352 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 16, NO.

3, MARCH 2019

Seismic Fault Detection Using Convolutional

Neural Networks Trained on Synthetic
Poststacked Amplitude Maps
Axelle Pochet , Pedro H. B. Diniz, Hélio Lopes , and Marcelo Gattass

Abstract— Fault detection is a crucial step in reservoir charac- putation, and, alone, are not suited for efficient fault identifica-
terization. Despite the many tools developed in the past decades, tion: a human interpreter must spend time finalizing the study
automation of this task remains a challenge. We investigate the manually. Consequently, many authors propose to postprocess
application of convolutional neural networks (CNNs) to seismic the attribute maps to extract fault location automatically. For
fault detection. CNN is a deep learning method growing in
interest in the computer vision community, due to its high example, Gibson et al. [12] use semblance to create a set of
performances in a great variety of object detection tasks. One of high faultiness points that are joined to build fault surfaces;
the constraints of this method is the need to provide a massive Zhang et al. [13] apply a skeletonization on the coherence
number of interpreted data, a requirement particularly difficult cube to extract fault sticks; and Wang and AlRegib [14], [28]
to attend in the seismic area. To this end, we built a synthetic data use the Hough Transform to extract fault locations from
set with simple fault geometries. The input of our network is the binarized continuity maps.
seismic amplitude only; the method does not require computing
any seismic attribute. We apply a strategy of patch classification
Another way to use attribute information is to combine
along the images, which requires a simple postprocess to extract them. Machine learning algorithms are particularly suited for
the exact fault location. Our network shows good results on this task, as they can efficiently find relationships between
synthetic data and encouraging results when tested on regions of a set of input features (seismic attributes) and the desired
a real section of The Netherland offshore F3 block in the North output value (fault location). The work of [15] uses a neural
Sea. network to combine a set of 12 attributes and generate a fault
Index Terms— Convolutional neural networks (CNNs), Hough probability map. More recently, Di et al. [16] combined 14
transform, seismic fault. attributes in a multiattribute support vector machine, a power-
ful supervised learning technique. Such methods give accurate
I. I NTRODUCTION results but are still computationally expensive: in addition to
the set of attributes that must be computed for each new
T HE past decades have seen the development of many tools
for computer-aided fault detection. The vast majority of
these methods are based on the use of seismic attributes. Those
point classification, supervised techniques need as input a large
amount of interpreted data that can be consequent depending
measurements, usually made in the poststack stage, allow on the algorithm applied.
enhancing possible fault location by looking at the local conti- The need for a large amount of interpreted data could
nuity of the seismic signal (coherence [1], [2], semblance [3], explain why more powerful algorithms such as deep learn-
variance [4], chaos [5], and edge detection [6]), or at the ing methods are still scarce in the seismic literature. Such
geometry of the reflectors (curvature [7]–[9] and flexure [10]). algorithms can automatically extract meaningful information
An alternative is to use the information of interpreted horizons from the original amplitudes through their combination in the
to find fault locations (horizons dip and azimuth maps, [11]). network’s hidden layers, creating new features dynamically
Each seismic attribute has its pros and cons, and fails at during training. Despite the usually substantial time to train
enhancing faults only; numerous artifacts remain, other struc- those networks, once the training is done, further outputs are
tures appear. Seismic attributes usually require massive com- obtained very efficiently. There are different approaches to
apply such methods to the problem of fault detection in seismic
Manuscript received April 2, 2018; revised June 11, 2018 and data. Di et al. [30] created a data set by manually labeling
September 18, 2018; accepted October 9, 2018. Date of publication crosslines from a seismic cube, and then achieved fault detec-
November 8, 2018; date of current version February 27, 2019. This work tion in other sections of the same cube. Araya-Polo et al. [17]
was supported in part by Shell Brazil through the “Coupled Geomechanics”
project at Tecgraf Institute (PUC-Rio) and in part by the ANP “Compromisso trained deep networks with thousands of synthetic volumetric
com Investimentos em Pesquisa e Desenvolvimento” through the Research velocity models, obtained by approximating acoustic wave
and Development Levy Regulation. (Corresponding author: Axelle Pochet.) equation to generate wave fields as time-series signals with
A. Pochet and H. Lopes are with the Department of Informatics, Pontifical predefined acquisition geometry. Huang et al. [18] proposed
Catholic University of Rio de Janeiro, Rio de Janeiro 22451-900, Brazil
(e-mail: [email protected]). to train their networks on top of several seismic attributes,
P. H. B. Diniz and M. Gattass are with the department of using synthetic seismic cubes with simple fault configurations.
Informatics, Pontificia Universidade Catolica do Rio de Janeiro, Despite achieving good results, the authors trained and clas-
Rio de Janeiro 22451-900, Brazil.
Color versions of one or more of the figures in this letter are available sified data in the same field, making it difficult to judge the
online at http://ieeexplore.ieee.org. generalization capability (and thus the practical usability) of
Digital Object Identifier 10.1109/LGRS.2018.2875836 the proposed methods.
1545-598X © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Halliburton Company. Downloaded on March 22,2025 at 22:31:23 UTC from IEEE Xplore. Restrictions apply.
POCHET et al.: SEISMIC FAULT DETECTION USING CNNS 353

