Journal of Applied Geophysics 137 (2017) 55–62

Contents lists available at ScienceDirect

Journal of Applied Geophysics

journal homepage: www.elsevier.com/locate/jappgeo

Seismic detection method for small-scale discontinuities based on

dictionary learning and sparse representation
Caixia Yu a, Jingtao Zhao b,⁎, Yanfei Wang a
a
Key Laboratory of Petroleum Resources Research, Institute of Geology and Geophysics, Chinese Academy of Sciences, Beijing 100029, China
b
State Key Laboratory of Coal Resources and Safe Mining, China University of Mining and Technology, Beijing 100083, China.

a r t i c l e i n f o a b s t r a c t

Article history: Studying small-scale geologic discontinuities, such as faults, cavities and fractures, plays a vital role in analyzing
Received 23 March 2016 the inner conditions of reservoirs, as these geologic structures and elements can provide storage spaces and
Received in revised form 7 November 2016 migration pathways for petroleum. However, these geologic discontinuities have weak energy and are easily
Accepted 2 December 2016
contaminated with noises, and therefore effectively extracting them from seismic data becomes a challenging
Available online 05 December 2016
problem. In this paper, a method for detecting small-scale discontinuities using dictionary learning and sparse
Keywords:
representation is proposed that can dig up high-resolution information by sparse coding. A K-SVD (K-means
Small-scale discontinuities clustering via Singular Value Decomposition) sparse representation model that contains two stage of iteration
Sparse representation procedure: sparse coding and dictionary updating, is suggested for mathematically expressing these seismic
Dictionary learning small-scale discontinuities. Generally, the orthogonal matching pursuit (OMP) algorithm is employed for sparse
Orthogonal matching pursuit coding. However, the method can only update one dictionary atom at one time. In order to improve calculation
efficiency, a regularized version of OMP algorithm is presented for simultaneously updating a number of atoms at
one time. Two numerical experiments demonstrate the validity of the developed method for clarifying and
enhancing small-scale discontinuities. The field example of carbonate reservoirs further demonstrates its
effectiveness in revealing masked tiny faults and small-scale cavities.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction
Many significant researches on detecting geologic discontinuities
have been conducted in seismic exploration and the well-known
coherency algorithms (Bahorich and Farmer, 1995; Marfurt et al.,
1998, 1999; Gersztenkorn and Marfurt, 1999) have been updated
to the third generation. Some detection methods based on signal
processing are also explored for reflection enhancement and edge
sharpness, such as matching pursuit algorithm (Mallat and Zhang,
1993; Castagna et al., 2003; Liu and Marfurt, 2005), spectral decomposi-
tion algorithms (Partyka et al., 1999; Puryear et al., 2012; Gao, 2013)
and edge-preserving processing (Luo et al., 2002; Fehmers and
Höecker, 2003). Until recently, there is some lack of knowledge about
using sparse optimization theory to detect seismic small-scale geologic
discontinuities. Studying these small-scale discontinuities belongs to
the scope of high-resolution recovery, with the weak information
masked by strong background and noises. A strategy is to reconstruct
them using dictionary learning and sparse representation.
The method of dictionary learning and sparse representation has
demonstrated successful applications in the community of image recov-
ery and feature extraction. The key of such a method is to create a
⁎ Corresponding author.
E-mail address: diffzjt@163.com (J. Zhao).
dictionary for sparse representation of signals and the dictionary can ei-
ther be chosen as a predefined set of basic functions or the one learned
from a set of given examples. Sometimes a predefined transformation is
really an attractive choice because of its speediness in estimating sparse
representation, such as wavelet (Freeman and Adelson, 1991; Shan
et al., 2009), curvelet (Cao et al., 2015) and wedgelet (Donoho, 1998).
However, the success of such a predefined dictionary depends on its
fitting to the sparsely described signals. On the contrary, the learned
dictionary can deal with more complicated data and has already been
extensively used in many fields, such as image processing (Yang et al.,
2016; Yeganli et al., 2016), speech signal representation (Jafari and
Plumbley, 2011; You et al., 2014) and seismic data processing (Zhou
et al., 2014; Chen et al., 2016). Since seismic small-scale discontinuities
that are mainly associated with faults and cavities possess a weak
spatial correlation, effectively detecting and recovering them from seis-
mic data is a challenge to the predefined transformations. In addressing
such an inverse problem by the Bayesian approach, a prior acknowledge
based on some simplifying assumptions, such as spatial smoothness
or entropy, is usually proposed, and the quality of these methods
strongly depend on a guess of the prior information. However, the
example-based technique suggests learning this prior information
from seismic data and merger it in training a dictionary. Therefore, the
way of creating dictionaries by training them from a given example
set is employed in this paper.

