to
the automatic fault extraction commercial software. To quantify
Figure 11. The 3D view showing (a) a vertical slice through the
seismic amplitude volume with a time slice through a commercial
swarm-intelligence result with variance as input and (b) the result-
ing extracted fault patches. Note that the swarm-intelligence tech-
nique not only enhances faults, but it also enhances footprint and
incoherent noise.
O33
the size of the fault patches extracted by these two alternative work-
flows, we compute the histograms and display them in Figure 13.
Note that the fault probability workflow results in a larger percentage
of fault patches that exceed 1000 m2 than that provided by swarm
intelligence. These smaller fault patches require considerable manual
intervention. Although some of them may correlate to small faults,
others correlate to noise (such as the acquisition footprint), whereas
Figure 12. The 3D view showing (a) a vertical slice through seismic
amplitude volume with a time slice through the fault probability and
(b) the automatically extracted fault patches with fault probability
as input. Note that faults shown on the fault probability are accu-
rately extracted by the automatic fault extraction technique. The
corresponding fault patches are easier to interpret, and fewer fault
patches need to be modified, which means it is more efficient for
interpretation compared with Figure 11b.
 O34
the majority need to be subsequently edited by the interpreter to form
a larger patch.
Because our objective is to accelerate the fault interpretation
process, it makes sense to use human-interpreted faults as the
ground truth (Figure 14). Note that the fault patches computed from
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
our fault attribute correlate very well with those from traditional
interpretation. Such fault extraction does not replace the human in-
terpreter; rather, the task of the interpreter is similar to that of using
a horizon autotracker — to quality control and edit the computer-
generated surfaces.
Figure 13. Histograms showing percentage of total fault surface
area represented by fault patches of a specific size, using (a) the
fault patches with the fault probability as input and (b) the fault
patches with the swarm intelligence as input. Note that the work-
flow shown in Figure 1 results in a greater amount of larger fault
patches.
Qi et al.
Application to the North Carnarvon Basin, Australia
The second example is located on the Northern Carnarvon Basin
in the northwest shelf of Australia. The seismic data, which were
time migrated with a bin size of 12.5 × 25 m, show a different noise
level compared with the first Taranaki Basin example. Although the
Carnavon Basin survey exhibits less random noise and acquisition
footprint than the older Taranki survey, it exhibits a much greater
amount of stratigraphic discontinuities. Figure 15a shows coherence
computed from the original seismic amplitude volume. Normal
faults with strong stair-step artifacts can be seen in the shallow area
(1.4–2 s). The siltstone and shale sediments (approximately 2.5 s)
show strong incoherent, laterally varied, chaotic anomalies, where
faults are hidden between those layers. Figure 15b shows a time
slice at 2.5 s through the original coherence. Most faults in the silt-
stone and shale layers are difficult to identify. Figure 16a and 16b
Figure 14. The 3D view showing a time slice through a cropped
seismic amplitude volume with (a) interpreter picked faults and
(b) fault patches with the fault attribute as input. Note that the fault
patches are as accurate as the interpreter-picked faults.
 Seismic attribute preconditioning O35

shows the same vertical slice and the time slice but through the fault
probability computed from our workflow. Compared with the origi-
nal coherence (Figure 15), we can easily recognize fault patterns
that cut through the siltstone and shale layers. We also note that
lateral discontinuities and noise are suppressed, and the stair-step
artifacts are smoothed. After our workflow, higher contrast, sharper,
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Given these limitations, our results show that the new fault attrib-
ute exhibits more continuous fault features, less noise, and higher
vertical resolution of fault trends. The preconditioned fault attribute
is validated by inputting into an automatic fault extraction technique
for comparison with other attributes. Although we find that the gen-