fault crossing the section entirely, modifying randomly fault

angle, position and throw, shearing slope, folding amplitude
and frequency, wavelet peak, and amount of noise. Resulting
images present amplitude values between −1 and 1. Along
with the seismic amplitude information, we generated for each
image its corresponding binary mask, indicating in white the
location of the fault. Fig. 1 shows an example of such a pair.

B. Patches Extraction
Many machine learning techniques use features extracted
from images as input. Features are relevant information that
we think could be efficiently combined to achieve the desired
Fig. 1. Synthetic Seismic Image. (a) Example of a synthetic seismic image classification. One of the advantages of CNN is that it does not
from our data set. (b) Corresponding binary mask.
require an explicit feature extraction step. Instead, the neural
network uses the image itself as input and attempts to extract
In this letter, we train convolutional neural networks (CNNs)
the best features implicitly.
with synthetic data and try to apply the classifiers on real data
Applying these principles to the seismic imagery area,
to study the generalizability of the model. Designing a robust,
good feature candidates are naturally any fault enhancing
general classifier would indeed save the heavy step of training
seismic attribute. Such attributes are computed using a small
and tuning a CNN for each new data set. CNNs combine
neighborhood of seismic amplitude values. Seismic amplitude
a series of convolution steps with a fully connected neural
is thus at the core of the fault detection problem, and a small
network to perform classification, and recently proved to be
neighborhood of amplitude values can be used as input to a
powerful in various computer vision tasks (see [19], [20]),
CNN, that will hopefully find and compute the best seismic
among which seismic objects detection (see [18], [29], [30]).
attributes dynamically, without the need of explicitly passing
Compared to real data, synthetic data present the great advan-
them as input. This small neighborhood is what we call here
tage to provide total control on the ground truth and are easily
a patch.
scalable in terms of the number of inputs. Our input is the
Since faults may be located anywhere in the seismic image,
seismic amplitude only, and our method does not require any
all pixels are fault candidates. Our approach seeks to classify
prior seismic attribute computation. Our main contribution is
all pixels as a fault or a nonfault pixel. A patch is composed by
hence to propose the use of an error free, easily generated
the candidate pixel itself at the center and its neighbor pixels.
synthetic data to train CNNs with the final goal to classify
The classification of a pixel is the classification of its patch.
real data, in order to avoid the tedious task to mark many
To separate the pixels in our two target classes, fault and
real data for training. We use a patch classification scheme on
nonfault, we use binary mask images, which contain the
synthetic data for network training, as explained in Section II.
marking of the faults. If a pixel in the seismic image is masked
Section III shows the results of our implementation of the
by a white pixel in the binary image, this pixel is considered as
proposed method on synthetic and real data, and in Section IV,
fault. Similarly, black pixels in the binary masks are considered
we draw some conclusions.
Non-fault. In addition, if a pixel is nonfault but the fault passes
somewhere in its patch (partial faulting), we discard the patch:
II. M ETHOD in this letter, such patches are simply not trained. Fig. 2 shows
We propose a methodology in four steps. First, we generate an example of one Fault patch and one nonfault patch extracted
synthetic seismic images where we control the location of the from a synthetic seismic image.
faults. Second, we extract fault and nonfault patches from the Follows how the sets of patches are used as input to the
generated data set. Then, we train and fine-tuned different CNN.
CNN architectures focusing on maximizing quality metrics. 1) For training images, we extract all possible fault patches
Finally, we classify pixels in new images (synthetic and real) and one nonfault patch every 23 pixels. This generates a
and postprocess the results for fault segmentation. balanced number of patches for the two classes, which
is desirable for training.
A. Synthetic Data Set Generation 2) For validation images, we extract all possible fault
patches and one nonfault patch every 10 pixels,
The difficulty in obtaining many good quality fault inter- to account for classes’ natural imbalance in practice, and
pretations on real seismic images led us to investigate the use thus obtain interpretable quality metrics.
of synthetic data with known fault positions. The open source 3) For test images, we extract one patch every 3 pixels,
code IPF from Hale [25] allowed us to reproduce the results of regardless of the binary mask. This small pixel step
migrated, poststacked seismic data. Beginning with a randomly ensures the fault will be crossed, while efficiently gen-
generated reflectivity model extended along the section, simple erating classifications suited for visualization.
image transformations recreate sequential rock deformations
along time: shearing, folding, and faulting. We can then apply
convolution with a Ricker wavelet and add random noise. Each C. CNN Training
step of the process can be parameterized. We built a data set The common architecture of a CNN is described in Fig. 3.
of 500 images of 572 × 572 pixels, all containing one straight The convolutional layers have trainable filters applied

