IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 17, NO.

9, SEPTEMBER 2020 1637

Seismic Random Noise Reduction Using Adaptive

Threshold Combined Scale and Directional
Characteristics of Shearlet Transform
Jicheng Liu , Ya Gu, Yongxin Chou , and Jianfei Gu

Abstract— Random noise attenuation is an important step in The rank of the Hankel matrix corresponds to the number
seismic processing. A method with an adaptive threshold based of seismic data events, and the rank-reduction algorithm is
on scale and directional characteristics of shearlet transform is selected to remove random noise. Singular spectrum analysis
proposed to attenuate seismic random noise. The method can
adaptively calculate threshold according to the decomposed scales (SSA) [11], damped SSA (DSSA) [12], and multistep SSA
and directions, where the directional characteristics are embodied (MSSA) [13], and so on, belong to this kind of methods.
by using the L1 norm of coefficients. Compared to the universal The fourth method is based on signal decomposition. The
threshold and the threshold based on scale methods only, the scale seismic data is decomposed into different components, and
and directional threshold methods can remove random noise and the main components are selected to reconstruct the useful
reserve an effective signal to the utmost extent. The synthetic and
field data examples demonstrate that the proposed method can signal. Empirical mode decomposition (EMD) proposed by
significantly improve the denoising effect and has the stronger Huang et al. [14] can adaptively decompose a nonstationary
denoising ability. signal into a series of narrowband amplitude and frequency-
Index Terms— Adaptive threshold, seismic random noise, modulated signals which are named as intrinsic mode func-
shearlet transform, sparsity. tions (IMFs). The seismic denoising methods based on EMD
and its derivations have been widely discussed in many works
I. I NTRODUCTION of literature [15]–[18]. With the development of exploration
technology and the complexity of exploration targets, some
R ANDOM noise attenuation is extremely important in
almost every aspect of seismic exploration, as it not only
improves the signal-to-noise ratio (SNR) of seismic signal
new denoising methods, such as mathematical morphology
filtering (MMF) [19], patch-based methods (PBMs) [20], and
but also influences the subsequent seismic data processing. deep learning (DL) [21] are continually investigated.
During the last several decades, various methods have been Shearlet transform proposed by Guo and Labate [22] com-
proposed and obtained better noise reduction results. Usually, bines complex wavelet theory and multiscale analysis tech-
these methods can be classified into five categories. The first nology and applies the framework of affine-like systems to
is based on domain transform and filtering theory, the seismic efficiently capture the geometry of multidimensional signals.
data is transformed from observation space into domain space, The shearlets are generated through parabolic scaling and
and most of the noises can be removed by filtering [1]–[4]. translation operators as well as shearing operators to the
The second is the methods based on sparse transform. The basic functions. Shearlet transform has optimal approxima-
threshold coefficients are transformed inversely into the time– tion properties for a large class of 2-D images and shows
space domain. Many transforms such as wavelet transform [5], a performance superior to wavelet transform and curvelet
seislet transform [6], curvelet transform [7], radon transform transform in a variety of applications. Shearlet transform is
[8], and so on, have been employed to attenuate seismic very suitable to handle seismic data which contains much
random noise. The third utilizes a rank reduction theory to directional information, such as textures, contours, and edges.
suppress random noise [9], [10]. The Hankel matrix of noise- Some noise attenuation methods based on shearlet transform
free data is low rank, but the matrix of the noisy data is have been proposed [23]–[25], but many of them just utilize
nonsingular because of the noncoherence of random noise. the signal sparsity in the shearlet domain and do not consider
fully its different scale and directional characteristics.
Manuscript received May 2, 2019; revised August 5, 2019 and Septem- Most of the shearlet-based denoising methods are based
ber 22, 2019; accepted October 17, 2019. Date of publication November 7,
2019; date of current version August 28, 2020. This work was supported on the universal “VisuShrink” threshold and its derivations
in part by the Natural Science Foundation of Jiangsu Province of China proposed by Donoho [26]. In this letter, we propose a novel
under Grant BK20181033 and Grant BK20170436 and in part by the Science adaptive threshold which combines the multiscale and mul-
Foundation of Changshu Institute Technology under Grant KYZ201600082.
(Corresponding author: Jicheng Liu.) tidirectional characteristics of shearlet transform to reduce
The authors are with the School of Electric and Automatic Engi- seismic random noise. The threshold is calculated adaptively
neering, Changshu Institute of Technology, Changshu 215500, China according to the different decomposed scales and the different
(e-mail: [email protected]; [email protected]; [email protected];
[email protected]). directions at each scale. The L1 norm of the coefficients
Digital Object Identifier 10.1109/LGRS.2019.2949806 at every scale and different directions, which expresses the
1545-598X © 2019 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: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on November 08,2024 at 05:50:20 UTC from IEEE Xplore. Restrictions apply.
 1638 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 17, NO. 9, SEPTEMBER 2020