erated fault patches within our fault attribute agree closely with the
and smoothed faults are exhibited in the fault probability. human-interpreted fault surface using traditional fault interpretation
on seismic amplitude volumes, the interpreter should still quality
DISCUSSION control the results for incorrectly joined fault segments or picked
artifacts, just as we do when examining autopicked horizons.
In the older Taranaki Basin data volume, footprint suppression Our first example is of approximately 400 Mbytes. We use 20
significantly reduces linear artifacts in the seismic data resulting cores, and the total cost wall clock time was approximately 2.35 h
in a “cleaner” seismic amplitude volume for subsequent edge de- (Figure 17). Structure-oriented filtering and footprint suppression that
tection. The noise level of seismic data defines the smoothness takes 57% of total cost. Because we compute multispectral coherence
and sharpness of structure-oriented filtering. If the data quality is that is an efficient energy-ratio coherence algorithm, it only depends
poor, we would avoid using overlapping Kuwahara windows as de- on the data size, which occupies a small percentage of the space. The
scribed by Davogustto and Marfurt (2011), thereby decreasing the cost of computing iterative energy-weighted LoG filtering is as low as
sharpening and increasing the smoothing. If the
data quality is good, we use aggressive structure-
oriented filtering with full Kuwahara subwind-
ows to generate a sharper fault image. For this
data volume, it appears that multispectral coher-
ence differentiates fault discontinuities that appear
at multiple frequencies corresponding to through-
going faults from those that appear at only the
higher frequencies corresponding to local dis-
continuities within the mass-transport deposit.
The use of a median filter to suppress coherence
anomalies parallel to the reflector dip is geology
dependent. In the first example, the faults dip
between 45° and 90° and will not be adversely af-
fected. However, in a survey characterized by lis-
tric faults, such filtering would incorrectly remove
the base of the fault as it soles out into softer
shales.
Discontinuities in reflectors are convolved
with the source wavelet such that this discontinu-
ity is continued perpendicular to the reflector, re-
sulting in stair-step anomalies. Such stair-step
anomalies often result in a lateral shift of coher-
ence anomalies seen on time slices compared
with a human interpreted fault on a vertical slice
through the seismic amplitude. Weighting the
second-order moment tensor by energy favors
that part of the wavelet al.gned with the reflector
discontinuity in the subsequent fault probability
workflow, thereby minimizing the stair-step arti-
facts. Although we believe that most stair-steps
features seem on 3D coherence volumes to be
artifacts, the reader should be forewarned that
Ferrill et al. (2016a) document such stair-step
features in outcrop.
A pitfall in fault enhancement is to be overly
aggressive and fill the gap between vertical fault
segments that in reality should be disjoint, or to
overly smooth faults that “refract” in response to
variable rock strength. We refer the reader to a
second paper by Ferrill et al. (2016b) that dispar- Figure 15. (a) Vertical slice and (b) time slice through the coherence computed from the
ages some of the myths associated with normal original seismic amplitude volume on the Northern Carnarvon Basin. Note that the silt-
faulting. stone and shale layers exhibit highly incoherent and laterally varied noise.
 O36 Qi et al.

computing regular image LoG filtering, which

is also approximated by a Gaussian filtering.

CONCLUSION
We have introduced a seismic attribute precon-
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ditioning workflow for automatic fault extrac-

tion. Our objective was to suppress coherence
anomaly artifacts, suppress coherence (noise or
stratigraphic) anomalies subparallel to reflector
dip, and sharpen fault and other discontinuities
that cut reflectors, preconditioning the fault
images for subsequent automatic extraction. To
avoid interpreting acquisition footprint as arti-
facts, we suppressed these wavenumber compo-
nents using a 2D continuous wavelet transform
applied time slice by time slice. To further sup-
press crosscutting noise as well as sharpen fault
edges, we applied a principal component edge-
preserving structure-oriented filter. The condi-
tioned amplitude volume served as input to an
eight-component multispectral coherence algo-
rithm, which enhanced through-going faults seen
at most frequencies and suppressed local discon-
tinuities seen at only a few frequencies.
Coherence algorithms show not only tectonic
and stratigraphic edges that cut reflectors but also
unconformities, condensed sections, and low sig-
nal-to-noise ratio shale-on-shale reflectors that are
subparallel to reflectors. We apply a median filter
to the coherence volume enhance, and we then
subtract anomalies parallel to the reflector dip,
leaving discontinuities that correlate to fault and
stratigraphic edges in the residual. Finally, we in-
put this residual through three iterations of an en-
ergy-weighted directional LoG filter to sharpen
and smooth the faults. Energy weighting reduces
Figure 16. The same slices through the fault probability computed from our workflow. stair-step fault anomalies commonly seen in co-
Note that the faults are shaper and more continuous between the siltstone and shale herence volumes. Using an iterative application
layers compared with Figure 15. of small-window LOG filters accommodates cur-
vilinear fault surfaces and avoids joining faults
across large gaps. The results after this attribute preconditioning
workflow are improved fault images amenable to automatic fault ex-
traction. We find the resulting fault images to compare very favorably
with respect to traditional human-interpreter generated on vertical sli-
ces through the seismic amplitude volume.

