Vous êtes sur la page 1sur 17

Journal of Geophysics and Engineering

ACCEPTED MANUSCRIPT

Reconstruction of seismic data with missing traces using normalized


gaussian weighted filter
To cite this article before publication: Lihua Fu et al 2018 J. Geophys. Eng. in press https://doi.org/10.1088/1742-2140/aac31c

Manuscript version: Accepted Manuscript


Accepted Manuscript is “the version of the article accepted for publication including all changes made as a result of the peer review process,
and which may also include the addition to the article by IOP Publishing of a header, an article ID, a cover sheet and/or an ‘Accepted
Manuscript’ watermark, but excluding any other editing, typesetting or other changes made by IOP Publishing and/or its licensors”

This Accepted Manuscript is © 2018 Sinopec Geophysical Research Institute.

During the embargo period (the 12 month period from the publication of the Version of Record of this article), the Accepted Manuscript is fully
protected by copyright and cannot be reused or reposted elsewhere.
As the Version of Record of this article is going to be / has been published on a subscription basis, this Accepted Manuscript is available for reuse
under a CC BY-NC-ND 3.0 licence after the 12 month embargo period.

After the embargo period, everyone is permitted to use copy and redistribute this article for non-commercial purposes only, provided that they
adhere to all the terms of the licence https://creativecommons.org/licences/by-nc-nd/3.0

Although reasonable endeavours have been taken to obtain all necessary permissions from third parties to include their copyrighted content
within this article, their full citation and copyright line may not be present in this Accepted Manuscript version. Before using any content from this
article, please refer to the Version of Record on IOPscience once published for full citation and copyright details, as permissions will likely be
required. All third party content is fully copyright protected, unless specifically stated otherwise in the figure caption in the Version of Record.

View the article online for updates and enhancements.

This content was downloaded from IP address 128.111.121.42 on 19/06/2018 at 07:45


Page 1 of 16 AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1

1
2
3
4
Reconstruction of Seismic Data with Missing Traces using
5 Normalized Gaussian Weighted Filter
6
7 Lihua Fu1, Meng Zhang2, Zhihui Liu1, Hongwei Li1*
8 1. School of Mathematics and Physics, China University of Geosciences

pt
9
No. 388, Lumo Road, Wuhan 430074, P. R. China
10
11 2. Department of Computer Science, Central China Normal University
12 No. 152, Luoyu Road, Wuhan 430079, P. R. China
13 *Corresponding author: hwli@cug.edu.cn

cri
14
15 Abstract: Missing traces complicate the seismic data processing and may cause difficulty in
16
geological interpretation. We present a simple but efficient normalized Gaussian weighted filter
17
18 (NGWF) method for seismic data interpolation that is suitable for reconstruction despite a large
19
number of missing traces in the data, and has low computational complexity. The missing data are
20

us
21 filled with locally retained pixel information via the Gaussian weight. Numerical tests show that the
22
reconstructed result using the NGWF method is better than that using dictionary learning, total
23
24 variation, partial differential equation, and economic orthogonal rank-one matrix pursuit. In addition,
25
the proposed approach is also applied to pre-stack and post-stack seismic section, and the results
26
27
28
29
an
indicate that the new approach is applicable to the recovery of seismic data with missing traces.
Keywords: seismic data; reconstruction; Gaussian weighted filter; missing trace

30 1. Introduction
31
32
dM
Seismic data quality is vital to geophysical application. Unfortunately, obstacles in the field,
33
34 construction conditions and economic factors constrain seismic recording acquisition, which can
35 lead to seismic data with missing traces. Many seismic processes, such as offset vector tilt (OVT)
36
37 (Ling et al 2016), reverse time migration (RTM) (Zhou et al 2011, Li et al 2016), full waveform
38 inversion (FWI) (Anderson et al 2012, Chi et al 2015), amplitude variation with offset (AVO) (Hunt
39
40 et al 2010), etc., are greatly influenced by missing data. Complete and regular data are needed for
41 these applications to yield results of high quality. Therefore, seismic data reconstruction applied
42
43 during the preprocessing stages becomes increasingly important.
pte