Authorized licensed use limited to: Halliburton Company. Downloaded on March 22,2025 at 22:31:23 UTC from IEEE Xplore. Restrictions apply.
354 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 16, NO. 3, MARCH 2019

TABLE I
CNN A RCHITECTURE T HAT A CHIEVED THE B EST R ESULTS ON
S YNTHETIC D ATA , A FTER 50 E POCHS . W ITH : #F = N UMBER OF
F ILTERS ; MS = M ASK S IZE ; #N = N UMBER OF N EURONS ;
C = C ONVOLUTION , MP = M AX -P OOLING ,
FC = F ULLY C ONNECTED L AYER

We used 400 seismic images for training (381 079 patches),

50 images for validation (148 632 patches), and kept the
Fig. 2. Extracting patches. (a) Seismic image crossed by a fault. The fault
location is highlighted. (b) Fault patch. (c) Nonfault patch. remaining 50 images to perform tests.
For each configuration, we estimate the quality of the
classifier on six common criteria, considering the Fault class
as the positive class: accuracy, sensitivity, specificity, F1-score,
area under the ROC curve (AUC) and a visual evaluation on
entire sections of the test set.
Sensitivity, the capacity of the network to output true posi-
tives, should be high enough to underline the faults. Specificity
should be as close to 1 as possible as even a small number
of false positives tend to give poor visual results on the test
sections.
Fig. 3. Architecture of a CNN, with: [@] = trainable filters and
[w] = trainable weights. Activation and dropout are optional layers. Flatten
operation converts the 2-D list of matrices into a 1-D list of neurons. The D. CNN Classification and Postprocessing
Feature extraction step can contain any number of convolution and pooling
layers. Once a good model is obtained, new images can be classi-
fied. Depending on the image size and fault scale, one can
throughout the patches. Pooling layers perform nonlinear choose to classify all pixels or just a subset. In any case,
downsampling. We use max-pooling, which yields maximum the result will be a binary image with sets of white pixels at
values over a neighborhood of feature maps. Activation layers the predicted fault location. Extracting the fault thus requires a
apply nonlinear functions on input neurons. Here, we use post-processing. We chose a simple sequence of morphological
the rectified linear unit, since it provides several times faster operations.
training than other activation functions [22], usually avoids the 1) Dilation and erosion, to join separated fault pixels and
vanishing gradient problem and promotes model sparsity [27]. erase isolated ones.
The last part of the network is a set of fully connected layers: 2) Thinning, in order decrease the number of fault pixels
the number of input neurons is defined by the number of pixels without losing information.
resulting from the previous layer. Dropout layers prevent the Those steps clean the image and result in smaller fault point
generation of overfitted models, which are common in com- sets. We apply the Hough [26] transform to finally extract the
plex networks. They are generally placed deep in the network, main straight lines, with restrictions on the fault angles, which
after layers learning a large set of parameters, since this is should be subvertical.
where the overfitting risk is high. Supervised training is carried
out using a form of stochastic gradient descent to minimize the III. R ESULTS
discrepancy between the desired output and the current output After training several CNN architectures on synthetic data,
of the network, based on some loss function [21]. We choose we find that the smallest patch size that gives satisfying
the Softmax loss function in the last layer, which outputs a results is 45 × 45. In this section, we present two different
normalized probability for each class. The final classification CNN architectures, the first giving the highest metrics on the
is the class with the highest score. synthetic test set, the second giving the most satisfactory visual
We applied and tested different CNN architectures and result when classifying real data patches.
hyper parameters, along with the input patch size and res-
olution. Beginning with a common LeNet [21], we added
complexity, applying results from the VGG-Net [23], since A. Results on Synthetic Data
deeper networks are better at differentiating classes. Another The best network for fault detection on synthetic data was
aspect is that with deeper networks, small convolution masks obtained with the architecture described in Table I. It gave an
of 3 × 3 pixels should give good results. However, adding accuracy of 0.98, a sensitivity of 0.95, a specificity of 0.99,
too many layers to a network can have negative effects: the F1-score of 0.97, and AUC of 0.99.
training time increases dramatically, and the classification can Fig. 4 (left column) shows the classification and fault
fall into overfitting the training data. Consequently, there is extraction of a section from the test set. Classification is
a tradeoff to find. All training sessions, additionally, shared performed every three pixels. The fault is clearly highlighted,
some parameters: we used a learning rate value of 0.001, but still coarse because patches crossing the fault partially
momentum value of 0.9, and input batch of 30 patches. are classified as fault. Since the training step did not include