ACKNOWLEDGMENTS
We would like to thank New Zealand Petroleum and Minerals for
providing the data to be used in research. We would also like to thank
Schlumberger for the license of Petrel provided to the OU for re-
search and education. Finally, we thank the sponsors of the OU
Attribute-Assisted Processing and Interpretation Consortium for their
guidance and their financial support.

DATA AND MATERIALS AVAILABILITY

Figure 17. Diagram of the computation cost for the whole work- Data associated with this research are available and can be
flow. obtained by contacting the corresponding author.
 Seismic attribute preconditioning
REFERENCES
Aarre, V., and B. Wallet, 2011, A robust and compute-efficient variant of the
Radon transform: Proceedings of the 31st Annual GCSSEPM Foundation
Bob F. Perkins Research Conference, 550–586.
Admasu, F., S. Back, and K. Toennies, 2006, Autotracking of faults on 3D
seismic data: Geophysics, 71, no. 6, A49–A53, doi: 10.1190/1.2358399.
Downloaded 05/08/19 to 116.254.121.50. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Alaei, B., 2012, Improved fault imaging by integration of frequency decom-
position and fault attributes, example from Mid Norwegian Sea: 74th An-
nual International Conference and Exhibition, EAGE, Extended Abstracts,
doi: 10.3997/2214-4609.20143024.
Alali, A., G. Machado, and K. J. Marfurt, 2018, Attribute-assisted footprint
suppression using a 2D continuous wavelet transform: Interpretation, 6,
no. 2, T457–T470, doi: 10.1190/INT-2017-0175.1.
Alaudah, Y., and G. AlRegib, 2017, A directional coherence attribute for seis-
mic interpretation: 87th Annual International Meeting, SEG, Expanded Ab-
stracts, 2102–2105, doi: 10.1190/segam2017-17739097.1.
AlBinHassan, N. M., and K. J. Marfurt, 2003, Fault detection using Hough
transforms: 73rd Annual International Meeting, SEG, Expanded Abstracts,
1719–1721, doi: 10.1190/1.1817639.
Aqrawi, A. A., and T. H. Boe, 2011, Improved fault segmentation using a dip
guided and modified 3D Sobel filter: 81st Annual International Meeting,
SEG, Expanded Abstracts, 999–1003, doi: 10.1190/1.3628241.
Bahorich, M. S., and S. L. Farmer, 1995, 3-D seismic discontinuity for faults
and stratigraphic features: The Leading Edge, 14, 1053–1058, doi: 10
.1190/1.1437077.
Bakker, P., 2002, Image structure analysis for seismic interpretation: Ph.D.
thesis, Delft University of Technology.
Barnes, A. E., 2006, A filter to improve seismic discontinuity data for fault
interpretation: Geophysics, 71, no. 3, P1–P4, doi: 10.1190/1.2195988.
Boe, T. H., 2012, Enhancement of large faults with a windowed 3D Radon
transform filter: 82nd Annual International Meeting, SEG, Expanded Ab-
stracts, doi: 10.1190/segam2012-1008.1.
Chopra, S., and K. J. Marfurt, 2007, Seismic attributes for prospect identi-
fication and reservoir characterization: SEG.
Cohen, I., and R. R. Coifman, 2002, Local discontinuity measures for 3-D
seismic data: Geophysics, 67, 1933–1945, doi: 10.1190/1.1527094.
Cohen, I., N. Coult, and A. A. Vassiliou, 2006, Detection and extraction of