44 Various approaches have been developed to deal with missing data in seismic records. Wave-
45
46 equation based reconstruction takes advantage of the physical properties of seismic wave
47 propagation to reconstruct wave-field. However, it is computational expensive in some applications
48
49 (Ronen 1987, Stolt 2002, Kaplan et al 2010). Prediction filters make full use of the property of
50 linear predictability of the signal in the F-X domain and F-K domain. Unfortunately, it requires the
51
ce

52 seismic signal data to be equally spaced (Spitz 1991, Porsani 1999, Naghizadeh and Sacchi, 2009).
53 Sparse dictionary learning (DL) methods have received considerable attention in seismic data
54
55 recovery (Yu et al 2015, Yu et al 2016, Zhu et al 2017). These methods suggest that seismic data be
56 approximated with a few atoms in a dictionary. The dictionary can be predefined (wavelets,
Ac

57
58 curvelets, contourlets, seislets, etc.) or, learned from training sets, the latter being more effective in
59 practice. With the learned dictionary, seismic patches can be constructed by identifying their sparse
60
AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1 Page 2 of 16

1
2
3
4 coefficients and the data can be recovered from the patches. However, the major drawback of these
5 DL-based approaches are their high computational complexity.
6
7 Rank-reduction (RR) methods have been successfully used for seismic reconstruction
8 problems. Multiple singular spectrum analysis (MSSA) has been recently attracting attention

pt
9
10 (Trickett et al 2010, Oropeza et al 2011, Kreimer et al 2013, Rekapalli et al 2017). This technique
11 needs to reorganize spatial data at a given temporal frequency into a Hankel matrix, and through the
12
13 recovery of a fewer number of singular values during the singular value decomposition (SVD) stage,

cri
14 seismic data is reconstructed. A main drawback of MSSA is that it is significant slower because of
15
16 the computation of SVD. Thus far, several techniques have been adopted to accelerate the speed.
17 Economic orthogonal rank-one matrix pursuit (EOR1MP) is one efficient and scalable algorithm
18
19 (Wang et al 2015).
20 For two-dimensional (2-D) seismic data, the reconstruction task is similar to the image-hole-

us
21
22 filling or inpainting problem, thus, many image-processing approaches can be used to solve it. The
23 two most common types are global methods and local methods. The former uses patch operation to
24
25 exploit image self-similarities, which may afford high quality but is not suitable in facing a large
26 number of missing traces, as there are nearly no connected pixels from which one could obtain
27
28
29
an
patches. Total variation (TV) and partial differential equation (PDE) are popular local-image-
inpainting techniques (Galic et al 2010, Getreuer 2012). However, the image contours are easily
30
31 blurred in the TV algorithm because of the nature of the TV regularizer, while the PDE approach
32 can preserve edge information well but it has a high computational cost (Achanata et al 2017).
dM
33
34 Despite the complexity of PDE, it can be implemented in an efficient manner using coarse-to-fine
35 strategies (Grewenig et al 2010).
36
37 In general, many seismic data interpolation methods, such as DL and RR, depend on the global
38 information, which exploit the self-similarities of all the traces. Each missing data prediction needs
39
40 the computation involving all the seismic records. Inevitably, the computational load will be huge,
41 especially in the large dataset applications. For example, the complexity of SVD algorithm in typical
42
43 RR method will reach ( N 3 ) , in which N denotes the dataset size. Besides, the classical dictionary
pte

44 learning method, K-SVD need call T times of SVD operators to update the dictionary, where T is
45
46 the size of dictionary, greater than N to make sure of the dictionary redundancy. In this paper, we
47 present a simple but efficient approach termed as normalized Gaussian weighted filter (NGWF)
48
49 which can fill missing traces using locally retained pixel information via the Gaussian weight. The
50 complexity of our algorithm is only ( N ) , so it is suitable for reconstruction with large datasets. We
51
ce

