IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.

60, 2022 5913613

Erratic Noise Attenuation Using Double Sparsity

Dictionary Learning Method
Lakshmi Kuruguntla , Graduate Student Member, IEEE,
Vineela Chandra Dodda , Graduate Student Member, IEEE, Anup Kumar Mandpura, Member, IEEE,
and Karthikeyan Elumalai , Member, IEEE

Abstract— In seismic data processing, attenuation of erratic the incoherent noise is randomly distributed along the data
noise is a challenging task due to the unknown noise distribution. and therefore it is difficult to estimate. The incoherent noise
Erratic noise consists of high amplitude peaks and conventional can be either Gaussian distributed or non-Gaussian distributed.
sparse transforms based on least-square (LS) approach that is
not appropriate for erratic noise attenuation. An alternative The noise which is uncorrelated from trace to trace and
approach, where the data with erratic noise are transformed follows a Gaussian distribution is termed as random noise. The
into pseudodata and then denoised based on fast discrete curvelet incoherent noise with non-Gaussian distribution is also known
transform with structure-oriented space-varying median filtering, as Erratic noise [2]. The pre-stack seismic data contain non-
is performed and achieves better attenuation. However, the fast Gaussian noise (NGN) due to various factors such as poorly
discrete curvelet transform with a fixed basis lacks the adaptivity
for various data patterns of seismic data. Hence, in this article, coupled geophones, recording errors, wind, and rain. The types
we propose a double sparsity dictionary learning (DSDL) method of erratic noise present in the seismic data include swell
which performs denoising and also preserves the original features noise, power line noise, noise bursts, traffic noise, harmonic
of seismic data. The proposed method combines the strength noise, and polarity reversals. These noises compose of spatially
of the analytical transform and adaptive transform to attenuate isolated events and spikes which change from one section to
both random noise and erratic noise in data. The performance of
proposed DSDL method is studied on synthetic datasets and field another in seismic data. Attenuation of these erratic noises
datasets. The numerical results demonstrate that the proposed in the recorded seismic data is paramount to improve the
method gives a better signal-to-noise ratio (SNR), a lower mean- signal-to-noise ratio (SNR). However, attenuation methods
squared error, and energy values for the denoised data in that improve the SNR while preserving the useful signal are
comparison to the existing methods. required for better interpretation of earth layers’ information.
Index Terms— Denoising, double sparse dictionary learning, In past decades, various studies have been conducted
erratic noise, median filtering. assuming that the noise is random with Gaussian distribution
and various seismic data denoising methods were developed
I. I NTRODUCTION in [3]–[8]. However, these denoising methods are not robust in

T HE knowledge of the earth subsurface layers is required

for the exploration of oil, gas, water, and other min-
eral resources. Seismic exploration is a cost-effective method
the presence of non-Gaussian distributed erratic noises. There-
fore, various methods were proposed in literature to attenuate
the erratic noise based on filtering and thresholding [9], rank
to identify the earth subsurface layers’ information upto a reduction [10]–[13] and sparsity-promoting filters [2], [14].
certain depth [1]. The seismic data recorded using seismic In [9], a f -x projection filtering (FXPF)-based method
exploration methods are afflicted by noise. There are two was studied to attenuate both random and NGN with hybrid
types of noise present in seismic data: coherent noise and l1 /l2 norm. In FXPF method, alternating minimization scheme
incoherent noise. Coherent noise occurs due to the multiple was used, where the estimation of prediction error filter
reflections of the source wavelet between layers [2]. However, and estimation of noise sequence was done iteratively until
the tolerance was met. However, the FXPF method has
Manuscript received November 18, 2021; revised February 7, 2022;
accepted March 15, 2022. Date of publication April 4, 2022; date of current two major limitations: high-computational cost and artificial
version April 28, 2022. This work was supported by the Science and events in the projection filter results. In [10], robust singular
Engineering Research Board, Department of Science and Technology, India, spectrum analysis (RSSA) algorithm based on M-estimate
under the Core Research Grant CRG/2019/001234. (Corresponding author:
Karthikeyan Elumalai.) formulation was proposed. The RSSA is a rank reduction
Lakshmi Kuruguntla, Vineela Chandra Dodda, and Karthikeyan Elumalai method for random noise attenuation in which the seismic data
are with the Department of Electronics and Communication Engineering, are represented as a Hankel matrix. Rank reduction of the
SRM University, Amaravati, Andhra Pradesh 522502, India (e-mail:
[email protected]; [email protected]; Hankel matrix followed by anti-diagonal averaging provides
[email protected]). the denoised data. In conventional singular spectrum analy-
Anup Kumar Mandpura is with the Department of Electrical Engineer- sis (SSA) method, a matrix factorization technique named
ing, Delhi Technological University, New Delhi 110042, India (e-mail:
[email protected]). as truncated singular value decomposition (TSVD) was used
Digital Object Identifier 10.1109/TGRS.2022.3164460 based on least square (LS) minimization. However, the TSVD
1558-0644 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 5913613 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