sparsity of the coefficients and indicates the concentration of

effective signal, is used to calculate the directional factors.
The adaptive threshold can preserve an effective signal at
utmost when removing noise. Synthetic and field examples
demonstrate the validity of the proposed method.

II. T HEORY
A 2-D seismic data can be expressed as follows:
d(t) = s(t) + n(t) (1)
where s(t) is the noise-free data and n(t) is the random
noise. Decomposing seismic data based on shearlet transform,
we obtain shearlet coefficients ψ j,l,k , where j , l, and k are
scale, direction, and translation, respectively. According to the Fig. 1. (a) Original data. (b) Noisy data.
sparse transform theory, shearlet coefficients of the effective
signal are sparse and have a larger value. However, random
noise is not sparse and its shearlet coefficients are smaller. The In (6), J is the number of scales, A j,l is the L1-norm
thresholding method suppresses noise by removing the small of shearlet coefficients on scale j and along the direction
coefficients. The denoised data can be obtained by transform l, max (A j,l ) expresses the maximum L1-norm of shearlet
1≤ j ≤ J
the thresholded coefficients into the time-spatial domain. The coefficients on scale j . A larger L1-norm of coefficients on
universal threshold is a popular candidate [26] scale j and along direction l demonstrates that the effective
 signal concentrates along the direction l, a smaller threshold
T H u = σ 2 ln(N) (2)
should be utilized to protect the effective signal. Otherwise,
where N is the total sampling points and σ is the noise the random noise is dominant, and a larger threshold should
variance. The value of the threshold should guarantee that the be adopted to suppress noise. Thus, the adaptive threshold
small coefficients corresponding to noise components can be T H d has superior noise attenuation capacity and reserves the
wiped out under a higher probability. Usually, σ is estimated effective signal to the utmost extent.
by a robust estimator based on the components’ median
III. N UMERICAL E XAMPLES
median(|ψ j,l,k |, j, l ∈ Z , k ∈ Z 2 )
σ̂ = . (3) First, a synthetic seismic data set is used to compare the
0.6745
denoising performance of the proposed threshold with the
The deficiency of the universal threshold T H u is that it is universal threshold and adaptive threshold based on scales.
unable to change with the varying scales and directions, so the The results demonstrate the validity of the proposed method.
denoising results are not good. Therefore, a threshold based Second, numerical examples of real data further demonstrate
on scales is proposed [27] the validity of the proposed method.

