Seismic fault detection based on multi-attribute support vector machine analysis Haibin Di*, Muhammad A. Shafiq, and Ghassan AlRegib

Center for Energy and Geo Processing (CeGP), Georgia Institute of Technology

Summary

Reliable fault detection is one of the major tasks of

subsurface interpretation and reservoir characterization from

three-dimensional (3D) seismic surveying. This study

presents an innovative workflow based on multi-attribute

support vector machine (SVM) analysis of a seismic volume,

which consists of four steps. First, three groups of seismic

attributes are selected and computed from the volume of

seismic amplitude, including edge-detection, geometric, and

texture, all of which clearly highlight the seismic faults in

the attribute images. Second, two sets of training samples are

prepared by manually picking on the faults and the non-

faulting zones, respectively. Third, the SVM analysis is

performed on the training datasets that builds an optimal

classification model for volumetric processing. Finally,

applying the SVM model to the whole seismic survey leads

to a binary volume, in which the presence of a fault is

labelled as ones. The added values of the proposed method

are verified through applications to the seismic dataset over

the Great South Basin in New Zealand, where the dominant

features are polygonal faults of varying sizes and

orientations. The results demonstrate not only good match

between the detected faults and the original seismic images,

but also great potential for quantitative fault interpretation,

such as semi-automatic/automatic fault extraction, to aid

structural framework modeling and reservoir simulation in

the exploration areas of numerous faults and fractures

Introduction

Faults and fractures are often of important geologic

implications for investigating hydrocarbon accumulation

and migration in a petroleum reservoir in the subsurface, and

the presence of a fault can be visually recognized as a

lineament of abrupt changes in reflection patterns from

three-dimensional (3D) reflection seismic data. However,

quantitative fault interpretation, such as patch extraction, is

a time-consuming and lab-intensive process and remains as

a challenging topic, especially for an exploration area

featured with numerous faults. In the past decades, great

efforts have been devoted into developing new attributes and

methods/algorithms to help improve the accuracy of seismic

fault detection.

From the perspective of seismic attribute analysis, both the

edge-detection and geometric attributes are applicable for

fault detection, due to the apparent lateral variation of

seismic waveform and/or amplitude across a fault. The

seismic edge-detection analysis was first presented as the

coherence attribute for highlighting the faults and

stratigraphic features from a seismic cube (Bahorich and

Farmer, 1995), and since then, such attribute and its

derivatives has been improved for better detection resolution

and noise robustness (e.g., Luo et al., 1996; Marfurt et al.,

1998; Tingdahl and de Rooij, 2005; Al-Dossary et al, 2014;

Di and Gao, 2014; Wang et al., 2016). Then for more robust

fault detection and fracture characterization from 3D seismic

data, the seismic geometric attributes are developed by

quantifying the lateral variations of the geometry of seismic

reflectors, including the second-order curvature (e.g., Lisle,

1994; Roberts, 2001; Al-Dossary and Marfurt, 2006), and

the third-order flexure attributes (e.g., Gao, 2013; Yu, 2014;

Di and Gao, 2017a). Meanwhile, computer-aided extraction

of fault patches has been the research focus since 1990s. For

example, Meldahl et al. (1999) presented a semi-automatic

approach for detecting seismic chimneys by combining

directive attributes and neural network, and such approach

was later adapted to fault extraction from 3D seismic data by

Tingdahl and de Rooij (2005). Gibson et al. (2005) and

Zhang et al. (2014) proposed grouping fault points into the

local planar patches and then merging these small patches

into larger fault surfaces under certain geometric constraints.

Hale (2013) proposed a dynamic time warping algorithm to

generate fault surfaces based on the boundary constraints

derived from the thinned discontinuity images. Wang and

AlRegib (2014) introduced the ideas of motion vectors in

video coding and processing to extract fault surfaces.

Unfortunately, all these algorithms are unable to

simultaneously achieve both high resolution in extracting

true faults and high accuracy in avoiding non-fault artifacts.