method was sensitive to NGN. Hence RSSA method replaced the sparsity-promoting methods with curvelet transform are
the LS estimate with robust M-estimate equations. In RSSA not suitable for field data due to a fixed basis. Hence
method, bisquare loss function also known as Tukey loss func- Chen et al. [17] further proposed a data-driven constrained
tion was used as a robust M-estimate. An iterative reweighted method (DDCM) to overcome the problems in RCS model.
least square (IRLS) method along with alternating minimiza- In DDCM, data-driven tight frame (DDTF) was used as
tion was implemented in RSSA for numerical computation of a dictionary and the Huber loss function was utilized in
robust matrix factorization. However, the computational cost the optimization problem for dictionary learning (DL). The
of RSSA algorithm is high compared to conventional SSA optimization problem of DL method was solved by an iterative
method. method which consists of two steps. In the first step, sparse
Later in [11], a robust principle component analysis (RPCA) coefficients were determined by hard thresholding algorithm.
method was proposed with two different constraints for recov- A selective threshold value based on noise level (or) sparsity
ery of data from NGN. In RPCA method, NGN is represented level is considered. The second step was dictionary update step
with sparse matrix and the data in f -x domain are represented with the assumption of fixed sparse coefficients. However, the
with a low-rank matrix. The RPCA method adapted nuclear optimization problem was non-linear due to Huber criterion.
norm constraint and l1 norm constraint simultaneously to Hence Chen et al. [17] used pseudo data rather than the
separate the signal from erratic noise. Besides these two original noisy data in the minimization problem. The denoised
norms, Frobenius norm was also used for random noise data from constrained DDTF method had an improved signal
suppression. The cost function with these norms is divided into quality when compared to conventional DDTF method and are
two subproblems and solved separately based on sub-gradient robust to NGN. However, the computational cost of DDCM
method. The denoised data of RPCA method have shown is higher than RISTA-based seismic denoising.
improved SNR with the attenuation of both random noise and Later Zhao et al. [14] proposed an iterative robust sparsity-
NGN. However, a proper assumption of rank is required in promoting filter (IRSPF) algorithm for erratic noise attenuation
RPCA. To address this issue, Akhondi-Asl and Nelson [13] with good preservation of signal. In IRSPF algorithm, LS
proposed a M-estimate robust PCA (MRPCA) method for criterion was replaced by Huber criterion with sparse domain
seismic data denoising with M-estimate penalty function. thresholding strategy. Moreover, curvelet transform was used
In MRPCA, Huber loss function was used as loss function as a sparsity-promoting filter with soft thresholding strategy.
and the optimization problem was solved with alternating The soft thresholding finds the sparse coefficients with respect
direction method of multipliers (ADMM). The denoised results to a threshold which was determined based on noisy data
with MRPCA have a marked improvement in SNR when or desired sparsity level. The threshold selection plays a
compared to RPCA. However, proper tuning of parameters vital role in sparsity-based denoising algorithms which impact
is necessary for Huber loss function to get the denoised data the results greatly. Henceforth, a robust standard deviation
with improved SNR. was selected as a threshold and it was estimated by median
Further, Zhang et al. [12] proposed a robust low-rank absolute deviation (MAD). Another important parameter in
approximation (RLA) method for incoherent noise suppression IRSPF algorithm is the stopping criteria which were selected
with a mixed model which consists of both nuclear norm and empirically to stop the computation. Like [16], IRSPF algo-
l1 norm. In RLA method, TSVD method was replaced by an rithm also used the Huber criterion which was robust to
orthonormal subspace learning (OSL) for matrix factorization. erratic noise and the denoising performance results were
The RLA method was shown to suppress the noncoherent better than FX-prediction filtering method and RSSA method.
noise effectively in comparison to RPCA method. However, However, the computational cost is very high when compared
errors are observed in the antidiagonal averaging process of to conventional sparse domain thresholding filters.
RLA method, and therefore outliers are present in the denoised Thereafter Carozzi and Sacchi [18] proposed a parallel
data. matrix factorization (PMF) algorithm with robust measures
Furthermore, Zhou and Zhang [15] proposed a denoising of misfit for suppression of erratic noise. In PMF algorithm,
method to attenuate random noise and spiky noise based l1 /l2 error norm was used as a robust misfit and the algorithm
on nuclear norm minimization (NNM) problem and trace was implemented iteratively with an alternation of two stages.
prediction strategy. The NNM problem was solved by a In the first stage, new estimate of data was obtained and in
singular value thresholding (SVT) algorithm. However, the the second step, rank reduction was done with tensor unfolding
NNM model must satisfy the assumption that the origi- of new estimated data. In PMF method, rank reduction was
nal data are low-rank matrix. Afterward in [16], a robust performed with randomized QR-decomposition to simplify the
compressive sensing (RCS) reconstruction model was pro- parameter selection. However, the selection of parameters such
posed to suppress the random noise and NGN. Based on as tensor rank and standard deviation of noise are important
the compressive sensing model, authors developed a robust for reconstruction and denoising of multidimensional seismic
iterative shrinkage thresholding algorithm (RISTA) to recon- data. Further in [2], an extension to IRSPF algorithm [14]
struct the seismic data. In RCS reconstruction model, robust was proposed with structure-oriented space-varying median
Huber loss function was selected instead of LS loss function. filtering step which attenuates the erratic noise in complex-
In RISTA, sparsity-promoting thresholding transform plays a structured seismic data.
vital role in denoising. A curvelet transform was used which Later in [19], a robust dictionary learning (RDL) method
separates the useful signal from an erratic noise. However, was proposed for reconstruction of seismic data. In general,

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 KURUGUNTLA et al.: ERRATIC NOISE ATTENUATION USING DOUBLE SPARSITY DICTIONARY LEARNING METHOD 5913613