T H s = σ 2 ln(N) · C (4) To evaluate the denoising quality, we define SNR, mean
square error (MSE), peak signal-to-noise ratio (PSNR), and
where C is a scale factor, and it can be calculated as C = denoised ability (DA). The smaller MSE means the better
2(k−K )/2 , where k is the current scale index and K is the removal of noise. The larger SNR, PSNR, and DA indicate
number of scale. The value of T H s varies with different scales the better denoising result [28].
and the denoising results are superior than that of T H u . But SNR is defined as
T H s still do not include the directional sensibility of shearlet   2
i j di, j
transform. If the direction of the shearlet basis function is near SNR =   (7)
j (di, j − d̂i, j )
2
to that of an effective signal, we can obtain larger shearlet i
coefficients, vice versa. Thus, the coefficients’ absolute values
where di, j is noise-free data and d̂i, j is the denoised data.
of the effective signal along the special direction are much
MSE and PSNR can be estimated by
larger than the value of noise. Nevertheless, T H s is a constant 
on each scale and does not change with direction. MSE = (di, j − d̂i, j )2 /(M · N) (8)
According to the compressive sensing (CS) theory, L0-norm i j
can be relaxed to L1-norm to describe the sparsity. On the 255
basis of T H s , an adaptive threshold combined with directional PSNR = 20 log10   (9)
j (di, j − d̂i, j ) /(M · N)
2
information is defined as follows: i
 In (8) and (9), M and N are the number of trace and samples,
T H d = σ 2 ln(N) · C · λ (5)
respectively.
where λ is a directional factor and can be calculated as DA is defined as
max (A j,l ) SNR − SNR0
1≤ j ≤ J DA =     (10)
λ= . (6) (d − d̃ ) 2− (d − d̂ ) 2
A j,l i j i, j i, j i j i, j i, j

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on November 08,2024 at 05:50:20 UTC from IEEE Xplore. Restrictions apply.
 LIU et al.: SEISMIC RANDOM NOISE REDUCTION 1639

Fig. 3. Removed noise corresponding to Fig. 2(a)–(d).

Fig. 2. Denoised results of Fig. 1(b) by using (a) F-X deconvolution,
(b) T H u threshold, (c) T H s threshold, and (d) T H d threshold.

where d̃i, j is the noisy data, and SNR and SNR0 are the SNR
of the denoised data and the noisy data, respectively.

A. Synthetic Data Tests

Fig. 1(a) shows the noise-free synthetic seismic data, includ-
ing 256 traces with 256 sampling points per trace with 4 ms
as time sampling interval. Fig. 1(b) shows the noisy data that
the SNR is 2 dB, where the effective events are submerged in
noise.
Fig. 2 demonstrates the denoising results used F-X decon- Fig. 4. Field data.
volution, T H u , T H s , and T H d , respectively. We implemented TABLE I
the shearing on four scales and the level of shearing occurring C OMPARISON OF THE F OUR D ENOISING M ETHODS IN
on each scale is (1, 2, 2, 2). Compared to Fig. 1(b), all the T ERMS OF SNR, MSE, PSNR, AND DA
methods can filter out noise and enhance the data’s quality,
while the proposed method outperforms the other three meth-
ods. Table I shows the statistical comparison of the denoised
results of the four methods in terms of SNR, MSE, PSNR, and
DA. The proposed method improves SNR, PSNR, and DA
better than the other methods and reduces MSE effectively,
which indicates the superior ability to preserve the effective
signal and suppress the random noise. Fig. 3 shows the residual
profiles corresponding to the different methods. Compared to B. Field Data Tests
the residual of T H u the thresholding method, the residual of Fig. 4 shows a real post-stack seismic data, including
the F-X deconvolution method contains a much more useful 201 traces with 350 sampling points per trace with 4 ms
signal. The method using T H s removes most of the noise as time sampling interval. The random noise submerges the
and preserves a more effective signal, but the denoising result weaker events, which seriously influences the resolution of
using T H d is the best and retains the most effective signal. the profile.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on November 08,2024 at 05:50:20 UTC from IEEE Xplore. Restrictions apply.
 1640 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 17, NO. 9, SEPTEMBER 2020

threshold is a key problem. In this letter, an adaptive thresh-

old based on the scale and directional characteristics of the
shearlet transform is proposed. According to the difference
between the coefficients’ sparsity of effective signal and noise
on scale and direction, the threshold can change adaptively
with the sparsity. Compared to the universal threshold and
the threshold based only on scale, the scale and directional
threshold methods can remove random noise and reserve an
effective signal to the utmost extent. Both synthetic and field
data examples demonstrate that the proposed method can
improve the denoising effect significantly and has the stronger
denoising ability.