http://dx.doi.org/10.1016/j.jappgeo.2016.12.005
0926-9851/© 2016 Elsevier B.V. All rights reserved.
56 C. Yu et al. / Journal of Applied Geophysics 137 (2017) 55–62

The structure of the paper is organized as follows. First, a sparsity-

constraint model for expressing seismic small-scale discontinuities is
presented. Second, an effective implementation scheme of the K-SVD
algorithm is suggested for solving this model. Third, two numerical
experiments and one field application to test the feasibility of the
proposed method are provided. A conclusion and discussion will end
this paper.

2. Seismic detection method for small-scale discontinuities

2.1. Sparse model for representing small-scale discontinuities

The basic model for sparse representation indicates that signals can
be effectively expressed as a linear combination of atoms and the
corresponding linear coefficients are sparse (Rubinstein et al., 2010;
Wang et al., 2012). For recovering the small-scale discontinuities from
seismic data, the classic sparse representation model is used:

min kY − DXk2F
D;X

ð1Þ
s:t:kXi k0 ≤ δ;
D j
¼ 1; i ¼ 1; 2; …; M; j ¼ 1; 2; …; R
2 Fig. 2. Prestack depth image of modified Marmousi model with the second and third
strings of cavities shown except for the first smallest scale one.
where matrix Y is seismic data with the size of N × M after the
plane-wave destruction (PWD) (Fomel, 2002; Yu et al., 2015), matrix
D is a dictionary with the size of N × R and X is a sparse representation
matrix with the size of R × M. The notation ‖·‖2F is Frobenius norm. Vec-
tors Xi and Dj stand for the i-th column and the j-th column of the ma- Step 1 Input matrix Y, initial dictionary D0, desired sparsity level δ,
trices X and D, respectively. Parameter δ controls the sparsity level the maximum iteration number n and initial loop index k = 0;
and notation ‖·‖0 counts non-zero elements. Step 2 Calculate sparse representation matrix X by solving the following
In order to efficiently train the sparse dictionary from given sub-problem:
examples, the K-SVD scheme that is the generalization of the K-means
cluster is adopted because of its fast convergence.

Xi ¼ arg min kYi − Dx k22 ; s:t:kx k0 ≤ δ

2.2. K-SVD scheme for the sparse model x

The K-SVD algorithm (Elad and Aharon, 2006; Rubinstein et al., where Xi , Yi , i = 1 , 2 , … , M are the i-th column of matrices X and Y,
2008) that trains a dictionary from given examples has already been respectively;
successfully used in many signal processing tasks. However, this meth-
od requires great memory when the dimensions of dictionary matrix in- Step 3 Update dictionary by individually calculating its every column:
crease or the training samples become large. A strategy for promoting
its calculation efficiency is suggested by modifying dictionary updating Dj = 0; (j =1 , 2 , … , R is the j-th column of matrix D)
steps and sparse-coding process. In this paper, a regularized version of
OMP to sparsely code large signals is employed. The corresponding pro-
cedure for iteratively solving the proposed sparse model (1) is listed:

Algorithm 1. (Solving the sparse representation model)

Fig. 3. Statistical analysis for showing the convergence of the K-SVD algorithm in modified
Fig. 1. Modified Marmousi model with three strings of different scales of cavities. Marmousi model with 100 times of iterations.
 C. Yu et al. / Journal of Applied Geophysics 137 (2017) 55–62 57