Authorized licensed use limited to: Halliburton Company. Downloaded on March 22,2025 at 22:31:23 UTC from IEEE Xplore. Restrictions apply.
POCHET et al.: SEISMIC FAULT DETECTION USING CNNS 355

Fig. 4. Classification and fault extraction on synthetic test sections. (a) Input
section with expected fault marked in dashed lines. (b) Raw classification
of 1 over 3 pixels. (c) Results of erosion, dilation, and thinning on (b).
(d) Extracted faults using the Hough transform on (c).

such patches, as stated in Section II-B, those results are

unsurprising. Note also that such patches did not enter in the
calculation of the quality metrics, over-estimating all values
except sensitivity.
We observed similar results on all test sections, includ-
ing sections with configurations that were not shown during
training: varying fault throw [Fig. 4(center column)], faults Fig. 5. Classification and segmentation on real test sections.
crossing each other [Fig. 4(right column)]. Indeed, as patches (a)–(c), (Top Left) Input section with expected fault location (manual picking).
contain local image information, a variety of fault geometries (Top Right) Classification of all pixels, with superposition of picking for better
can be detected even when not specifically trained. visualization. (Bottom) Extracted faults using our postprocessing procedure.
(a) Single straight fault. (b) Several subvertical faults. (c) Several subvertical
faults with noise.
B. Results on Real Data TABLE II
We test our CNNs on a real data of the North Sea, CNN A RCHITECTURE T HAT A CHIEVED THE B EST R ESULTS ON
the F3 cube [24]. The selected section presents different fault S YNTHETIC D ATA , A FTER 60 E POCHS . W ITH : #F = N UMBER OF
regions. Manual interpretation of the faults in each region can F ILTERS ; MS = M ASK S IZE ; #N = N UMBER OF N EURONS ;
C = C ONVOLUTION , MP = M AX -P OOLING ,
be observed in Fig. 5(a)–(c) (top left). To test our classifier, FC = F ULLY C ONNECTED L AYER
we apply the following process: first, we set the amplitude
values of the entire section between −1 and 1, clipping the
histogram under a threshold to enhance visualization, in order
to set the section as close as possible to the training conditions.
Second, we extract one image per region of interest. Third,
for each image we extract patches with a visual close to
the synthetic patches, because the horizons’ scale in the real Fig. 5(a) shows the detection of a single straight fault.
section is different from the one in the synthetic data set. Fig. 5(b) shows several parallel, subvertical and well-defined
Hence, extracting directly 45 × 45 patches from the real data straight faults, detected by extracting the main peaks in the
can lead to poor results. We assume that the horizon’s scale Hough transform process. Fig. 5(c) presents similar results
should be homogeneous along the regions of interest, and thus on a noisy, multifaulted region. These multifaulted regions
visually select a single patch size per region. We resize those using smaller patches (8 × 8 resized to 45 × 45) allowed to
patches to the expected classification size of 45 × 45, using a separate the faults despite the coarse classification. In addition,
bicubic interpolation. We classify every pixel in the images. in these regions, there was no need to apply morphological
The best CNN architecture in this case was chosen looking operations described in Section II-D before applying the
only at the visual results and is summarized in Table II. Hough Transform.