52 compare the proposed method to DL, TV, PDE and EOR1MP in seismic reconstruction and present
53 its advantages through numerical and field examples.
54
55 2. Method
56
Ac

57 Due to the local similarity of the seismic records, this paper proposes to approximate the
58 missing trace by a linear combination of all the retained neighborhood traces. Especially, in
59
60 predicting a missing trace, the adjacent trace should make more contribution than the distant one.
Page 3 of 16 AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1

1
2
3 That is, the weight of each trace is a monotony decrease function of its distance to the missing one.
4
5 Specially, normalized Gaussian weighted filter is applied to obtain the weight.
6
7 2.1 Normalized Gaussian weighted filter
8

pt
9 A Gaussian filter is effective at removing noise, and smoothing and enhancing an edge
10 (Achanata et al 2017). During the process of filtering, each individual data component is replaced
11
12 with the original intensity value by convolution with a Gaussian function as follow:
13

cri
14  x2  y2 
G ( x , y )  exp  
  , (1)
15  2 2 
16
17 where  is the standard deviation of the distribution.
18 For the Gaussian filter with a window size (2M  1)  (2M  1) , we consider 2D seismic data on
19
20 a regular grid S  S ( m , n ), m  1, , N y , n  1,, N x and the discrete normalized

us
21 convolution at the pixel [m, n] is defined as follow:
22 M M M M
1 1
23 ( S  G )[ m, n ]    S [ m  i, n  j ]G[i, j ]    S [i , j ]G[ m  i , n  j ] (2)
24 W i  M j  M W i  M j  M
25
26 And
27
28
29
W 
an
M
 
i  M j  M
M
G[i , j ]

30 is the sum of the filter coefficients to normalize the weights. Each pixel in S is scanned using a
31
32
template, and the Gaussian kernel gives less weight to pixels further from the center of the window.
dM
33 The pixel at [ m , n ] is replaced by the weighted average values in the neighborhood as follow:
34
35 1 M M  (m  i)2  (n  j)2 
Sˆ[m, n]  (S  G)[m, n]    exp    S[i, j] . (2)
36 W iM j M 2 2
 
37
38
In addressing the missing trace problem, only retained data need be considered in calculating
39
40 W as follow:
41 M M
42 W   G[i, j ] (3)
43
pte

i M j M
44 (i , j )
45
46 where  is the set of the existing data in S . Thus, the Gaussian filter is adjusted to a normalized
47 Gaussian weighted filter (NGWF) as follow:
48 ( m i ) 2  ( n  j ) 2
M M
49 
2 2
50   e S[i, j ]
51 i  M j  M
ce

S[m, n] 
(i , j )
52 . (4)
M M ( m  i )2  ( n  j )2
53 
2 2
54   e
55 i  M j  M
(i , j )
56
Ac

57
58
2.2 Seismic reconstruction based on the NGWF method
59
60 For a 2-D seismic data S , the NGWF method can be used to recover the missing data. The
AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1 Page 4 of 16

1
2
3
4 algorithm description can be found in Tab.1. For each retained data, we use NGWF to calculate its
5 weights in predicting the missing ones in line 5 of the table. In line 6 and 7, according to Equation
6
7 (4), both numerator F and denominator G are calculated iteratively.
8 The computational complexity involved is dominated by the calculation of Gaussian weights

pt
9
10 in line 5. For each retained data, Gaussian function needs to be called for about 4M 2 times. Therefore
2
11 the complexity of NGWF for the retained dataset is about (1 r)Nx Ny  4M , in which r (0<r<1) is
12
13 the missing ratio. The complexity of NGWF is only  ( N ) , with N  N x N y .