the basis or atoms in the adaptive dictionary are learned from cumulative noise in the seismic trace can be expressed as
noisy data by minimizing the error. In RDL method, DDTF
N(n) = G(n) + E(n). (2)
was used as a dictionary and projected operator (PO) with
Huber misfit and was applied to attenuate the erratic noise with The recorded seismic trace X (n) is mathematically
high amplitude levels. Hence, the RDL method is also called expressed as
as projected DDTF (PDDTF). The PO converts the nonlinear
Huber misfit to linear with the use of sub-gradients of residual X (n) = S(n) + N(n). (3)
in the optimization problem. PO provides the restriction of Let D ∈ R L×M be the seismic data that consist of M seismic
abnormal portions in data so that it is robust to erratic noise. traces given as
However, the computational cost of RDL method is high.
The transform-domain thresholding-based filtering methods D = [S1 , S2 , . . . , S M ] (4)
work best for random noise attenuation. However, in this
where Si ∈ R L×1
is the noise-free trace. The noisy seismic
work, we have considered the presence of random noise in
data with M traces can be mathematically represented as
the data along with erratic noise which is characterized by
sudden large amplitude values. We extend the method in [2] X = D+N (5)
by introducing the additional sparsity by K-SVD to make
it double sparse. Iterative structure-oriented space-varying where D indicates the noiseless data and N denotes the noise
median filter (ISOSVMF) used in the first step is able to matrix.
attenuate erratic noise; however, there is some residual noise
remaining. In the second step, we use an adaptive DL method B. Huber Criterion
for the removal of the residual noise. In adaptive DL method, In conventional denoising methods, LS criterion is com-
the atoms in the dictionary are learned from the training data; monly used for error minimization between the denoised
therefore, the useful seismic signal is preserved and we obtain data and true data. However, LS criterion is effective in the
a low-signal leakage in the final denoised result. The K-SVD attenuation of random noise with Gaussian distribution and is
method is used as an adaptive DL method which can represent sensitive to non-Gaussian (or) erratic noise. Huber criterion
the seismic data in sparse form with an explicit dictionary. The is a robust error minimization criterion in comparison to LS
atoms in the dictionary are updated sequentially by singular criterion and provides a better attenuation of erratic noise [21].
value decomposition and provide high efficiency in the random Let e indicates error between the denoised data and true data,
noise attenuation compared to analytic transforms [20]. and t be a tuning parameter, the Huber criterion is then defined
Rest of this article is organized as follows. In Section II, as follows:
we provide the mathematical preliminaries which includes ⎧
⎨1 2
signal model in Section II-A, Huber criterion in Section II-B, e , if |e| ≤ t
ρt (e) = 2 (6)
robust sparsity-promoting filter in Section II-C and dou- ⎩t (|e| − t), if |e| > t.
ble sparsity dictionary learning (DSDL) in Section II-D.
Section III describes our proposed method for denoising noisy The derivative of Huber criterion ρ(e) is defined as
seismic data. In Section IV, we compared our proposed 
e, if |e| ≤ t
method on two synthetic and field data examples with existing φt (e) = (7)
methods and the results are given in Tables I, II, and IV, tsign(e), if |e| > t.
respectively. Finally, conclusion is given in Section V. The value of tuning parameter acts as threshold which is
selected empirically as t = 1.345σ [22], where σ is the
II. M ATHEMATICAL P RELIMINARIES standard deviation (SD) of noise. The SD of noise is estimated
In this section, we provide the mathematical preliminaries based on median absolute deviation (MAD) [23] which can be
that are used in the subsequent sections. expressed as
MAD median(|e − median(e)|)
A. Signal Model σ ≈ = . (8)
0.6745 0.6745
Assuming r (n) is the reflection coefficients and w(n) is the
seismic wavelet. The seismic trace S(n) is the convolution of C. Robust Sparsity-Promoting Filter
seismic wavelet and reflection coefficient and is mathemati- Sparse representation of a seismic signal plays a vital role
cally expressed as follows: in denoising application to attenuate the noise present in the
signal. Wavelets [24], curvelets [25], shearlets [4], contourlets
S(n) = r (n) ∗ w(n) (1)
[5], and seislets [26] are examples of sparsity-promoting filters
where n is the discrete time index and ∗ indicates the convo- that have been used to attenuate the random noise in seismic
lution operator. data based on LS criterion. However, those are not robust for
Let N(n) be the additive noise in the seismic trace during erratic noise attenuation due to that LS criterion is sensitive for
the acquisition process. The noise component with Gaussian erratic noise [10], [16]. Therefore, robust sparsity-promoting
distribution is denoted by G(n) and the non-Gaussian dis- filter was used in [14]. In robust sparsity-promoting filter,
tributed erratic noise component is denoted by E(n). The LS criterion was replaced by Huber criterion.

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 5913613 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 1. Block diagram for the proposed DSDL method.

Let B be an analytic transform used as a sparsity-promoting

filter to attenuate the additive noise N in the observed seismic
data X given in (5). The denoising model with robust sparsity-
promoting filter can be mathematically expressed as follows:

d̂ = ρt (X − d) + λBd1 (9)

where d denotes the denoised data, ρt (X − d) is the Huber

norm of (X − d) which is defined in (6). The above opti-
mization problem with Huber norm criterion is non-linear and
it is difficult to obtain a solution. Hence, in [27], theoret-
Fig. 2. Synthetic seismic data-1. (a) Clean synthetic seismic data-1. (b) Noisy
ical pseudodata were introduced to transform the nonlinear data.
optimization problem into a linear LS optimization problem.
The pseudodata corresponding to a seismic data X are defined
in [14] as Therefore, the sparse representation coefficients of observed
 signal based on double sparsity can be expressed as
X, if |e| ≤ t
P= (10) Â = arg min  A0
d̂ + tsign(R), if |e| > t A
1
where R represents the removed noise obtained from sparsity- s.t. X − H A22  η (13)
2
promoting filter and t is the tuning parameter. Therefore, where η is error threshold which is proportional to noise level.
changing the noisy data into pseudodata is important to
attenuate the erratic noise using linear LS optimization.
E. Local Similarity
In [14], an iterative robust sparsity-promoting filter was
proposed with the Huber norm criterion and pseudodata The local similarity measure was defined by Fomel in [30].
concept. Let c(t) and d(t) be two signals, then the global correlation
coefficient is defined as
c(t), d(t)
D. Double Sparsity Dictionary Learning β= √ (14)
c(t), c(t)d(t), d(t)
DSDL method provides double sparsity using both analytic where c(t), d(t) is the dot product between two signals c(t)
and adaptive transforms [28]. The DSDL method is better and d(t). The above equation provides a single value for
than an analytic transform or an adaptive transform and the whole signal and is not suitable for the analysis of local
provides better efficiency and adaptability for seismic data changes in the signal. Let c and d be the vector notation for
denoising [29]. In DSDL method, the sparsity is implemented the signals c(t) and d(t), respectively. β1 and β2 be the LS
by a cascade combination of analytic transform and adaptive inverses of vectors c and d, respectively, and are given by
transform. Let B be an analytic transform and W an adaptive
transform, then the transform for DSDL method denoted by β1 = (c T c)−1 (c T d)
H is defined as β2 = (d T d)−1 (d T c). (15)