The common result is either an aggressive case with high

resolution (most/all true faults extracted) but low accuracy

(many artifacts introduced), or a conservative one with high

accuracy (few artifacts introduced) but low resolution (few

true faults extracted) (Di and Gao, 2017b). Therefore, semi-

automatic/automatic fault extraction is still in the

experimental phase for testing and not ready for practical

implementation and application to industrial projects.

For resolving such limitation, this paper first proposes a new

fault-detection workflow based on semi-supervised multi-

attribute support vector machines (SVM) analysis, and then

applies it to the 3D seismic dataset over the Great South

Basin (GSB) in New Zealand, where polygonal faults of

varying sizes and orientations are observed in the

subsurface.

Algorithm description

This study adapts the binary SVM classification to work for

multi-attribute seismic fault detection, and the proposed

© 2017 SEG SEG International Exposition and 87th Annual Meeting

Page 2039

Dow

nloa

ded

08/2

2/17

to 1

43.2

15.1

48.5

7. R

edis

trib

utio

n su

bjec

t to

SEG

lice

nse

or c

opyr

ight

; see

Ter

ms

of U

se a

t http

://lib

rary

.seg

.org

/

SVM-based fault detection

method consists of four steps. First, three groups of seismic

attributes are selected and calculated from a seismic dataset,

which expands the seismic domain into a multi-dimensional

attribute domain and represents every sample with a vector.

Second, two sets of training data are manually picked at both

the faults and the non-faulting areas. Third, a SVM classifier

is trained by the picked training datasets. Finally, the created

classifier is applied to the whole seismic volume, and a

binary volume is generated with the potentials faults labelled

as ones.

For the convenience of illustrating the proposed workflow,

we expand each step in details, using the 3D seismic dataset

over the Great South Basin (GSB) in New Zealand, where

the subsurface geology is dominated by polygonal faulting

in varying sizes and orientations. Figure 1 displays the time

section of original seismic amplitude at 1152 ms, in which a

set of faults are clearly imaged.

A. Attribute selection

For efficient fault detection, the selected attributes are

expected capable of enhancing the faults to be more

distinguishable from the non-fault features (such as

horizons) in the attribute images. Among all possible seismic

attributes, we select and generate fourteen attributes from the

amplitude volume, which are categorized into three groups,

Geometric attribute. It measures the lateral variation of

the geometry of seismic reflectors, including the

reflector dip, the curvature (Roberts, 2001), the flexure

(Di and Gao, 2017a), and the geometric fault (Di and

AlRegib, 2017).

Edge-detection attribute. It evaluates the lateral

changes in seismic waveform and/or amplitude using

various edge-detection operators, including the

coherence (Bahorich and Farmer, 1995), the Sobel filter

(Luo et al., 1996), the semblance (Marfurt et al., 1998),

the Canny edge (Di and Gao, 2014), the similarity, and

the variance (Pedersen et al., 2002).

Texture attribute. It describes the local distribution of

seismic texture using various statistical operators,

including the GLCM contrast and homogeneity (Gao,

2003; Eichkitz et al., 2013), gradient of texture (GoT)

(Shafiq et al., 2017a), and seismic saliency (Shafiq et

al., 2017b).

After extracting all fourteen attributes, every sample in the

seismic volume is represented by a vector of 14 elements,

which allows classifying the samples in the 14-dimensional

attribute domain and estimating a 13-dimensional

hyperplane that well separates these samples into two

classes. To be clear, all the selected attributes are

normalized, due to their different units of measurement and

moreover distinct histograms.

B. Training sample labelling

To ensure accurate SVM classification, two sets of training

samples are manually picked on the faults (171 picks) and

the non-faulting zones (722 picks) (Figure 2), which are

further used in two ways. First, cross plotting of the selected

attributes over the manual pickings could be used for

verifying the step of attribute selection, in which the fault

pickings and the non-fault pickings are partitioned into two

groups with a border between them. Second, the manual

pickings can help check the accuracy of the training SVM

Figure 2: The manual pickings on the vertical section of

crossline 2800, including 171 pickings on the faults (denoted

as cyan dots) and 722 pickings on the surrounding non-

faulting features (denoted as magenta dots).

Figure 1: The time slice of original seismic amplitude at

1152 ms over the Great South Basin (GSB) in New Zealand,

where the dominant structural features are polygonal

faulting in varying sizes and orientations.