Authorized licensed use limited to: Halliburton Company. Downloaded on March 22,2025 at 22:31:23 UTC from IEEE Xplore. Restrictions apply.
356 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 16, NO. 3, MARCH 2019

When applying this second CNN on the synthetic validation [5] T. Randen, S. I. Pedersen, and L. Sønneland, “Automatic extraction of
set, we obtain lower quality metrics than the CNN presented fault surfaces from three-dimensional seismic data,” in Proc. SEG Tech.
Program Expanded Abstr., 2001, pp. 551–554.
in Table I (accuracy: 0.94, sensitivity: 0.69, specificity: 0.99, [6] H. B. Di and D. L. Gao, “Gray-level transformation and Canny edge
F1-score: 0.80, and AUC: 0.96). We assume that despite a loss detection for 3D seismic discontinuity enhancement,” Comput. Geosci.,
in quality for Table II network when applied to the synthetic vol. 72, pp. 192–200, Nov. 2014.
data, this CNN has actually a greater generalization capacity. [7] R. J. Lisle, “Detection of zones of abnormal strains in structures
using Gaussian curvature analysis,” AAPG Bull., vol. 78, no. 12,
pp. 1811–1819, 1994.
IV. C ONCLUSION [8] A. Roberts, “Curvature attributes and their application to 3D interpreted
horizons,” First Break, vol. 19, no. 2, pp. 85–100, 2001.
This letter presented a methodology for the detection and [9] S. Al-Dossary and K. J. Marfurt, “3D volumetric multispectral esti-
segmentation of faults in seismic images, using a patch- mates of reflector curvature and rotation,” Geophysics, vol. 71, no. 5,
based CNN approach for training and classification, and line pp. P41–P51, 2006.
[10] D. Gao, “Integrating 3D seismic curvature and curvature gradient
Hough Transform for post-processing. CNNs were trained attributes for fracture characterization: Methodologies and interpreta-
with synthetic data only and yielded good results when applied tional implications,” Geophysics, vol. 78, no. 2, pp. O21–O31, 2013.
to new synthetic data, with a perfect match of the predicted [11] E. J. H. Rijks and J. C. E. M. Jauffred, “Attribute extraction: An impor-
tant application in any detailed 3-D interpretation study,” Lead. Edge,
and ground truth fault after applying the postprocess. They also vol. 10, no. 9, pp. 11–19, 1991.
revealed promising results when classifying preprocessed real [12] D. Gibson, M. Spann, J. Turner, and T. Wright, “Fault surface detection
data. Interestingly, the CNN with the highest performance in in 3-D seismic data,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 9,
both synthetic and real case is different. Table II CNN revealed pp. 2094–2102, Sep. 2005.
[13] B. Zhang, Y. Liu, M. Pelissier, and N. Hemstra, “Semiautomated fault
a greater generalization capacity than Table I CNN, despite interpretation based on seismic attributes,” Interpretation, vol. 2, no. 1,
poorest quality metrics on the synthetic validation set. pp. SA11–SA19, 2014.
The use of CNN allowed us to train with the seismic [14] Z. Wang and G. AlRegib, “Fault detection in seismic datasets using
amplitude map as the only input feature: explicit feature Hough transform,” in Proc. IEEE Int. Conf. Acoust., Speech Signal
Process. (ICASSP), May 2014, pp. 2372–2376.
extraction and selection steps were unnecessary. However, [15] K. M. Tingdahl and M. de Rooij, “Semi-automatic detection of faults
we highlighted the large number of empirical parameters used in 3D seismic data,” Geophys. Prospecting, vol. 53, no. 4, pp. 533–542,
in CNNs, which makes model fine-tuning difficult. 2005.
[16] H. Di, M. A. Shafiq, and G. AlRegib, “Seismic-fault detection based
There are a few changes that can be applied to achieve on multiattribute support vector machine analysis,” in Proc. SEG Tech.
higher performance, especially in real seismic images. First, Program Expanded Abstr., 2017, pp. 2039–2044.
the patch size in the case of real data could be extracted auto- [17] M. Araya-Polo, T. Dahlke, C. Frogner, C. Zhang, T. Poggio, and
matically, considering whether the scale of the horizons (using D. Hohl, “Automated fault detection without seismic processing,” Lead.
Edge, vol. 36, no. 3, pp. 208–214, 2017.