Fig. 4. Seismic small-scale discontinuity profile revealing the smallest string of cavities and Fig. 6. Prestack depth image of the 3D SEG/EAGE overthrust model (inline direction) with
velocity inhomogeneities. large-scale layers and macro-scale faults shown, except for small-scale geologic details.
(For interpretation of the references to color in this figure, the reader is referred to the
web version of this article.)
g = XTj , I (I means indices of column in matrix Y which use the
j-th atom)

d ¼ YI g − DðXI gÞ
Step II Identification. Choose a set J by picking δ biggest coordinates in the
d :¼ d=kdk2
magnitude of the vector u=Dr.
g ¼ YTI d − ðDXI ÞT d
Step III Regularization. Divide the set J into subsets J0, according to
Dj ¼ d
conditions: | u(i)| ≤ 2 | u(j)| for i , j ∈ J0, and choose the one with
X j;I ¼ gT
the maximal energy ‖u|J0‖2.
Step 4 If k exceeds the user-defined maximum number of iterative steps, Step IV Updating. Add subset J0 to the index set p, and solve a least squares
output D and calculate Y = DX, STOP; Otherwise, set k = k + 1 problemXi ¼ min ∈ℜ p kYi −Dxk2 to get a sparse coefficient vector
x
and GOTO Step 2. Xi. The corresponding new fitting residual is updated by r=Yi −DXi.
Step V If n exceeds the user-defined maximum number of iterative steps or
The main computation cost lies in Step 2 and a regularized version
the maximal energy ‖u|J0‖2 ≤ ξ, output Xi and STOP; Otherwise,
of OMP algorithm (Needell and Vershynin, 2009) is used for solving
set n: =n+1 and GOTO Step II.
this sub-problem. The main target is to generate a vector Xi of sparse
coefficients, with the given input dictionary D, data Yi, sparsity level δ
and termination parameter ξ. The iteration procedure of this algorithm Remark. The implementations of the identification in Step II and the
is listed as follows. regularization in Step III have advantages in computation performance,
as these procedures can be performed in a linear time.
Algorithm 2. (Solving the sub-problem in Algorithm 1)
Step I Initialization. Set iteration index n = 0, fitting residual r = Yi,
tolerance ξN 0, and index set p=0 (p means index set correspond-
ing to indices of non-zero values in the sparse coefficient vector Xi);

Fig. 7. Statistical analysis for showing the convergence of the K-SVD algorithm in 3D
Fig. 5. Velocity model of the 3D SEG/EAGE overthrust model (inline direction). SEG/EAGE overthrust model with 100 times of iterations.
58 C. Yu et al. / Journal of Applied Geophysics 137 (2017) 55–62
Fig. 8. Seismic small-scale discontinuity profile of 3D SEG/EAGE overthrust model (inline
direction): black arrows, red circles and red square indicate faults, edges and channel,
respectively. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)
The dominant operations of the Algorithm 2 at each iteration
includes the complexity of applying D (2NR operations), back-
substitution process for solving Xi (N2 operations), calculating the
residual r (2NR + N operations). The computational cost of K iterations
of the Algorithm 2 is given by
 