© 2017 SEG SEG International Exposition and 87th Annual Meeting

Page 2040

Dow

nloa

ded

08/2

2/17

to 1

43.2

15.1

48.5

7. R

edis

trib

utio

n su

bjec

t to

SEG

lice

nse

or c

opyr

ight

; see

Ter

ms

of U

se a

t http

://lib

rary

.seg

.org

/

SVM-based fault detection

model. Ideally, the built model should be capable of

labelling these pickings accurately; however, due to the mis-

pickings of the seeds and/or the noises present in the

attributes, errors are often observed.

C. SVM model training and testing

Based on the labelled training datasets, the SVM model is

built for connecting the selected fourteen attributes and the

fault labelling. Before applying for volumetric processing,

the model is first verified through the training datasets.

Figure 3 displays the classification of the 893 pickings.

Compared to the training samples (Figure 2), we notice that

the majority of the pickings are labelled corrected, except

those on the zones where the seismic expressions are similar

to those on the faults (denoted by circles).

D. Volumetric processing

After verifying the accuracy of the built SVM classification

model, it is applied to the whole seismic survey for

volumetric processing, generating a binary volume with fault

labelled as ones.

Results and applications

We apply the proposed multi-attribute SVM classification to

the GSB seismic dataset, and for the convenience of result

interpretation, we then adjust the opacity of the binary fault

volume and overlay them on the original seismic images.

The results are displayed in different ways, including the 3D

view (Figure 4), the corresponding time slice at 1152 ms

(Figure 5), and four randomly-selected vertical sections

(Figure 6), all of which demonstrate good match between the

detected faults and the original seismic images, indicating

the accuracy of the proposed SVM fault classification.

Besides the qualitative mapping of the faults from a seismic

dataset, the binary volume could also serve as input to more

advanced quantitative interpretation. For example, seimi-

automatic/automatic fault extraction can be performed on

the fault volume, and Figure 7 displays the extracted fault

patches from the binary volume, clipped to the time slice at

1152 ms.

Figure 4: 3D view of the fault detection from the Great

South Basin (GSB) in New Zealand, overlaying the original

seismic amplitude. Note the good match between the

detected faults and the seismic images.

Figure 3: The re-classification of the 893 pickings by the

built SVM model, overlaying the vertical section of crossline

2800, with the cyan dots representing the faults and the

magenta dots representing the non-faulting features.

Figure 5: The time slice of the detected faults at 1152 ms by

the proposed multi-attribute SVM classification. Note the

good match between the polygonal faults and the original

seismic image.

© 2017 SEG SEG International Exposition and 87th Annual Meeting

Page 2041

Dow

nloa

ded

08/2

2/17

to 1

43.2

15.1

48.5

7. R

edis

trib

utio

n su

bjec

t to

SEG

lice

nse

or c

opyr

ight

; see

Ter

ms

of U

se a

t http

://lib

rary

.seg

.org

/

SVM-based fault detection

Conclusions

Reliable detection of subsurface faults and fractures from 3D

seismic data is essential for reservoir characterization and

modeling. This study has presented a new workflow for

seismic fault detection based on multi-attribute support

vector machine (SVM) classification, which consists of three

major components. First, three attribute groups, including

the edge-detection attributes, geometric attributes, and the

texture attributes, are generated from the amplitude volume

that help differentiate the faults of interpretational interest

from the surrounding non-faulting features. Second, two sets

of manual pickings are used for training a SVM model in the

attribute domain and building an optimal model for

volumetric processing. Finally, applying the built

classification model to the whole seismic survey provides a

binary volume, which not only clearly delineates the faults,

but also holds the potential for more advanced fault

interpretation, such as semi-automatic/automatic fault

extraction.

Such multi-attribute seismic feature classification could be

further improved by incorporating more fault attributes,

increasing the amount of training samples, and applying

more advanced machine learning algorithms (e.g., CNN).

However, it might require more time and efforts for data

preparation and computation.