cri
14 TABLE 1. NGWF for the retained data
15
NGWF for the retained data
16
INPUT: S -----2-D seismic data with the sampled matrix  ;
17
18 (2 M  1)  (2 M  1) -----Gaussian filter size;
19  ----- standard deviation of Gaussian filter
20 OUTPUT: S -----recovered seismic data

us
21 Algorithm:
22 1 Initialization: G and F are zero matrix with the same size as S
23
2 for every retained data k  with its row number rk and column number ck
24
3 Obtain its (2 M  1)  (2 M  1) neighborhood ind_neib
25
26 4 for i  ind _ neib
27
28
29
5 w(ri, ci)  exp  
 2 2
an
( ri rk )2 (ci ck )2 


30 6 G ( ri , ci )  G ( ri , ci )  w ( ri , ci )
31 7 F ( ri , ci )  F ( ri , ci )  w( ri , ci )  S ( rk , ck )
32 8 end for
dM
33 9 end for
34 F (i, j )
35 10 S(i, j ) 
G (i, j )
36
37
38 Generally, a single execution of NGWF is capable to complete the trace records. But, in some
39 extreme case in which there are too few samples retained, it may be not enough to recover all the
40
41 missing data to run NGWF just once. NGWF for the missing data needs to be executed as in Tab.2.
42 In the second execution procedure, each missing data is replaced by the weighted average of the
43
pte

44 neighbor window. Actually, the output Ŝ of Tab.2 is considered as the refinement of S .


45
TABLE 2. NGWF for the missing data
46
NGWF for the unknown data
47
48 INPUT: S ----- recovered 2-D seismic data using a NGWF with the missing label set 
;
49  
(2 M  1)  (2 M  1) -----Gaussian filter size;
50  ----- standard deviation of Gaussian filter
51
ce

OUTPUT: Sˆ ----- recovered seismic data


52
Algorithm:
53  with its row number rk and column number ck
54
for every known data k  
55 Obtain its (2M  1)  (2M  1) neighborhood ind_neib
56 Initialization: G and F are zero matrix with the same size as H
Ac

57 for i  ind _ neib


58
If S( ri, ci)   0
59
60 G (ri, ci)  G (ri, ci) , F (ri, ci)  F (ri, ci)
Page 5 of 16 AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1

1
2
3 else
4
( ri  rk )2  ( ci  ck )2 
5 w(ri, ci )  exp   
 2 2 
6
7 G ( ri , ci )  G ( ri , ci )  w ( ri , ci )
8 F ( ri , ci )  F ( ri , ci )  w( ri , ci )  H ( rk , ck )

pt
9 end
10 end for
11  ( F (ri, ci ))
12 iind _ neib
Sˆ (rk , ck ) 
13  (G (ri, ci ))

cri
14 iind _ neib
15 end for
16
17 2.3 Window size and Gaussian-filter scale
18
19 Window size (i.e. Gaussian-filter size) is an important parameter in the completion algorithm.
20 A large window size will lead to heavy computation, while small window size may not capture the

us
21
22 salient and local features. As for the NGWF algorithm listed in Tab. 1, the window size is set to a
23 9  9 patch ( M  4 ). If the missing ratio is great, we should use the ‘NGWF for the unknown data’
24
25 algorithm to refine the reconstruction data. At this time, the window is always set to a 7  7 patch
26   3 ). And the suggested Gaussian-filter scale is   (2M  1) 3 .
(M
27
28
29
3. Numerical examples
an
30 To verify the feasibility of the NGWF algorithm, an example of the recovery of a synthetic
31 seismic model is considered, and the reconstruction result is compared to the results of the DL, TV,
32
dM
PDE and EOR1MP methods. The model is not realistic and contains two nonlinear events. Figure
33
34 1(a) shows the original synthetic data with 450 traces, 128 sampling points, a 2 ms sampling interval
35 and the spatial sampling interval of 40 m. The horizontal axis denotes spatial-trace sampling and
36
the vertical axis denotes temporal sampling. Figure 1(b) depicts its corresponding spectra. The
37
38 corrupted data has 70% randomly missing traces whose indices are randomly selected between 1
39 and 128. The spectra are shown in Figure1 (c) and (d). Figure1 (e), (g), (i) and (k) present the
40 recovery results using the DL, TV, PDE and EOR1MP methods, respectively. The corresponding
41
42 spectra are shown in Figure1 (f), (h), (j) and (l). We performed the NGWF algorithm using the
43 parameters M  4 and   3 . The reconstructed data are presented in Figure 1(m) and Figure1 (n)
pte