image texture measurements), or the seismic signal frequency [18] L. Huang, X. Dong, and T. E. Clee, “A scalable deep learning platform
(patches should have a similar frequency range). Ideally, for identifying geologic features from seismic attributes,” Lead. Edge,
we could try to generate a data set containing the commonly vol. 36, no. 3, pp. 249–256, 2017.
[19] W. Rawat and Z. Wang, “Deep convolutional neural networks for image
seen real seismic signal frequencies. Second, the architecture classification: A comprehensive review,” Neural Comput., vol. 29, no. 9,
of the CNN could be improved. To do so, we could try to pp. 2352–2449, 2017.
visualize and better understand the generated features, or we [20] W. Zhiqiang and L. Jun, “A review of object detection based on
could use evolutionary or other optimization algorithms to find convolutional neural network,” in Proc. 36th Chin. Control Conf. (CCC),
Jul. 2017, pp. 11104–11109.
the best architecture automatically. The question of patches [21] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based
with partial faulting should also be addressed. Including such learning applied to document recognition,” Proc. IEEE, vol. 86, no. 11,
patches as a new label in the classification could be a way to pp. 2278–2324, Nov. 1998.
refine classification around the fault. Another line of research [22] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification
with deep convolutional neural networks,” in Proc. Adv. Neural Inf.
could be the improvement of our postprocessing, following Process. Syst., 2012, pp. 1097–1105.
better skeletonization algorithms, as [31], for example. Finally, [23] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
while promising results on real data show the relevance of large-scale image recognition,” in Proc. Int. Conf. Learn. Represent.,
2014, pp. 1–14.
the synthetic training methodology, it is clear that it may [24] Netherland Offshore F3 Block Complete. Accessed: Sep. 2017.
be more efficient to use real seismic images for training. [Online]. Available: https://www.opendtect.org/osr/Main/
Therefore, we are currently trying fine-tuning and transfer NetherlandsOffshoreF3BlockComplete4GB
learning methods to include some real data in the training [25] D. Hale. (2014). Seismic Image Processing for Geologic Faults.
[Online]. Available: https://github.com/dhale/ipf
and better adapt our CNN to real data. [26] P. V. C. Hough, “Method and means for recognizing complex patterns,”
U.S. Patent 3 069 654, Dec. 18, 1962.
[27] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier
R EFERENCES neural networks,” in Proc. 14th Int. Conf. Artif. Intell. Stat., 2011,
[1] M. Bahorich and S. Farmer, “3-D seismic discontinuity for faults and pp. 315–323.
stratigraphic features: The coherence cube,” Lead. Edge, vol. 14, no. 10, [28] Z. Wang and G. AlRegib, “Interactive fault extraction in 3-D seismic
pp. 1053–1058, 1995. data using the Hough transform and tracking vectors,” IEEE Trans.
[2] Y. Luo, W. G. Higgs, and W. S. Kowalik, “Edge detection and strati- Comput. Imag., vol. 3, no. 1, pp. 99–109, Mar. 2017.
graphic analysis using 3D seismic data,” in Proc. SEG Tech. Program [29] A. U. Waldeland and A. H. S. S. Solberg, “Salt classification using deep
Expanded Abstr., 1996, pp. 324–327. learning,” in Proc. 79th EAGE Conf. Exhib., 2017.
[3] K. J. Marfurt, J. R. L. Kirlin, S. L. Farmer, and M. S. Bahorich, [30] H. Di, Z. Wang, and G. AlRegib, “Seismic fault detection from post-
“3-D seismic attributes using a semblance-based coherency algorithm,” stack amplitude by convolutional neural networks,” in Proc. 80th EAGE
Geophysics, vol. 63, no. 4, pp. 1150–1165, 1998. Conf. Exhib., 2018.
[4] P. P. van Bemmel and R. E. F. Pepper, “Seismic signal processing [31] Z. Wang, D. Temel, and G. Alregib, “Fault detection using color
method and apparatus for generating a cube of variance values,” blending and color transformations,” in Proc. IEEE Global Conf. Signal
U.S. Patent 6 151 555, Nov. 21, 2000. Inf. Process. (GlobalSIP), Dec. 2014, pp. 999–1003.

Authorized licensed use limited to: Halliburton Company. Downloaded on March 22,2025 at 22:31:23 UTC from IEEE Xplore. Restrictions apply.