Acknowledgments

This work is supported by the Center for Energy and Geo

Processing at Georgia Tech and King Fahd University of

Petroleum and Minerals. We thank the New Zealand

Petroleum and Minerals (NZP&M) for providing the 3D

seismic dataset over the Great South Basin (GSB). The SVM

classification algorithm is provided by the Turi GraphLab

CreateTM under an academic license.

Figure 6: Four randomly-selected vertical sections of the

fault detection by the proposed multi-attribute SVM

classification. Note the good match between the detected

faults and the original seismic image.

Figure 7: The fault patches extracted from the binary fault

volume by the proposed multi-attribute SVM classification,

clipped to the time slice at 1152 ms.

© 2017 SEG SEG International Exposition and 87th Annual Meeting

Page 2042

Dow

nloa

ded

08/2

2/17

to 1

43.2

15.1

48.5

7. R

edis

trib

utio

n su

bjec

t to

SEG

lice

nse

or c

opyr

ight

; see

Ter

ms

of U

se a

t http

://lib

rary

.seg

.org

/

EDITED REFERENCES

Note: This reference list is a copyedited version of the reference list submitted by the author. Reference lists for the 2017

SEG Technical Program Expanded Abstracts have been copyedited so that references provided with the online

metadata for each paper will achieve a high degree of linking to cited sources that appear on the Web.

REFERENCES

Al-Dossary, S., and K. J. Marfurt, 2006, 3D volumetric multispectral estimates of reflector curvature and

rotation: Geophysics, 71, no. 5, P41–P51, http://dx.doi.org/10.1190/1.2242449.

Al-Dossary, S., Y. E. Wang, and M. McFarlane, 2014, Estimating randomness using seismic disorder:

Interpretation, 2, no. 1, SA93–SA97, http://dx.doi.org/10.1190/INT-2013-0088.1.

Bahorich, M., and S. Farmer, 1995, 3-D seismic coherency for faults and stratigraphic features: The

Leading Edge, 14, 1053–1058, http://dx.doi.org/10.1190/1.1437077.

Di, H., and G. AlRegib, 2017, Volumetric fault imaging based on seismic geometry analysis: AAPG

Annual Convention and Exhibition.

Di, H., and D. Gao, 2014, Gray-level transformation and Canny edge detection for 3D seismic

discontinuity enhancement: Computers & Geosciences, 72, 192–200,

http://dx.doi.org/10.1016/j.cageo.2014.07.011.

Di, H., and D. Gao, 2017a, 3D seismic flexure analysis for subsurface fault detection and fracture

characterization: Pure and Applied Geophysics, 174, 747–761, http://dx.doi.org/10.1007/s00024-

016-1406-9.

Di, H., and D. Gao, 2017b, Seismic attribute-aided fault detection in petroleum industry: A review, in D.

Martion, ed., Fault detection: Methods, Applications and Technology: Nova Science Publishers,

53–80.

Eichkitz, C. G., J. Amtmann, and M. G. Schreilechner, 2013, Calculation of gray level co-occurrence

matrix-based seismic attributes in three dimensions: Computers & Geosciences, 60, 176–183,

http://dx.doi.org/10.1016/j.cageo.2013.07.006.

Gao, D., 2003, Volume texture extraction for 3D seismic visualization and interpretation: Geophysics, 68,

1294–1302, http://dx.doi.org/10.1190/1.1598122.

Gao, D., 2013, Integrating 3D seismic curvature and curvature gradient attributes for fracture detection:

Methodologies and Interpretational implications: Geophysics, 78, no. 2, O21–O38,

http://dx.doi.org/10.1190/geo2012-0190.1.

Gibson, D., M. Spann, J. Turner, and T. Wright, 2005, Fault surface detection in 3D seismic data: IEEE

Transactions on Geoscience and Remote Sensing, 43, 2094–2102,

http://dx.doi.org/10.1109/TGRS.2005.852769.

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, http://dx.doi.org/10.1190/geo2012-0331.1.

Lisle, R. J., 1994, Detection of zones of abnormal strains in structures using Gaussian curvature analysis:

AAPG Bulletin, 78, 1811–1819, http://dx.doi.org/10.1306/A25FF305-171B-11D7-

8645000102C1865D.