R EFERENCES
[1] K. Chen and M. D. Sacchi, “Robust f − x projection filtering for
simultaneous random and erratic seismic noise attenuation,” Geophys.
Prospecting, vol. 65, no. 3, pp. 650–668, 2017.
[2] S. Huo, W. Zhu, and T. Shi, “Iterative dip-steering median filter,” J. Appl.
Geophys., vol. 144, pp. 151–156, Sep. 2017.
[3] A. J. Ogah and A. D. Chinedu, “The beam-forming technique for
enhancement of noisy seismic refraction data,” Int. J. Geosci., vol. 3,
no. 24, pp. 866–871, 2012.
[4] L. Moreau et al., “Improving ambient noise correlation functions with an
SVD-based Wiener filter,” Geophys. J. Int., vol. 211, no. 1, pp. 418–426,
2017.
[5] J. Mao, J. Gao, and W. Chen, “On the denoising method of prestack
seismic data in wavelet domain,” Chin. J. Geophys., vol. 25, no. 1,
pp. 1155–1163, 2006.
[6] S. Fomel and Y. Liu, “Seislet transform and seislet frame,” Geophysics,
vol. 75, no. 3, pp. V25–V38, 2010.
[7] W. Liu, S. Cao, Y. Chen, and S. Zu, “An effective approach to attenuate
random noise based on compressive sensing and curvelet transform,”
J. Geophys. Eng., vol. 13, no. 2, pp. 135–145, 2016.
[8] Y. Xue, J. Yang, J. Ma, and Y. Chen, “Amplitude-preserving nonlinear
adaptive multiple attenuation using the high-order sparse radon trans-
form,” J. Geophys. Eng., vol. 13, no. 3, pp. 207–219, 2016.
[9] Y. Zhou, C. Shi, H. Chen, J. Xie, G. Wu, and Y. Chen, “Spike-like
blending noise attenuation using structural low-rank decomposition,”
IEEE Geosci. Remote Sens. Lett., vol. 14, no. 9, pp. 1–5, Sep. 2017.
[10] Y. Chen, Y. Zhou, W. Chen, S. Zu, W. Huang, and D. Zhang, “Empirical
low-rank approximation for seismic noise attenuation,” IEEE Trans.
Geosci. Remote Sens., vol. 55, no. 8, pp. 4696–4711, Aug. 2017.
[11] V. Oropeza and M. Sacchi, “Simultaneous seismic data denoising and
reconstruction via multichannel singular spectrum analysis,” Geophysics,
vol. 76, no. 3, pp. V25–V32, 2011.
[12] W. Huang, R. Wang, Y. Chen, H. Li, and S. Gan, “Damped multi-
Fig. 5. Denoised results of the field data example using (a) F-X deconvolu- channel singular spectrum analysis for 3D random noise attenuation,”
tion, (c) T H s threshold, (e) T H d threshold. (b), (d), and (f) Corresponding Geophysics, vol. 81, no. 4, pp. V261–V270, 2016.
residual profiles. [13] D. Zhang, Y. Chen, W. Huang, and S. Gan, “Multi-step damped
multichannel singular spectrum analysis for simultaneous reconstruction
Fig. 5 demonstrates the denoised results and the corre- and denoising of 3D seismic data,” J. Geophys. Eng., vol. 13, no. 5,
pp. 704–720, 2016.
sponding residual profiles using F-X deconvolution, and T H s [14] N. E. Huang et al., “The empirical mode decomposition and the Hilbert
and T H d thresholds, respectively. The F-X deconvolution spectrum for nonlinear and non-stationary time series analysis,” Proc.
method has the worst denoising results, where much more Roy. Soc. London Ser. A, Math., Phys. Eng. Sci., vol. 454, no. 1971,
pp. 903–995, Mar. 1998.
useful information remained in the residual profile as shown [15] J. L. Gómez and D. R. Velis, “A simple method inspired by empirical
in Fig. 5(b). The T H s threshold method suppresses random mode decomposition for denoising seismic data,” Geophysics, vol. 81,
noise evidently. The continuity of events is improved, and no. 6, pp. V403–V413, 2015.
[16] W. Chen, D. Zhang, and Y. Chen, “Random noise reduction using
the borders become clearer. Compared with Fig. 5(c) and (e), a hybrid method based on ensemble empirical mode decomposition,”
the latter reflects the strata structure more clearly. The removed J. Seismic Explor., vol. 26, no. 3, pp. 227–249, 2017.
noise shown in Fig. 5(f) is more random than the other two [17] Y. Chen and J. T. Ma, “Random noise attenuation by f x empirical-
mode decomposition predictive filtering,” Geophysics, vol. 79, no. 3,
methods. Besides, the removed noise of the proposed method pp. V81–V89, 2014.
hardly displays useful structure signals. [18] Y. Chen, C. Zhou, J. Yuan, and Y. J. Zhao, “Applications of empirical
mode decomposition in random noise attenuation of seismic data,”
IV. C ONCLUSION J. Seismic Explor., vol. 23, no. 6, pp. 481–495, 2014.
[19] W. Huang, R. Wang, D. Zhang, Y. Zhou, W. Yang, and Y. Chen,
The thresholding denoising method is widely used in seis- “Mathematical morphological filtering for linear noise attenuation of
mic random noise reduction. However, selecting a suitable seismic data,” Geophysics, vol. 82, no. 6, pp. V369–V384, 2017.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on November 08,2024 at 05:50:20 UTC from IEEE Xplore. Restrictions apply.
 LIU et al.: SEISMIC RANDOM NOISE REDUCTION 1641