44 showing their spectra. It is clear that missing traces were properly recovered, and the analysis of the
45
spectra, therefore, suggests that the NGWF recovered data is more similar to the original (shown in
46
47 Figure 1(a) and (b)).
48 To estimate the reconstruction quality, the signal-to-noise ratio (SNR) is defined as eq. (5) as
49
follow:
50
2
51
ce

S F
52 SNR  10 log 2
, (5)
53 S  Sˆ
F
54 where S and Sˆ denote the original data and its recovery, respectively.
55
The SNR values of the recovered data and central processing unit (CPU) times using the
56
Ac

57 different aforementioned methods are shown in Table 1. All of the experiments were performed
58 using a Lenovo ThinkPad W550s Workstation with Intel Core i7-5500U Processors and 8.00GB of
59 RAM. For each experiment, the parameters were taken empirically to obtain the best results. The
60
AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1 Page 6 of 16

1
2
3 DL method achieves high-quality reconstruction with an SNR value of 21.1173 dB, while the CPU
4
5 time was 35.7234s which was a result of the dictionary training with the K-SVD algorithm. The
6 PDE method was very efficient with a CPU times of just 0.0413s. In this experiment, the DL and
7 EOR1MP approaches obtained higher recovery than that of the TV and PDE method. The proposed
8

pt
9 NGWF method indeed achieved a higher SNR value and a lower computational cost in this case.
10 More experiments were performed in which the percentage of the missing traces ranged from
11 10%-80%. The SNR performance curves are presented in Figure 2. The reconstruction results using
12
the NGWF method yielded significantly higher quality recovery than that using the DL, TV, PDE
13

cri
14 and EOR1MP methods. Meanwhile, the CPU times of the NGWF method were less than those of
15 the DL and TV methods, and they decreased with an increasing percentage of missing traces.
16
17
18
19
20

us
21
22
23
24
25
26
27
28
29
an
30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49
50
51
ce

52
53
54
55
56
Ac

57
58
59
60
Page 7 of 16 AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1

1
2
3
4
5
6
7
8

pt
9
10
11
12
13

cri
14
15
16
17
18
19
20

us
21
22
23
24
25
26
27
28
29
an
30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49
50 Figure 1. Recovery results for 70% missing traces: (a) original data; (b) spectra of (a); (c) missing
51
ce

data; (d) spectra of (c); (e) result using DL-based method; (f) spectra of (e); (g) result using TV-
52
53 based method; (h) spectra of (g); (i) result using PDE-based method; and (j) spectra of (k); (k) result
54 using EOR1MP-based method; (l) spectra of (k); (m) result using NGWF-based method; and (n)
55 spectra of (m).
56
Ac

57 Table 1. SNR and CPU time comparison of reconstructed data for the four methods
58 Reconstruction methods SNR(dB) CPU time (s)
59
DL 21.1173 35.7234
60
AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1 Page 8 of 16

1
2
3
TV 19.3335 10.7813
4
5 PDE 17.3742 1.4683
6 EOR1MP 20.1281 3.9063
7
NGWF 22.8261 4.5591
8

pt
9
10
11
12
13

cri
14
15
16
17
18
19
20

us
21
22
23
24
25
26
27
28
29
an
Figure 2. SNR value versus percentage of missing traces for the five methods.
30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49 Figure 3. CPU time versus percentage of missing traces for the five methods.
50
51
ce