Luo, Y., W. G. Higgs, and W. S. Kowalik, 1996, Edge detection and stratigraphic analysis using 3-D

seismic data: 66th Annual International Meeting, SEG, Expanded Abstracts, 324–327,

https://doi.org/10.1190/1.1826632.

Marfurt, K. J., R. L. Kirlin, S. L. Farmer, and M. S. Bahorich, 1998, 3-D seismic attributes using a

semblance-based coherency algorithm: Geophysics, 63, 1150–1165,

http://dx.doi.org/10.1190/1.1444415.

Meldahl, P., R. Heggland, A. H. Bril, and P. F. M. de Groot, 1999, The chimney cube, an example of

semi-automated detection of seismic objects by directive attributes and neural networks - Part 1:

© 2017 SEG SEG International Exposition and 87th Annual Meeting

Page 2043

Dow

nloa

ded

08/2

2/17

to 1

43.2

15.1

48.5

7. R

edis

trib

utio

n su

bjec

t to

SEG

lice

nse

or c

opyr

ight

; see

Ter

ms

of U

se a

t http

://lib

rary

.seg

.org

/

Methodology: 69th Annual International Meeting, SEG, Expanded Abstracts, 931–934,

https://doi.org/10.1190/1.1821262.

Pedersen, S. I., T. Randen, L. Sonneland, and O. Steen, 2002, Automatic fault extraction using artificial

ants: 72nd Annual International Meeting, SEG, Expanded Abstracts, 512–515,

http://dx.doi.org/10.1190/1.1817297.

Roberts, A., 2001, Curvature attributes and their application to 3D interpreted horizons: First break, 19,

85–100, http://dx.doi.org/10.1046/j.0263-5046.2001.00142.x.

Shafiq, M. A., T. Alshawi, Z. Long, and G. AlRegib, 2017a, The role of visual saliency in the automation

of seismic interpretation: Accepted in Geophysical Prospecting.

Shafiq, M. A., Z. Wang, G. AlRegib, A. Amin, and M. Deriche, 2017b, A texture-based interpretation

workflow with application to delineating salt domes: Interpretation, 5, no. 3, SJ1-SJ19,

http://dx.doi.org/10.1190/INT-2016-0043.1.

Tingdahl, K. M., and M. de Rooij, 2005, Semi-automatic detection of faults in 3D seismic data:

Geophysical Prospecting, 53, 533–542, http://dx.doi.org/10.1111/j.1365-2478.2005.00489.x.

Wang, S., S. Yuan, B. Yan, Y. He, and W. Sun, 2016, Directional complex-valued coherence attribute for

discontinuous edge detection: Journal of Applied Geophysics, 129, 1–7,

http://dx.doi.org/10.1016/j.jappgeo.2016.03.016.

Wang, Z., Z. Long, G. AlRegib, A. Asjad, and M. A. Deriche, 2014, Automatic fault tracking across

seismic volumes via tracking vectors: IEEE International Conference on Image Processing

(ICIP), 5851–5855, https://doi.org/10.1109/icip.2014.7026182.

Wang, Z., T. Hegazy, Z. Long, G. AlRegib, 2015, Noise-robust detection and tracking of salt domes in

postmigrated volumes using texture, tensors, and subspace learning: Geophysics, 80, no. 6,

WD101–WD116, http://dx.doi.org/10.1190/geo2015-0116.1.

Yu, J., 2014, Using cylindrical surface-based curvature change rate to detect faults and fractures:

Geophysics, 79, no. 5, O1–O9, http://dx.doi.org/10.1190/geo2014-0003.1.

Zhang, B., Y. Liu, M. Pelissier, and N. Hemstra, 2014, Semiautomated fault interpretation based on

seismic attributes: Interpretation, 2, no. 1, SA11-SA19, http://dx.doi.org/10.1190/INT-2013-

0060.1.

© 2017 SEG SEG International Exposition and 87th Annual Meeting

Page 2044

Dow

nloa

ded

08/2

2/17

to 1

43.2

15.1

48.5

7. R

edis

trib

utio

n su

bjec

t to

SEG

lice

nse

or c

opyr

ight

; see

Ter

ms

of U

se a

t http

://lib

rary

.seg

.org

/