H = BW. (11) From [31, eq. (10)] the squared correlation coefficient is
defined as
The noisy seismic data in terms of double sparsity dictionary β 2 = β1 β2 . (16)
is formulated as
Using shaping regularization, the LS inverses in scalar notation
X = H A + N. (12) are transformed into vectors and localized. Let S be the

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 KURUGUNTLA et al.: ERRATIC NOISE ATTENUATION USING DOUBLE SPARSITY DICTIONARY LEARNING METHOD 5913613

Fig. 3. Comparative results of proposed DSDL method with state-of-the-art denoising methods on synthetic seismic data-1: Denoised data using (a) curvelet
method, (b) iterative curvelet method, (c) single-step median filtering method, (d) iterative structure-oriented space-varying median filtering with sparsity
constraint, and (e) proposed DSDL method; noise section for (f) curvelet method, (g) iterative curvelet method, (h) single-step median filtering method,
(i) iterative structure-oriented space-varying median filtering with sparsity constraint, and (j) proposed DSDL method; and local similarity map (k) between
(a) and (f), (l) between (b) and (g), (m) between (c) and (h), (n) between (d) and (i), and (o) between (e) and (j).

shaping regularization operator, λ be a scaling parameter, C robust sparsity-promoting filter is chosen as an analytic trans-
and D are two diagonal operators derived from elements of c form and K-SVD is chosen as an adaptive transform which
and d, respectively. Then the vector notation of β1 and β2 is includes sparse coding and dictionary update. We propose
denoted as e1 and e2 a two-step method in which first we perform the denoising
of seismic data X using ISOSVMF with sparsity-promoting
e1 = [λ2 I + S(C T C − λ2 I )]−1 SC T d curvelet transform [2]. This step is repeated till the stopping
e2 = [λ2 I + S(D T D − λ2 I )]−1 S D T c. (17) criterion is reached. In the second step, we again perform
the residual noise attenuation based on K-SVD method with
The componentwise multiplication of e1 and e2 gives the local a dictionary initialized from the obtained denoised data
correlation of signals also known as local similarity and is of ISOSVMF; therefore, our method is termed as double
mathematically expressed as [8] sparsity DL.

The block diagram of the proposed method is shown in
βlocal = e1T e2 . (18)
Fig. 1. Let F be a fast discrete curvelet transform-based
sparsity-promoting filter, d be the denoised data obtained
from robust sparsity-promoting filter defined in (9) and di
III. P ROPOSED M ETHOD
be the denoised data at i th iteration. The denoised data
In this section, we present our proposed DSDL method to for first iteration is initialized as zero, i.e., d1 = 0
denoise the seismic data corrupted with random and erratic and pi indicates pseudodata which is updated and con-
noise. In our method, an ISOSVMF [31] constrained with strained by ISOSVMF at every iteration. The iteration in

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 5913613 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 4. Comparative results with highlighted portions of proposed DSDL method with state-of-the-art denoising methods on synthetic seismic data-1. (a) Clean
data and its highlighted portion. (b) Noisy data and its highlighted portion. Denoised data and highlighted portions of (c) curvelet method, (d) iterative curvelet
method, (e) single-step median filtering method, (f) iterative structure-oriented space-varying median filtering with sparsity constraint, and (g) proposed DSDL
method.

where i denotes the iterations from 1, 2, . . . , N, and γ is the

regularization parameter.
The stopping criteria of the first step is given by
di+1 − di  < (20)
where is defined as [32]
√
= Gσ M. (21)
Here G is the gain factor and is considered as 1.15 [33], M
represents the size of noisy seismic data X and σ is noise
level defined in (8).
The removed noise R is defined as the difference of noisy
Fig. 5. Synthetic seismic data-2. (a) Clean synthetic seismic data-2. (b) Noisy
data. data X and denoised data di as follows:
R = X − di . (22)
the first step of the proposed method can be mathematically
expressed as The updated pseudodata pi is given by
1
di+1 = arg min  pi − di 22 + γ Fdi 1 (19) pi = di + φt (R) (23)
di 2
Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 KURUGUNTLA et al.: ERRATIC NOISE ATTENUATION USING DOUBLE SPARSITY DICTIONARY LEARNING METHOD 5913613

Fig. 6. Comparative results of proposed DSDL method with state-of-the-art denoising methods on synthetic seismic data-2: Denoised data using (a) curvelet
method, (b) iterative curvelet method, (c) single-step median filtering method, (d) iterative structure-oriented space-varying median filtering with sparsity
constraint, and (e) proposed DSDL method; noise section for (f) curvelet method, (g) iterative curvelet method, (h) single-step median filtering method,
(i) iterative structure-oriented space-varying median filtering with sparsity constraint, and (j) proposed DSDL method; and local similarity map (k) between
(a) and (f), (l) between (b) and (g), (m) between (c) and (h), (n) between (d) and (i), and (o) between (e) and (j).

where φt is defined in (7). Further, the updated pseudodata pi where Ŵ , Â are the desired adaptive dictionary matrix and
is constrained by ISOSVMF can be expressed as sparse coefficient matrix, respectively, A is the sparse coef-
ficients matrix consists of V number of columns and each
pi = C( pi ) (24)
column is represented as αi with i = 1, 2, . . . , V . Where as
where C is the constraining operator. The updated pseudodata a j represents the columns in adaptive dictionary with j =
in the last iteration of ISOSVMF is denoted as U . 1, 2, . . . , M. Here s and l represent the sparsity levels in each
In the second step of proposed method, an overcomplete column of sparse coefficient matrix and adaptive dictionary
dictionary B ∈ R E×F is obtained from the denoised data matrix, respectively.
of the first step. The dictionary B is used to initialize the The sparse coefficient matrix in (25) is represented by the
dictionary for K-SVD. Different from conventional K-SVD optimization problem also called sparse coding problem can
method, in our proposed method sparse coding and dictionary be expressed as
update are performed alternatively on updated pseudodata 1
rather than noisy data. The learned dictionary updated by  = arg min U − BW A22 + λ A1 . (26)
A 2
K-SVD method [32] is represented as W . The dictionary
update step with an intial dictionary B and learned dictionary By substituting (11) in (26), the optimization problem for
W on updated pseudodata U using K-SVD method is given sparse coding is given as
as 1
 = arg min U − H A22 + λ A1 . (27)
(Ŵ , Â) = arg min U − BW A22 A 2
W,A
 The above equation is solved using the orthogonal matching
αi  ≤ s ∀i pursuit (OMP) algorithm [34] over the dictionary H = BW .
s.t.  0   (25)
a j  ≤ l,  Ba j  = 1 ∀ j OMP is a iterative greedy algorithm that selects the atom
0 2