4. Field examples
52
53 4.1 Pre-stack seismic data processing
54
55 The following example explores the proposed reconstruction algorithm using pre-stack seismic
56 data. The real data consists of a 2-D land line from east Texas, USA. Figure 4(a) shows the shot
Ac

57
gather with 33 traces, 1501 sampling points, and a 2-ms sampling interval. The horizontal axis
58
59 represents the offset of each seismic receiver from the source where each records a trace (vertical
60 axis). One can visualize that the wave-field is more complex including a direct wave, refraction,
Page 9 of 16 AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1

1
2
3 reflection, etc. Figure 4(b) shows the corrupted data with 30% missing traces obtained from Figure
4
5 4(a) after randomly sampling. We give in Figure 4(c)-4(j) the retrieved results and the error results
6 by the DL, TV, PDE and EOR1MP, respectively. The SNR values are shown in the figure captions.
7 The reconstructed data by the proposed NGWF method is shown in Figure 4(k). The SNR value for
8

pt
9 the recovery data using the NGWF method is 27.1394 dB. Figure 4(l) shows the difference between
10 the reconstructed data and the original data. In this experiment, DL and TV obtain better SNR values
11 than PDE and EOR1MP, and the NGWF method achieves the highest SNR value. From Figure 4(k)
12
and (l), one can see that the corrupted data with missing traces was recovered well. To see clearly
13

cri
14 the reconstruction result, we present a close-up of the patches that were taken from the original,
15 corrupted, and recovery data from the 301*13 pixels (red box in Figure 4) in the full data with the
16 offset ranging from -1100 to 1540, and the time ranging from 0.3980 to 0.9980 s. Figure 5(a) and
17
18 (b) display the original data and corrupted data, respectively. Figure 5(c) and (d) portray the
19 reconstructed result and its difference from the original data, respectively. It is evident that the error
20 based on the proposed approach is minimal indicating sound reconstruction accuracy. Meanwhile,

us
21
22 the CPU time of the DL, TV, PDE, EOR1MP and the new NGWF method were 46.5545s, 3.6563s,
23 0.0313s, 4.5000s and 4.0824s, respectively.
24
25
26
27
28
29
an
30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49
50
51
ce

52
53
54
55
56
Ac

57
58
59
60
AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1 Page 10 of 16

1
2
3
4
5
6
7
8

pt
9
10
11
12
13

cri
14
15
16
17
18
19
20

us
21
22
23
24
25
26
27
28
29
an
30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49
50
51
ce

52
53
54
55
56
Ac

57
58
59
60
Page 11 of 16 AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1

1
2
3
4
5
6
7
8

pt
9
10
11
12
13

cri
14
15
16
17
18
19 Figure 4. The reconstructed results of the pre-stack seismic data: (a) original data; (b) corrupted
20 data with 30% missing traces; (c) data recovered using the DL method (SNR=26.6119 dB); (d)

us
21
22 difference between reconstructed data and original data; (e) data recovered using the TV
23 method(SNR=25.2200 dB); (f) difference between reconstructed data and original data; (g) data
24 recovered using the PDE method (SNR=23.6414 dB); (h) difference between reconstructed data and
25
original data; (i) data recovered using EOR1MP method (SNR=24.4919 dB); (j) difference between
26
27
28
29
an
reconstructed data and original data; (k) data recovered using the NGWF method (SNR=27.1394
dB); and (l) difference between reconstructed data and original data.

30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49
50
51
ce

52
53
54
55 Figure 5. Close-up of results: (a) original data; (b) corrupted data; (c) recovered results using the
56
NGWF method; and (d) difference between reconstructed data and original data.
Ac

57
58
59 4.2 Post-stack seismic data processing
60
AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1 Page 12 of 16