[20] S. Amani, A. Gholami, and A. J. Niestanak, “Seismic random noise [25] H. Cheng, G. Chen, E. Wang, Z. Hou, and J. Fu, “Seismic
attenuation via 3D block matching,” J. Appl. Geophys., vol. 136, data de-noising method of adaptive threshold based on Shear-
pp. 353–363, Jan. 2017. let transform,” Acta Petrolei Sinica, vol. 39, no. 1, pp. 82–91,
[21] S. Yu and J. Ma, “Deep learning for denoising,” in Proc. Int. Geophys. Jan. 2018.
Conf., Beijing, China, 2018, pp. 461–464. [26] D. L. Donoho, “De-noising by soft-thresholding,” IEEE Trans. Inf.
[22] K. Guo and D. Labate, “Optimally sparse multidimensional representa- Theory, vol. 41, no. 3, pp. 613–627, May 1995.
tion using shearlets,” SIAM J. Math. Anal., vol. 39, no. 1, pp. 298–318, [27] M. Sheikh, I. Rabiul, X. Huang, and K. Le, “A novel
2007. adaptive shrinkage threshold on Shearlet transform for image
[23] N. Tang, X. Zhao, Y. Li, and D. Zhu, “Adaptive threshold shearlet denoising,” in Proc. Int. Conf. Neural Inf. Process., 2014,
transform for surface microseismic data denoising,” J. Appl. Geophys., pp. 127–134.
vol. 153, pp. 64–74, Jun. 2018. [28] L. Dong, D. Wang, Y. Zhang, and D. Zhou, “Signal enhancement based
[24] X. Y. Wang, Y. C. Liu, and H. Y. Yang, “An efficient remote sensing on complex curvelet transform and complementary ensemble empiri-
image denoising method in extended discrete shearlet domain,” J. Math. cal mode decomposition,” J. Appl. Geophys., vol. 144, pp. 144–150,
Imag. Vis., vol. 49, no. 2, pp. 434–453, 2014. Sep. 2017.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on November 08,2024 at 05:50:20 UTC from IEEE Xplore. Restrictions apply.