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 5913613 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 7. Zoomed section of denoised data for (a) curvelet method, (b) iterative curvelet method, (c) single-step median filtering method, (d) iterative
structure-oriented space-varying median filtering with sparsity constraint, and (e) proposed DSDL method.

in the dictionary at each iteration and correlates with the

updated residuals until the stopping criteria. Later, estimated
sparse coefficient matrix  is fixed and atoms in dictionary W
are updated sequentially through singular value decomposition
(SVD). The updated dictionary matrix in (25) can be solved
using the optimization problem expressed as
 2
Ŵ = arg min U − BW Â F . (28)
W

The sparse coding and the dictionary update are performed

alternatively upto the number of iterations met and the
denoised data of our proposed DSDL method is equal to the
linear combination of few atoms of updated dictionary B Ŵ
Fig. 8. Field datasets with random and erratic noise. (a) Field data-1. (b) Field
and sparse coefficient matrix Â. data-2.

IV. R ESULTS AND D ISCUSSION field datasets with four case studies. The performance metrics
In this section, we demonstrate the performance of our used for our method are SNR, mean square error (MSE) of
proposed method described in Section III on two synthetic and the denoised seismic signal, and local similarity map between

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 KURUGUNTLA et al.: ERRATIC NOISE ATTENUATION USING DOUBLE SPARSITY DICTIONARY LEARNING METHOD 5913613

Fig. 9. Comparative results of proposed DSDL method with state-of-the-art denoising methods on field data-1: Denoised data using (a) curvelet method,
(b) iterative curvelet method, (c) single-step median filtering method, (d) iterative structure-oriented space-varying median filtering with sparsity constraint, and
(e) proposed DSDL method; removed noise section for (f) curvelet method, (g) iterative curvelet method, (h) single-step median filtering method, (i) iterative
structure-oriented space-varying median filtering with sparsity constraint, and (j) proposed DSDL method; and local similarity map (k) between (a) and (f),
(l) between (b) and (g), (m) between (c) and (h), (n) between (d) and (i), and (o) between (e) and (j).

removed noise section and denoised data with energy values. The formula to calculate energy value is given as [35]
The SNR of denoised seismic signal is expressed as
 E = mean(mean(βlocal )) (31)
t (s (t)
2
SNR = 10 log10 (29) where βlocal is defined in (18).
t (x(t) − s(t))
2

where s(t) represents noise-free trace and x(t) is the noisy A. Case Study 1: Comparison of Proposed Method With
ˆ ) for
trace in seismic data. The MSE of the denoised trace S(i Existing Methods on Synthetic Seismic Data-1
N number of traces is formulated as
In this case study, we compared the performance of
N  2
1 our proposed method with four existing denoising methods
MSE = s(i ) − s(iˆ ) (30) based on curvelet transform [36], iterative curvelet trans-
N i=1
form [14], single-step median filtering [31], and iterative
where i indicates the trace number and s(i ) is the noise less structure-oriented space-varying median filtering with sparsity
trace. constraint [2]. We generate the synthetic seismic data-I with
Local similarity metric is used to measure the signal leak- linear events like in [2]. The synthetic seismic data-1 consists
age by calculating the energy value from similarity map of of 50 traces with sampling interval of 4 ms. Fig. 2(a) and (b)
denoised section and noise section [8], [8], [26]. The local shows the clean and noisy synthetic seismic data-I consisting
similarity between two signals is defined in Section II-E. of four linear events. The clean data are corrupted with
The energy value is calculated from the local simi- both random and erratic noise with high amplitude peaks
larity between denoised data and removed noise section. with resulting SNR of −1.25 dB. The proposed method

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 5913613 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 10. Comparative results of proposed DSDL method with state-of-the-art denoising methods on field data-2: Denoised data using (a) curvelet method,
(b) iterative curvelet method, (c) single-step median filtering method, (d) iterative structure-oriented space-varying median filtering with sparsity constraint, and
(e) proposed DSDL method; removed noise section for (f) curvelet method, (g) iterative curvelet method, (h) single-step median filtering method, (i) iterative
structure-oriented space-varying median filtering with sparsity constraint, and (j) proposed DSDL method; and local similarity map (k) between (a) and (f),
(l) between (b) and (g), (m) between (c) and (h), (n) between (d) and (i), and (o) between (e) and (j).

is applied on noisy seismic data for denoising using (25). TABLE I

The performance of the state-of-the-art denoising methods on C OMPARISON OF P ROPOSED M ETHOD ON S YNTHETIC D ATA -1 W ITH
S TATE - OF - THE -A RT D ENOISING M ETHODS IN T ERMS OF
noisy synthetic seismic data-1 for erratic and random noise SNR, MSE, AND E NERGY
attenuation is tabulated in Table I. From Table I, it can be
observed that the existing methods based on curvelet, iterative
curvelet, single-step median filtering, and ISOSVMF with
sparsity constraint improves the SNR of seismic data to 7.37,
8.55, 5.81, and 9.61 dB, respectively. The proposed method
based on double sparsity improves the SNR of seismic data
to 16.38 dB. In addition, we also evaluated the energy values
from the local similarity maps.
From Table I, we observe that the proposed method based
on double sparsity achieves a high SNR, lower MSE, and methods are shown in Fig. 3(f)–(j), respectively. From the
lower energy values in comparison to the other four meth- removed noise sections, we observe that all the methods other
ods. The denoised data of curvelet, iterative curvelet, single- than proposed method shown some signal components. Based
step median filtering, ISOSVMF with sparsity constraint and on denoised data and noise section, we also plotted the local
proposed methods for synthetic seismic data-1 are shown in similarity maps for each method in Fig. 3(k)–(o). In Fig. 3(o),
Fig. 3(a)–(e), respectively. Among all the methods, denoised we observe that the signal leakage is very low for the proposed
data of the proposed method are better with good signal preser- method and it recovers the linear events accurately from
vation. The removed noise sections of all the above-mentioned the erratic noisy data as compared to existing methods. The