fault surfaces in 3D seismic data: Geophysics, 71, no. 4, P21–P27, doi: 10
.1190/1.2215357.
Davogustto, O., and K. J. Marfurt, 2011, Footprint suppression applied to leg-
acy seismic data volumes, in K. J. Marfurt, D. Gao, A. Barnes, S. Chopra,
A. Corrao, B. Hart, H. James, J. Pacht, and N. C. Rosen, eds., Attributes:
New views on seismic imaging — Their use in exploration and production,
31st Annual GCSSEPM, 1–34.
Dewett, D. T., and A. A. Henza, 2016, Spectral similarity fault enhancement:
Interpretation, 4, no. 1, SB149–SB159, doi: 10.1190/INT-2015-0114.1.
Dorn, G. A., B. Kadlec, and P. Murtha, 2012, Imaging faults in 3D seismic
volumes: 82nd Annual International Meeting, SEG, Expanded Abstracts,
doi: 10.1190/segam2012-1538.1.
Fehmers, G, and C. F. W. Höcher, 2003, Fast structural interpretation with
structure-oriented filtering: Geophysics, 68, 1286–1293, doi: 10.1190/1
.1598121.
Ferrill, D. A., A. P. Morris, R. N. McGinnis, and K. J. Smart, 2016b, Myths
about normal faulting, in C. Childs, R. E. Holdsworth, C. A.-L. Jackson, T.
Manzocchi, J. J. Walsh, and G. Yielding, eds., The geometry and growth of
normal faults: Geological Society of London, Special Publications, 41–56.
Ferrill, D. A., A. P. Morris, R. N. McGinnis, K. J. Smart, S. S. Wigginton,
and N. J. Hill, 2016a, Mechanical stratigraphy and normal faulting: Jour-
nal of Structural Geology, 94, 275–302, doi: 10.1016/j.jsg.2016.11.010.
Gersztenkorn, A., and K. J. Marfurt, 1999, Eigenstructure based coherence
computations as an aid to 3D structural and stratigraphic mapping: Geo-
physics, 64, 1468–1479, doi: 10.1190/1.1444651.
Hale, D., 2013, Methods to compute fault images, extract fault surfaces, and
estimate fault throws from 3D seismic images: Geophysics, 78, no. 2,
O33–O43, doi: 10.1190/geo2012-0331.1.
Karimi, P., S. Fomel, L. Wood, and D. Dunlap, 2015, Predictive coherence:
Interpretation, 3, no. 4, SAE1–SAE7, doi: 10.1190/INT-2015-0030.1.
O37
King, P. R., and G. P. Thrasher, 1996, Cretaceous Cenozoic geology and
petroleum systems of the Taranaki Basin, New Zealand: Institute of Geo-
logical and Nuclear Sciences, 2.
Li, F., and W. Lu, 2014, Coherence attribute at different spectral scales:
Interpretation, 2, no. 1, SA99–SA106, doi: 10.1190/INT-2013-0089.1.
Li, F., J. Qi, B. Lyu, and K. J. Marfurt, 2018, Multispectral coherence:
Interpretation, 6, no. 1, T61–T69, doi: 10.1190/INT-2017-0112.1.
Li, F., J. Qi, and K. J. Marfurt, 2015, Attribute mapping of variable thickness
incised valley-fill systems: The Leading Edge, 34, 48–52, doi: 10.1190/
tle34010048.1.
Libak, A., B. Alaei, and A. Torabi, 2017, Fault visualization and identifi-
cation in fault seismic attribute volumes: Implications for fault geometric
characterization: Interpretation, 5, no. 2, B1–B16, doi: 10.1190/INT-2016-
0152.1.
Lin, T., T. Ha, K. J. Marfurt, and K. L. Deal, 2016, Quantifying the signifi-
cance of coherence anomalies: Interpretation, 4, no. 2, T205–T213, doi:
10.1190/INT-2015-0102.1.