T A2 ¼ K N2 þ 4NR þ N ð2Þ
The computation cost of the Algorithm 1 at each iteration requires
calculating the sparse representation matrix (MTA2 operations), apply-
ing D twice for multiplication in updating dictionary (4NR operations).
Applying D (2NR operations) for calculating Y in the final stage should
also be included.
Therefore, the dominate computational cost of the proposed
approach is.
T ¼ nðMT A2 þ 4NRÞ þ 2NR ð3Þ
Fig. 9. Seismic coherence profile of the 3D SEG/EAGE overthrust model (inline direction)
with large-scale discontinuities revealed but small-scale detailed information obscured.
Fig. 10. Velocity model of the 3D SEG/EAGE overthrust model (crossline direction).
3. Numerical experiments
In this section, two numerical experiments are given to individually
test the proposed method in detecting two-dimensional and three-
dimensional small-scale discontinuities. The first experiment is a
two-dimensional modified Marmousi model and the second one is a
three-dimensional SEG/EAGE Overthurst model. Seismic synthetic
data of both models are added with noises for demonstrating the
robustness of the proposed method.
3.1. Modified Marmousi model
For studying the carbonate reservoirs, clarifying and positioning the
cavities are usually the primary task. However, these geologic structures
are small-scale objects and their seismic responses are difficult to be
detected. In order to illustrate this problem, the classic Marmousi
model is modified by adding different scales of cavities. The original
Marmousi model includes 158 horizontal layers and a series of faults
and tilted blocks. Three strings of cavities were added in the deep part
in Fig. 1. The first string is composed of four cavities with the velocity
as 1500 m/s and the scale as 5 m. The second string consists of five
cavities with the velocity as 2000 m/s and the scale as 20 m. The last
string includes four cavities with the velocity as 1500 m/s and the
scale as 50 m.
Fig. 11. Prestack depth image of the 3D SEG/EAGE overthrust model (crossline direction)
showing large-scale layers except for small-scale geologic details.
 C. Yu et al. / Journal of Applied Geophysics 137 (2017) 55–62
Fig. 12. Seismic small-scale discontinuity profile of 3D SEG/EAGE overthrust model
(crossline direction): black arrows, red circles and red square indicate faults, edges and
channel, respectively and black circles indicate dramatic amplitude changes. (For
interpretation of the references to color in this figure legend, the reader is referred to
the web version of this article.)
The dominant frequency of seismic source wavelet is chosen as
30 Hz. Prestack depth image of the modified Marmousi model with its
synthetic data added noises is displayed in Fig. 2. The information of
large layers, faults and even the second and third strings of cavities
can be observed. However, the smallest string of cavities is completely
lost. In the stage of petroleum exploration, the locations of faults and
cavities are needed to clarify for well deployment. Based on this
consideration, the proposed seismic sparsity detection method for
small-scale discontinuities is used. The quality of RMSE (root mean
squared error) is used for evaluating the preformation of a solution
and is calculated by the flowing equation:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
kY − DXk2F
RMSE ¼
numðYÞ
where the num(Y) means the number of elements in a matrix.
Fig. 13. Seismic coherence profile of the 3D SEG/EAGE overthrust model (crossline
direction) revealing large-scale discontinuities except for small-scale detailed
information.
59
Fig. 14. Prestack time imaging profile of 3D land field data (inline direction).
Seismic high-resolution information that is associated with small-
scale geologic faults and cavities generally possesses a feature of discon-
tinuities or inhomogeneities. Trying to preserve the true amplitude of
this information has always been a challenging issue and sometimes a
solution with a relatively small error and a good performance of conver-
gence is acceptable. The RMSE of this example in Fig. 3 statistically
shows a uniform convergence of the K-SVD algorithm and the running