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 KURUGUNTLA et al.: ERRATIC NOISE ATTENUATION USING DOUBLE SPARSITY DICTIONARY LEARNING METHOD 5913613

TABLE II TABLE III

C OMPARISON OF P ROPOSED M ETHOD ON S YNTHETIC D ATA -2 W ITH C OMPARISON OF P ROPOSED M ETHOD W ITH S TATE - OF - THE -A RT
S TATE - OF - THE -A RT D ENOISING M ETHODS IN T ERMS OF D ENOISING M ETHODS FOR VARIOUS I NPUT SNR VALUES
SNR, MSE, AND E NERGY ON S YNTHETIC D ATA -1

highlighted portions of clean synthetic data-1, noisy data,m

and denoised results of all methods are shown in Fig. 4(a)–(g).

B. Case Study 2: Comparison of Proposed Method With

Existing Methods on Synthetic Seismic Data-2
In this case study, we generated the data with hyper-
bolic events which contains 128 traces with sampling inter-
val of 4 ms similar to [37]. The noise-free synthetic seismic
data-2 with five layers of hyperbolic events is shown in
Fig. 5(a). Random noise and erratic noise are added to the
noise-free seismic data to get the noisy data with SNR of
6.57 dB shown in Fig. 5(b). The proposed method is applied
on the noisy data having hyperbolic events and the denoising
performance is analyzed. The denoised results of proposed
method are also compared with the existing denoising meth-
ods: curvelet [36], iterative curvelet [14], single-step median
filtering [31], and ISOSVMF with sparsity constraint [2]. The
SNR, MSE, and energy of the denoised data are tabulated in
Table II.
From Table II, we observe that the proposed method based
on double sparsity improves the SNR and lower the MSE
and energy values of the denoised data in comparison to the
existing methods. The denoised data of curvelet [36], iterative
curvelet [14], single-step median filtering [31], ISOSVMF The denoised data metrics such as SNR, MSE, and energy
with sparsity constraint [2], and proposed method on noisy values for each method are tabulated in Table III. From
seismic data-2 are shown in Fig. 6(a)–(e), respectively. The Table III, we observed that the proposed method improves the
magnified versions of these denoised data plots with one layer SNR to high value as compared to other methods. The other
are shown in Fig. 7(a)–(e), respectively. We also provide the two metrics MSE and energy also very low for the proposed
noise sections and local similarity maps for each method method.
shown in Fig. 6(f)–(o), respectively. From the denoised data
plots and local similarity maps, we observe that the hyperbolic
D. Case Study 4: Application of Proposed Method on Field
events are clearly visible in the proposed method with high
Data
noise attenuation and very low-signal energy leakage. There-
fore, the erratic noise attenuation using double sparsity has In this case study, we analyze the feasibility of our proposed
a better performance when compared to the existing methods method on two field datasets which are corrupted by both ran-
(curvelet, iterative curvelet, single-step median filtering, and dom and erratic noise taken from [2]. The field data-1 consist
ISOSVMF with sparsity constraint). of 800 time samples and 220 traces with sampling rate of
4 ms shown in Fig. 8(a). The field data-2 consists of 256 time
samples and 256 traces with sampling interval of 4 ms is
C. Case Study 3: Comparative Analysis of Proposed Method shown in Fig. 8(b). We compare the performance of pro-
on Synthetic Data-1 With Different SNR Values posed method with the existing methods(curvelet [36], iterative
In this case study, we consider the noisy data with three curvelet [14], single-step median filtering [31], ISOSVMF
different SNR levels (−9, −3, and 0.3 dB) and perform with sparsity constraint [2]) on the two field datasets. For
the erratic and random noise attenuation using curvelet field data, SNR and MSE are not evaluated as noise-free
[36], iterative curvelet [14], single-step median filtering [31], data. Hence, we use the local similarity map between the
ISOSVMF with sparsity constraint [2], and proposed method. noise sections and denoised sections, and energy values are

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 5913613 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