1
2
3 Figure 6(a) portrays a 2D post-stack section of the east Texas seismic data that we use as the
4
5 original data comprising 60 traces and 1502 sampling points, and with a time sampling interval of
6 2 ms. Figure 6(b) is the corrupted data with 40% randomly missing traces. Figure 6(c)-(j) presents
7 the restoration results and the error results based on the DL, TV, PDE and EOR1MP method. The
8

pt
9 SNR values are shown in the figure captions. The reconstructed section by the NGWF method is
10 shown in Figure 6(k) with an SNR value of the reconstruction of 25.5553dB. It is evident that the
11 missing traces are recovered well, and the faults become clearer compared to that shown in Figure
12
6(a). The residual depicting the difference between the original data and recovered result is shown
13

cri
14 in Figure 6(l). It can be seen that the residual is minor and the effectiveness of the NGWF method
15 is confirmed.
16
17
18
19
20

us
21
22
23
24
25
26
27
28
29
an
30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49
50
51
ce

52
53
54
55
56
Ac

57
58
59
60
Page 13 of 16 AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1

1
2
3
4
5
6
7
8

pt
9
10
11
12
13

cri
14
15
16
17
18
19
20

us
21
22
23
24
25
26
27
28
29
an
30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49
50
51
ce

52
53
54
55
56
Ac

57
58
59
60
AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1 Page 14 of 16

1
2
3
4
5
6
7
8

pt
9
10
11
12
13

cri
14
15
16
17
18
19
20 Figure 6. The reconstructed results of post-stack seismic data: (a) original data; (b) corrupted data

us
21
22 with 40% missing traces; (c) data recovered using the DL method (SNR=23.2821dB); (d) difference
23 between reconstructed data and original data; (e) data recovered using the TV method
24 (SNR=17.6581 dB); (f) difference between reconstructed data and original data; (g) data recovered
25
using the PDE method (SNR=24.0629 dB); (h) difference between reconstructed data and original
26
27
28
29
an
data; (i) data recovered using the EOR1MP method (SNR=24.3462 dB); (j) difference between
reconstructed data and original data; (k) data recovered using the NGWF method (SNR=25.5553
dB); and (l) difference between reconstructed data and original data.
30
31 5. Conclusion
32
dM
33 We introduced a new NGWF-based method and showed how it could be used in reconstructing
34
seismic data. The computational complexity of the new algorithm is only  ( N ) . NGWF makes full
35
36 use of local similarity between closely located traces, while the other global methods, such as DL
37 and RR, pay more attention to exploiting the global self-similarity in the entire image. Therefore,
38
NGWF may outperform the global methods in dealing with the image with strong local similarity.
39
40 A synthetic simulation was utilized to test the performance. The NGWF method results were better
41 compared to those obtained using the DL, TV, PDE and EOR1MP methods. The SNR values were
42 higher, and the computational cost was lower. We found that the new method is robust and can
43
pte

44 tolerate high missing ratios. Real data from east Texas was also tested using the method. The results
45 suggest that the NGWF method is a viable alternative for seismic data reconstruction. NGWF holds
46 an obvious potential to address the problem in 3-D seismic data processing, which is under study
47
48 and will be reported in near future.
49
50 Acknowledge
51
ce

This work is supported by the National Natural Science Fund of China (grant no.: 61601417),
52
53 Science and technology research project of Hubei Provincial Department of Education (grant no.:
54 B2017597), and Open Research Project of the Hubei Key Laboratory of Intelligent Geo-Information
55 Processing (grant no.: KLIGIP2016A01, KLIGIP2016A02).
56
Ac

57 Reference
58
59 Achanata R, Arvanitopoulos N and Susstrunk S 2017 Extreme image completion. ICASSP 2017,
60 1333-1337
Page 15 of 16 AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1