Machado, G., A. Alali, B. Hutchinson, O. Olorunsola, and K. J. Marfurt,
2016, Display and enhancement of volumetric fault images: Interpreta-
tion, 4, no. 1, SB51–SB61, doi: 10.1190/INT-2015-0104.1.
Marfurt, K. J., 2017, Interpretational value of multispectral coherence:
79th Annual International Conference and Exhibition, EAGE, Extended
Abstracts, doi: 10.3997/2214-4609.201700528.
Marfurt, K. J., and T. M. Alves, 2015, Pitfalls and limitations in seismic attrib-
ute interpretation of tectonic features: Interpretation, 3, no. 1, A5–A15, doi:
10.1190/INT-2014-0122.1.
Marfurt, K. J., R. L. Kirlin, S. H. Farmer, and M. S. Bahorich, 1998, 3D
seismic attributes using a running window semblance-based algorithm:
Geophysics, 63, 1150–1165, doi: 10.1190/1.1444415.
Palmer, J., 1985, Pre-miocene Lithostratigraphy of Taranaki Basin, New
Zealand: New Zealand Journal of Geology and Geophysics, 28, 197–216,
doi: 10.1080/00288306.1985.10422220.
Pedersen, S., T. Randen, L. Sonneland, and O. Steen, 2002, Automatic 3D
fault interpretation by artificial ants: 72nd Annual International Meeting,
SEG, Expanded Abstracts, 512–515, doi: 10.1190/1.1817297.
Qi, J., F. Li, B. Lyu, O. Olorunsola, K. J. Marfurt, and B. Zhang, 2016a, Seis-
mic fault enhancement and skeletonization: 86th Annual International Meet-
ing, SEG, Expanded Abstracts, 1966–1970, doi: 10.1190/segam2016-
13876567.1.
Qi, J., F. Li, and K. J. Marfurt, 2017a, Multiazimuth coherence: Geophysics,
82, no. 6, O83–O89, doi: 10.1190/geo2017-0196.1.
Qi, J., T. Lin, T. Zhao, F. Li, and K. J. Marfurt, 2016b, Semisupervised multi-
attribute seismic facies analysis: Interpretation, 4, no. 1, SB91–SB106,
doi: 10.1190/INT-2015-0098.1.
Qi, J., G. Machado, and K. J. Marfurt, 2017b, A workflow to skeletonize
faults and stratigraphic features: Geophysics, 82, no. 4, O57–O70, doi:
10.1190/geo2016-0641.1.
Qi, J., B. Zhang, H. Zhou, and K. J. Marfurt, 2014, Attribute expression of
fault-controlled karst–Fort Worth Basin, TX: Interpretation, 2, no. 3,
SF91–SF110, doi: 10.1190/INT-2013-0188.1.
Randen, T., S. Pedersen, and L. Sønneland, 2001, Automatic extraction of fault
surfaces from three-dimensional seismic data: 71st Annual International
Meeting, SEG, Expanded Abstracts, 551–554, doi: 10.1190/1.1816675.
Van Bemmel, P. P., and R. E. Pepper, 2000, Seismic signal processing
method and apparatus for generating a cube of variance values: U.S. Pat-
ent 6,151,555.
Wu, X., 2017, Directional structure-tensor-based coherence to detect seismic
faults and channels: Geophysics, 82, no. 2, A13–A17, doi: 10.1190/
geo2016-0473.1.
Wu, X., and S. Fomel, 2018, Automatic fault interpretation with optimal sur-
face voting: Geophysics, 83, no. 5, O67–O82, doi: 10.1190/geo2018-0115
.1.
Wu, X., and D. Hale, 2015, 3D seismic image processing for faults: Geo-
physics, 81, no. 2, IM1–IM11, doi: 10.1190/geo2015-0380.1.
Zhang, B., Y. Liu, M. Pelissier, and N. Hemstra, 2014, Semiautomated fault
interpretation based on seismic attributes: Interpretation, 2, no. 1, SA11–
SA19, doi: 10.1190/INT-2013-0060.1.