TABLE IV [3] M. Xiong, Y. Li, and N. Wu, “Random-noise attenuation for seismic
C OMPARISON OF P ROPOSED M ETHOD ON F IELD D ATASETS W ITH data by local parallel radial-trace TFPF,” IEEE Trans. Geosci. Remote
S TATE - OF - THE -A RT D ENOISING M ETHODS IN T ERMS OF E NERGY Sens., vol. 52, no. 7, pp. 4025–4031, Jul. 2014.
[4] D. Kong, Z. Peng, Y. He, and H. Fan, “Seismic random noise attenuation
using directional total variation in the shearlet domain,” J. Seismic
Explor., vol. 25, no. 4, pp. 321–338, 2016.
[5] K. Li, J. Gao, W. Wang, and Y. Yang, “Local angle extraction and noise
attenuation for seismic image using contourlet transform,” in Proc. Int.
Workshop Educ. Technol. Training Int. Workshop Geosci. Remote Sens.,
vol. 1, Dec. 2008, pp. 349–352.
[6] P. Turquais, E. G. Asgedom, W. Söllner, and E. Otnes, “Dictionary learn-
ing for signal-to-noise ratio enhancement,” in SEG Technical Program
Expanded Abstracts. Tulsa, OK, USA: Soc. Explor. Geophysicists, 2015,
evaluated. The energy values of field data-1 and field data-2 pp. 4698–4702.
[7] D. Xu, L. Han, D. Liu, and Y. Wei, “Random seismic noise attenuation
are denoted by E1 and E2, respectively, are given in Table IV. by learning-type overcomplete dictionary based on K-singular value
From Table IV, we observe that the proposed method achieves decomposition algorithm,” Global Geol., vol. 19, no. 1, pp. 55–60,
less energy value for the two field datasets. Therefore, our 2016.
proposed method has less signal leakage compared to existing [8] Y. Chen and S. Fomel, “Random noise attenuation using local
signal-and-noise orthogonalization,” Geophysics, vol. 80, no. 6,
methods. The denoised data sections of each method on field pp. WD1–WD9, Nov. 2015.
data-I are shown in Fig. 9(a)–(e). The respective noise sections [9] K. Chen and M. D. Sacchi, “Robust f-x projection filtering for simulta-
are shown in Fig. 9(f)–(j). Local similarity maps of field data-1 neous random and erratic seismic noise attenuation,” Geophys. Prospect-
ing, vol. 65, no. 3, pp. 650–668, May 2017.
are shown in Fig. 9(k)–(o). From the noise sections and local [10] K. Chen and M. D. Sacchi, “Robust reduced-rank filtering for erratic
similarity maps, we clearly observed that signal leakage is very seismic noise attenuation,” Geophysics, vol. 80, no. 1, pp. V1–V11,
less in proposed method compared to other methods. Jan. 2015.
[11] J. Cheng, K. Chen, and M. D. Sacchi, “Application of robust princi-
Similarly, the denoised sections of field data-2 are shown pal component analysis (RPCA) to suppress erratic noise in seismic
in Fig. 10(a)–(e). Noise sections are shown in Fig. 10(f)–(j) records,” in SEG Technical Program Expanded Abstracts Tulsa, OK,
and local similarity maps are shown in Fig. 10(k)–(o). For this USA: Soc. Explor. Geophysicists, 2015, pp. 4646–4651.
field data also, the proposed method has less signal leakage. [12] M. Zhang, Y. Liu, H. Zhang, and Y. Chen, “Incoherent noise suppression
of seismic data based on robust low-rank approximation,” IEEE Trans.
Geosci. Remote Sens., vol. 58, no. 12, pp. 8874–8887, Dec. 2020.
[13] H. Akhondi-Asl and J. D. B. Nelson, “M-estimate robust PCA for
V. C ONCLUSION seismic noise attenuation,” in Proc. IEEE Int. Conf. Image Process.
(ICIP), Sep. 2016, pp. 1853–1857.
In this work, we proposed a novel method for attenuation [14] Q. Zhao, Q. Du, X. Gong, and Y. Chen, “Signal-preserving erratic noise
of erratic and random noise based on double sparsity DL attenuation via iterative robust sparsity-promoting filter,” IEEE Trans.
method. In our method, we initialized the dictionary from Geosci. Remote Sens., vol. 56, no. 6, pp. 3547–3560, Apr. 2018.
[15] Y. Zhou and S. Zhang, “Robust noise attenuation based on nuclear
the denoised data of ISOSVMF method with sparsity con- norm minimization and a trace prediction strategy,” J. Appl. Geophys.,
straint by curvelet transform. We again perform residual noise vol. 147, pp. 52–67, Dec. 2017.
attenuation on pseudodata updated in ISOVMF method with [16] Q. Zhao and Q. Z. Du, “Robust reconstruction of seismic data with
new sparsity constraint based on adaptive dictionary such as random and erratic noise,” in Proc. 79th EAGE Conf. Exhib., no. 1,
2017, pp. 1–5.
K-SVD. We tested the proposed double sparsity DL method [17] Y. Chen, M. Bai, and Y. Chen, “Obtaining free USArray data by multi-
on synthetic and field datasets. The results demonstrated that dimensional seismic reconstruction,” Nature Commun., vol. 10, no. 1,
the proposed method achieves better denoising performance in pp. 1–13, 2019.
[18] F. Carozzi and M. D. Sacchi, “Robust tensor-completion algorithm
terms of SNR, MSE, and energy as compared to the existing for 5D seismic-data reconstruction,” Geophysics, vol. 84, no. 2,
methods. pp. V97–V109, Mar. 2019.
[19] Q. Zhao, Q. Du, Q. Li, and L. Fu, “Robust dictionary learning for erratic
noise-corrupted seismic data reconstruction,” Acta Geophysica, vol. 68,
ACKNOWLEDGMENT no. 3, pp. 687–700, Jun. 2020.
[20] L. Zhu, E. Liu, and J. H. McClellan, “Seismic data denoising through
The authors would like to give special thanks to Prof. multiscale and sparsity-promoting dictionary learning,” Geophysics,
vol. 80, no. 6, pp. WD45–WD57, Nov. 2015.
Yangkang Chen and his research group at The University of
[21] R. K. W. Wong and T. C. M. Lee, “Matrix completion with noisy entries
Texas at Austin, Austin, TX, USA, for their great help in pro- and outliers,” J. Mach. Learn. Res., vol. 18, no. 1, pp. 5404–5428, 2017.
viding research data, professional suggestions, and comments. [22] P. J. Huber, “Robust estimation of a location parameter,” in
They are also thankful to the anonymous reviewers for their Breakthroughs in Statistics. New York, NY, USA: Springer, 1992,
pp. 492–518.
invaluable suggestions and comments. [23] P. W. Holland and R. E. Welsch, “Robust regression using iteratively
reweighted least-squares,” Commun. Statist.-Theory Methods, vol. 6,
no. 9, pp. 813–827, 1977.
R EFERENCES [24] G. Peyré, A Wavelet Tour of Signal Processing: The Sparse Way.
[1] D. W. Steeples and R. D. Miller, “Seismic reflection methods applied New York, NY, USA: Academic, 2009.
to engineering, environmental, and ground-water problems,” in Proc. [25] J. Ma and G. Plonka, “The curvelet transform,” IEEE Signal Process.
1st EEGS Symp. Appl. Geophys. Eng. Environ. Problems, 1988, Mag., vol. 27, no. 2, pp. 118–133, Mar. 2010.
pp. 409–461. [26] Y. Chen and S. Fomel, “EMD-seislet transform,” Geophysics, vol. 83,
[2] G. Huang, M. Bai, Q. Zhao, W. Chen, and Y. Chen, “Erratic noise no. 1, pp. A27–A32, Jan. 2018.
suppression using iterative structure-oriented space-varying median fil- [27] H.-S. Oh, D. W. Nychka, and T. C. M. Lee, “The role of pseudo data
tering with sparsity constraint,” Geophys. Prospecting, vol. 69, no. 1, for robust smoothing with application to wavelet regression,” Biometrika,
pp. 101–121, Jan. 2021. vol. 94, no. 4, pp. 893–904, 2007.