1
2
3 Anderson J, Tan L and Wang D 2012 Time-reversal checkpointing method for RTM and FWI.
4 Geophysics 77 S93-S103
5
6 Chi B, Dong L and Liu Y 2015 Correlation-based reflection full-waveform inversion. Geophysics
7 80 R189-R202
8
Li X, Du Y, Ling F and Li W 2016 Locally adaptive linear mixture model-based super-resolution

pt
9
land-cover mapping based on a structure tensor. International Journal of Remote Sensing 37
10
11
5802-5825
12 Ling Y, Wang X, Li F and Zhao Y 2016 Application of OVT processing to 3D seismic data in
13 western China SEG International Geophysical Conference 353-355

cri
14
15 Naghizadeh M and Sacchi M 2009 F-X adaptive seismic-trace interpolation. Geophysics 74 V9-
16 V16
17
Galic I, Weickert J, Welk M, Bruhn A, Belyaec Ar and Seidel H 2008 Image compression with
18
anisotropic diffusion. Journal of Mathematical Imaging and Vision 30 255-269
19
20 Getreuer P 2012 Total variation inpainting using split Bregman. Image Processing On Line 2 147-

us
21 157
22
23 Grewenig S, Weickert J and Bruhn A 2010 From box filtering to fast explicit diffusion. DAGM
24 Conference on Pattern Recognition 533-542
25
Hunt L, Downton J, Reynolds S, Hadley S, Trad D and Hadley M 2010 The effect of interpolation
26
27
28
29
an
on image and AVO: A viking case study. Geophysics 75 WB265-WB274
Kaplan S, Naghizadeh M and Sacchi M 2010 Data reconstruction with shot-profile least-squares
migration. Geophysics 75 WB121-WB136
30
31 Kreimer N, and Sacchi M 2013 Tensor Completion based on nuclear norm minization for 5D seismic
32 data reconstruction. Geophysics 78 V273-284
dM
33
Trickett S, Burroughs L, Milton A et al 2010 Rank-reduction-based trace interpolation. SEG
34
Expanded Abstracts 29 3829-3833
35
36 Oropeza V and Sacchi M 2011 Simultaneous seismic data denoising and reconstruction via
37 multichannel singular spectrum analysis. Geophysics 76 V25-V32
38
39 Porsani M 1999 Seismic trace interpolation using half-step prediction filters. Geophysics 64 1461-
40 1467
41
Rekapalli R, Tiwari R, Sen M and Vedanti N 2017 3D seismic data de-noising and reconstruction
42
using multichannel time slice singular spectrum analysis. Journal of Applied Geophysics 140
43
pte

145-153
44
45 Ronen J 1987 Wave-equation trace interpolation. Geophysics 52 973-984
46
47 Stolt R 2002 Seismic data mapping and reconstruction. Geophysics 3 890-908
48
Spitz S 1991 Seismic trace interpolation in the F-X domain. Geophysics 56 785-794
49
50 Wang Z, Lai M , Lu Z, Fan W and Davulcu H 2015 Orthogonal rank-one matrix pursuit for low
51
ce

rank matrix completion. SIAM Journal on Scientific Computing 37 A488-A514


52
53 Yu S, Ma J and Osher S 2016 Monte Carlo data-driven tight frame for seismic data recovery.
54 Geophysics 81 V327-V340
55
Yu S, Ma J, Zhang X and Sacchi M 2015 Interpolation and denoising of high-dimensional seismic
56
data by learning a tight frame Geophysics 80 V119-V132
Ac

57
58 Zhou Z, Howard M and Mifflin C 2011 Use of RTM full 3D subsurface angle gathers for subsalt
59 velocity update and image estimation: Case study at Shenzi field. Geophysics 76 WB27-WB39
60
AUTHOR SUBMITTED MANUSCRIPT - JGE-102058.R1 Page 16 of 16

1
2
3 Zhu L, Liu E and McClellan J 2017 Joint seismic data denoising and interpolation with double-
4 sparsity dictionary learning. Journal of Geophysics and Engineering 14 802-810
5
6
7
8

pt
9
10
11
12
13

cri
14
15
16
17
18
19
20

us
21
22
23
24
25
26
27
28
29
an
30
31
32
dM
33
34
35
36
37
38
39
40
41
42
43
pte

44
45
46
47
48
49
50
51
ce

52
53
54
55
56
Ac

57
58
59
60

Vous aimerez peut-être aussi