times implemented on Intel Core i7-3770 machine is 1 min.
The corresponding result of seismic small-scale discontinuities is
displayed in Fig. 4. In this profile, three series of faults corresponding
to titled blocks and two strings of cavities with the scale of 20 m and
50 m can be clearly seen. The masked string of cavities with the scale
of 5 m and velocity inhomogeneities are also revealed. The modified
Marmousi model demonstrates the effectiveness of proposed method
for high-quality detecting small-scale geologic objects and its potential
in reservoir analysis.
3.2. 3D SEG/EAGE overthrust model
The 3D SEG/EAGE overthrust model depicts a complex sedimentary
sequence, which includes unconformities, weathering zones, tilted
Fig. 15. Statistical analysis for showing the convergence of the K-SVD algorithm in 3D field
data with 100 times of iterations.
60 C. Yu et al. / Journal of Applied Geophysics 137 (2017) 55–62
Fig. 16. Seismic small-scale discontinuity profile of 3D field data (inline direction): black
arrows, red squares, and red circles indicate tiny faults, inside story of weather crusts
and resolved cavities, respectively. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)
faults, a kink folding and channels. Such discontinuous subsurface
features are targets of the proposed method. In order to test the stability
of the proposed method, some Gaussian noises are added in seismic
synthetic data. One inline velocity profile is displayed in Fig. 5 and its
prestack depth image is shown in Fig. 6. In this inline profile, there
are a kink folding in anticline, two main faults and channels. This
image can show large-scale layer information, basement and thrusts.
However, the small-scale discontinuous features are hard to be detect-
ed, and thus some strategy should be taken to reveal this information.
In Fig. 7, the RMSE statistically shows the uniform convergences of
the K-SVD algorithm in a 3D case and the cost of running times is
15 min. The corresponding seismic sparsity detection profile for
small-scale discontinuous features is shown in Fig. 8. Comparing Fig. 6
with Fig. 8, the boundaries of two faults can be clearly indicated by
black arrows. There are also some other interesting features with
red circles corresponding to the edges of reflectors and red square
indicating channels. After dictionary learning process, these small-
scale discontinuous features find a way to stand out.
Fig. 17. Seismic coherence profile of the 3D land field data (inline direction) with
large-scale discontinuities revealed but small-scale detailed information obscured.
Fig. 18. Prestack time imaging profile of 3D field data (crossline direction).
To further confirm these resolved small-scale discontinuities, the
seismic coherence profile is given in Fig. 9. Comparing with the
marked places in Fig. 8, Fig. 9 gives a clue but not high-resolution in-
formation about the small-scale discontinuities. One crossline direc-
tion velocity profile is provided in Fig. 10 and its corresponding
prestack depth image is given in Fig. 11. At the shallow part of this
crossline profile, the complex contact relations between different
layers that associate edges and vertexes can be seen. In the seismic
sparsity detection profile (Fig. 12), the information of edges and ver-
texes marked with red circles can be clearly defined. In the middle
part, the micro-scale faults are picked out marked with red arrows
and the places marked with black circles indicate the amplitude
changes in layer reflections. The channel information can be
obviously viewed in Fig. 12 marked with red square. However, this
information can only be guessed if only Fig. 11 is available. At
the bottom, the black arrows indicate the locations of faults. The
corresponding seismic coherency profile is shown in Fig. 13 to confirm
the revealed small-scale discontinuous geologic features.
Fig. 19. Seismic small-scale discontinuity profile of 3D field data (crossline direction):
black arrows and red circles indicate faults, dramatic amplitude changes of the layer,
respectively. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)
 C. Yu et al. / Journal of Applied Geophysics 137 (2017) 55–62 61