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.
 KURUGUNTLA et al.: ERRATIC NOISE ATTENUATION USING DOUBLE SPARSITY DICTIONARY LEARNING METHOD 5913613

[28] R. Rubinstein, M. Zibulevsky, and M. Elad, “Double sparsity: Learning Vineela Chandra Dodda (Graduate Student Mem-
sparse dictionaries for sparse signal approximation,” IEEE Trans. Signal ber, IEEE) received the B.Tech. degree from
Process., vol. 58, no. 3, pp. 1553–1564, Mar. 2010. Amrita Vishwa Vidyapeetham, Coimbatore, India,
[29] Y. Chen, J. Ma, and S. Fomel, “Double-sparsity dictionary for seis- in 2015, and the M.Tech. degree in embedded sys-
mic noise attenuation,” Geophysics, vol. 81, no. 2, pp. V103–V116, tems from KL University, Guntur, India, in 2019.
Mar. 2016. She is currently pursuing the Ph.D. degree with
[30] S. Fomel, “Local seismic attributes,” Geophysics, vol. 72, no. 3, the Department of Electronics and Communication
pp. A29–A33, May 2007. Engineering (ECE), SRM University, Amaravati,
[31] Y. Chen, S. Zu, Y. Wang, and X. Chen, “Deblending of simultaneous Andhra Pradesh, India, under the guidance of
source data using a structure-oriented space-varying median filter,” Dr. Karthikeyan Elumalai.
Geophys. J. Int., vol. 222, no. 3, pp. 1805–1823, Sep. 2020. She has worked as an Associate Software Engineer
[32] M. Aharon, M. Elad, and A. Bruckstein, “K-SVD: An algorithm for with Robert Bosch Engineering and Business Solutions Ltd., Coimbatore,
designing overcomplete dictionaries for sparse representation,” IEEE from 2015 to 2017. Her research interests include neural networks, machine
Trans. Signal Process., vol. 54, no. 11, pp. 4311–4322, Nov. 2006. learning, and seismic data processing.
[33] M. Elad and M. Aharon, “Image denoising via sparse and redundant
representations over learned dictionaries,” IEEE Trans. Image Process.,
vol. 15, no. 12, pp. 3736–3745, Dec. 2006.
[34] Y. Zhou, S. Li, J. Xie, D. Zhang, and Y. Chen, “Sparse dictionary Anup Kumar Mandpura (Member, IEEE) received
learning for seismic noise attenuation using a fast orthogonal matching the B.Tech. degree in electronics engineering from
pursuit algorithm,” J. Seismic Explor., vol. 26, no. 5, pp. 433–454, 2017. Banaras Hindu University, Varanasi, India, in 2006,
[35] O. M. Saad and Y. Chen, “Deep denoising autoencoder for seismic the M.Tech. degree in digital signal processing from
random noise attenuation,” Geophysics, vol. 85, no. 4, pp. V367–V376, IIT Guwahati, Guwahati, India, in 2009, and the
Jul. 2020. Ph.D. degree in wireless communication from IIT
[36] E. Candès, L. Demanet, D. Donoho, and X. Ying, “Fast discrete curvelet Delhi, Delhi, India, in 2018.
transforms,” Multiscale Model. Simul., vol. 5, no. 3, pp. 861–899, He is currently working as an Assistant Profes-
Sep. 2006. sor with the Department of Electrical Engineering,
[37] Y. Zhou, J. Yang, H. Wang, G. Huang, and Y. Chen, “Statistics-guided Delhi Technological University, New Delhi. His
dictionary learning for automatic coherent noise suppression,” IEEE research interests include performance analysis of
Trans. Geosci. Remote Sens., vol. 60 pp. 1–17, 2021. wireless communication systems, energy harvesting communication systems,
microwave/millimeter-wave systems, and signal processing.

Lakshmi Kuruguntla (Graduate Student Member,

IEEE) received the B.Tech. degree from Jawa- Karthikeyan Elumalai (Member, IEEE) received
harlal Nehru Technological University (JNTU), the B.E. and M.E. degrees from Anna University,
Hyderabad, India, in 2008, and the M.Tech. degree Chennai, India, in 2007 and 2009, respectively,
from JNTU, Kakinada, India, in 2012. She is and the Ph.D. degree from IIT Delhi, Delhi, India,
currently pursuing the Ph.D. degree with the in 2018.
Department of Electronics and Communication He is currently working as an Assistant Profes-
Engineering (ECE), SRM University, Amaravati, sor with the Department of Electronics and Com-
Andhra Pradesh, India, under the supervision of munication Engineering (ECE), SRM University,
Dr. Karthikeyan Elumalai. Amaravati, Andhra Pradesh, India. He received a
She has seven years teaching experience as an research grant from the Department of Science
Assistant Professor in JNTUK affiliated colleges. Her research interests and Technology (DST), India. His research interests
include seismic signal processing, compressive sensing, and dictionary include seismic signal processing, machine learning, and seismic modeling
learning. and inversion.

Authorized licensed use limited to: BENEMERITA UNIVERSIDAD AUTONOMA DE PUEBLA. Downloaded on February 16,2025 at 20:43:35 UTC from IEEE Xplore. Restrictions apply.