Fig. 20. Seismic coherence profile of the 3D field data (crossline direction) with large-scale
discontinuities revealed but small-scale detailed information obscured. (For interpretation
of the references to color in this figure, the reader is referred to the web version of
this article.)

The 3D SEG/EAGE over-thrust model demonstrates the ability of the

proposed seismic sparsity detection method for revealing the tiny faults,
edges and channel in complex geologic cases.

4. Field data application

Clarifying the small-scale faults and cavities plays a significant role in

analyzing carbonate reservoirs. The faults can provide effective paths for
hydrocarbon migration and cavities may generate storage spaces. The
research field is from northwest region of China and the seismic data
is collected on land with 3D survey. The aim of reservoir analysis is to
discriminate and locate faults and cavities for well deployment pro-
gram. After standard seismic processing, the prestack time migration
profile (inline direction) is shown in Fig. 14. The large-scale layers can
be clearly viewed and macro-scale cavities in the bottom also show
up. According to the available geologic investigation, there are many
tiny faults developed at the shallow part of the region but their real
paths are not clear. In the middle part, the Ordovician weathering crusts
are the main prospective targets but their inner conditions are also
ill-defined. At the bottom, there are many cavities, in the forms of single
ones or group ones. These small-scale geologic discontinuities can't be
interpreted from the conventional seismic image.
After analyzing the convergence behavior of the K-SVD algorithm
implemented on this 3D data in Fig. 15, the corresponding small-scale
information is shown in Fig. 16. The running time of 100 iterations is
Fig. 21. Seismic coherence maps along one layer: (a) Coherence attribute of prestack time
about 15 min. In Fig. 16, the black arrows indicate tiny faults that locate
imaging data; (b) Coherence attribute of seismic small-scale discontinuity data.
at two places of the profile and the red squares indicate inner features of
the weathering crusts that reveals the broken conditions and possible
existence of an underground river. At the bottom, many marked cavities for seal evaluation. The seismic coherence profile is also provided in
with red circles emerge in the form of single ones or group ones. The Fig. 20 for confirming these discovered small-scale discontinuities.
corresponding seismic coherence profile is provided in Fig. 17 that can In order to understand the planer distributions of revealed faults and
confirm the marked small-scale discontinuities in Fig. 16. Comparing cavities, the seismic attribute analysis is performed along one layer. The
with seismic coherence attributes, seismic sparsity detection profiles coherence attribute of conventional prestack time imaging data and
possess high-resolution abilities and exhibit waveform characteristics. seismic sparsity detection data are shown in Fig. 21(a) and Fig. 21(b),
One crossline direction profile of prestack time migration is presented respectively. Comparing the marked places with red arrows in
in Fig. 18 and its corresponding seismic sparsity detection profile is Fig. 21(a) with Fig. 21(b), some tiny faults are resolved by the proposed
displayed in Fig. 19. Many small-scale discontinuous geologic phenom- method. These tiny faults can assist geologists in studying the connec-
ena show up, such as faults marked by black arrows and dramatic tivity of carbonate reservoirs. Additionally, the discovered cavities
amplitude changes marked by red circles. Amplitude changes on the distributed in the northern region as indicated by red circles can provide
layers can reflect broken conditions and this information is usefully valuable information for well deployment program. Through this 3D
62 C. Yu et al. / Journal of Applied Geophysics 137 (2017) 55–62
land field application, the proposed method demonstrates its ability for
successfully detecting the hidden tiny-fault systems and small-scale
cavities.
5. Conclusion
Classifying and locating small-scale geologic discontinuities is a sig-
nificant target for analyzing reservoirs, as planning a well deployment
program depends on the distributions of these object. A challenging
problem for effectively extracting them lies in their weak energy and
easy contamination with noises. In order to dig up this high-resolution
information from seismic data, a small-scale discontinuity detection
method using dictionary learning and sparse representation is pro-
posed. The used K-SVD algorithm shows a uniform convergence and
its limited computation cost exhibits a great potential for processing
3D seismic data.
The modified 2D Marmousi and 3D SEG/EAGE over-thrust model
demonstrate the ability of the proposed method for resolving the
small-scale faults, cavities and channels. The 3D field application further
confirms its advantages in detecting masked tiny-fault system and
small-scale cavities. Comparing with seismic coherence attributes,
seismic sparsity detection profiles can reflect small-scale discontinuous
information and possess waveform characteristics. Application of the
proposed seismic small-scale discontinuity detection method can be
extended to other fields, such as coal mining for collapsed columns or
civil engineering for buried pipes.
Acknowledgements
This work is supported by National Natural Science Foundation of
China under grant numbers 41325016, 41604125 and the Strategic
Priority Research Program of the Chinese Academy of Sciences under
grant number XDB10020101. Chinese postdoctoral science fund under
grant number 2014 M550829 is also acknowledged.
References
Bahorich, M.S., Farmer, S.L., 1995. 3-D seismic discontinuity for faults and stratigraphic
features, the coherence cube. Lead. Edge 16, 1053–1058.
Cao, J.J., Zhao, J.T., Hu, Z.Y., 2015. 3D seismic denoising based on a low-redundancy
curvelet transform. J. Geophys. Eng. 12 (4), 566–576.
Castagna, J.P., Sun, S.J., Seigfried, R.W., 2003. Instantaneous spectral analysis: detection of
low-frequency shadows associated with hydrocarbons. Lead. Edge 22, 120–127.
Chen, Y., Ma, J., Fomel, S., 2016. Double-sparsity dictionary for seismic noise attenuation.
Geophysics 81 (2), V17–V30 (2016).
Donoho, D.L., 1998. Wedgelets: nearly minimax estimation of edges. Ann. Stat. 27 (3),
859–897.
Elad, M., Aharon, M., 2006. Image denoising via sparse and redundant representations
over learned dictionaries. IEEE Trans. Image Process. 15 (12), 3736–3745.
Fehmers, G.C., Höecker, F.W., 2003. Fast structural interpretation with structure-oriented
filtering. Geophysics 68 (4), 1286–1293.
Fomel, S., 2002. Application of plane-wave destruction filters. Geophysics 67 (6),
1946–1960.
Freeman, W.T., Adelson, E.H., 1991. The design and use of steerable filters. IEEE Pat. Anal.
Mach. Intell. 13 (9), 891–906.
Gao, D.L., 2013. Wavelet spectral probe for seismic structure interpretation and fracture
characterization: a workflow with case studies. Geophysics 78 (5), O57–O67.
Gersztenkorn, A., Marfurt, K.J., 1999. Eigenstructure-based coherence computations as an
aid to 3-D structural and stratigraphic mapping. Geophysics 64 (5), 1468–1479.
Jafari, M.G., Plumbley, M.D., 2011. Fast dictionary learning for sparse representations of
speech signals. IEEE J. Sel. Top. Sign. Proces. 5 (5), 1025–1031.
Liu, J.L., Marfurt, K.J., 2005. Matching pursuit decomposition using Morlet wavelets.
Expanded Abstracts, 75th SEG Meeting, Houston, Texas, USA, pp. 786–789.
Luo, Y., Marhoon, M., Dossary, S.A., Alfaraj, M., 2002. Edge-preserving smoothing and
applications. Lead. Edge 21, 136–158.
Mallat, S.G., Zhang, A.F., 1993. Matching pursuits with time-frequency dictionaries. IEEE
Trans. Signal Process. 41 (12), 3397–3415.
Marfurt, K.J., Kirlin, R.L., Farmer, S.L., Bahorich, M.S., 1998. 3-D seismic attributes using a
semblance-based coherence algorithm. Geophysics 63 (4), 1150–1165.
Marfurt, K.J., Sudhaker, V., Gersztenkorn, A., Crawford, K.D., Nissen, S.E., 1999. Coherence
calculations in the presence of strong structural dip. Geophysics 64 (1), 104–111.
Needell, D., Vershynin, R., 2009. Uniform uncertainty principle and signal recovery via
regularized orthogonal matching pursuit. Found. Comput. Math. 9, 317–334.
Partyka, G.A., Gridley, J.A., Lopez, J.A., 1999. Interpretational aspects of spectral decompo-
sition in reservoir characterization. Lead. Edge 18, 353–360.
Puryear, C.L., Portniaguine, O.N., Cobos, C.M., Castagna, J.P., 2012. Constrained least-
squares spectral analysis: application to seismic data. Geophysics 77 (5), V143–V167.
Rubinstein, R., Zibulevsky, M., Elad, M., 2008. Efficient implementation of the K-SVD
algorithm using batch orthogonal matching pursuit. Technical Report - CS Technion,
pp. 1–15.
Rubinstein, R., Zibulevsky, M., Elad, M., 2010. Double sparsity: learning sparse dictionaries
for spare signal approximation. IEEE Trans. Signal Process. 58 (3), 1553–1564.
Shan, H., Ma, J.W., Yang, H.Z., 2009. Comparisons of wavelets, contourlets and curvelets in
seismic denoising. J. Appl. Geophys. 69 (2), 103–115.
Wang, Y.F., Yang, C.C., Cao, J.J., 2012. On Tikhonov regularization and compressive
sensing for seismic signal processing. Math. Models Methods Appl. Sci. 22 (2)
(1150008–1–1150008-24).
Yang, L., Fang, J., Cheng, H., et al., 2016. Sparse Bayesian dictionary learning with a
Gaussian hierarchical model. Signal Process. 130, 93–104.
Yeganli, F., Nazzal, M., Unal, M., et al., 2016. Image super-resolution via sparse represen-
tation over multiple learned dictionaries based on edge sharpness. SIViP 10 (3), 1–8.
You, D., Han, J., Zheng, G., et al., 2014. Sparse representation with optimized learned
dictionary for robust voice activity detection. Circuits Syst. Signal Process. 33 (7),
2267–2291.
Yu, C.X., Wang, Y.F., Zhao, J.T., 2015. A seismic diffraction extraction method for the study
of discontinuous geologies using a regularisation algorithm. Explor. Geophys. http://
dx.doi.org/10.1071/EG15006.
Zhou, Y., Chen, W., Gao, J., 2014. Separation of seismic blended data by sparse inversion
over dictionary learning. J. Appl. Geophys. 106 (7), 146–153.