See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/303792656

Modified Sparse Multichannel Blind

Deconvolution

Conference Paper · June 2016

DOI: 10.3997/2214-4609.201601244

CITATIONS READS

0 66

3 authors:

Nasser Kazemi Ali Gholami

University of Alberta University of Tehran
25 PUBLICATIONS 23 CITATIONS 75 PUBLICATIONS 238 CITATIONS

SEE PROFILE SEE PROFILE

Mauricio Sacchi
University of Alberta
291 PUBLICATIONS 2,734 CITATIONS

SEE PROFILE

All content following this page was uploaded by Nasser Kazemi on 04 June 2016.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
 30 May – 2 June 2016 | Reed Messe Wien

We STZ1 09
Modified Sparse Multichannel Blind
Deconvolution
N. Kazemi* (University of Alberta), A. Gholami (University of Tehran) & M.
D. Sacchi (University of Alberta)

SUMMARY
Euclid deconvolution is a multichannel algorithm that leads to the estimation of the multichannel seismic
reflectivity via the solution of homogeneous system of equations. In the ideal case, the eigenvector
associated to the minimum nonzero eigenvalue of the homogenous system of equations is an estimator of
the multichannel reflectivity. However, small amounts of noise impinge on the identification of the
eigenvector associated to the impulse response. Recently, we proposed a method called SMBD that solves
the homogeneous system of equations arising in Euclid deconvolution by imposing sparsity on the
unknown multichannel impulse response. The method can accurately estimate the seismic reflectivity and
wavelet in the presence of a moderate amount of noise. However, it does not model the noise properly and
there is no automatic way for defining the regularization parameter. In this abstract, we tried to improve
the SMBD algorithm by including an extra term to handle additive noise. Moreover, in our new algorithm
the regularization parameters can be automatically estimated via line search and cross validation
procedures. The method is successfully tested on a realistic synthetic example and on marine and land real
datasets.

78th EAGE Conference & Exhibition 2016

Vienna, Austria, 30 May – 2 June 2016
 30 May – 2 June 2016 | Reed Messe Wien

Introduction

Euclid deconvolution is a member of a large group of methods that have been proposed for blind de-
convolution of seismic data. The method is discussed and used by the seismic data processing as well
as by the signal and image processing literature (Rietsch, 1997; Xu et al., 1995; Sroubek and Milanfar,
2012; Kazemi and Sacchi, 2014). The idea can be summarized as finding the common factors of the
z-transform of the source function embedded in a group of seismograms with different reflectivity se-
quences. The problem leads to the estimation of the multichannel seismic reflectivity via the solution of
an homogeneous system of equations. In the ideal case, the eigenvector associated to the minimum non-
zero eigenvalue of the homogenous system of equations is an estimator of the multichannel reflectivity.
However, small amounts of noise impinge on the identification of the eigenvector associated to the im-
pulse response. Kazemi and Sacchi (2014) proposed an improvement to Euclid deconvolution where the
homogeneous equation is satisfied by a sparse solution (sparse impulse response). The problem leads to
a non-quadratic minimization technique where the solution must be constrained to be on the unit sphere.
The method permits to obtain accurate estimates of the seismic reflectivity and wavelet in the presence
of a moderate amount of noise. The method has been recently extended to the surface consistent case
(Kazemi et al., in press). A major shortcoming of the method is the automatic selection of its regulariza-
tion parameter and its stability in the presence of noise. In this paper, we improve our original algorithm
by including a term to handle additive noise. We also discuss the estimation of regularization parameters
and show that they can be automatically estimated via line search and cross validation procedures.

Modified Sparse Multichannel Blind Deconvolution

The input-output relationship for earth system, assuming a stationary source wavelet, can be written as

dj = w∗rj +nj , j = 1...J (1)

where w is source wavelet, ∗ stands for convolution and d j , r j and n j are data, reflectivity and noise
time series of channel j, respectively. We also remind the readers that convolution can be represented
via the Z-transform as follows
D j (z) = W (z) R j (z) + N j (z) , j = 1, . . . , J (2)
and by virtue of equation (2), it is easy to show that
D p (z) Rq (z) − Dq (z) R p (z) =
Np (z) Rq (z) − Nq (z) R p (z) , ∀ p, q.
The latter can be rewritten in matrix-vector form as
D p rq − Dq r p = N p rq − Nq r p , (3)
where D p and N p in equation (3) represent the convolution matrices of channel p of data and noise terms,
respectively. The combination of all possible equations leads to the following system of equations
Ax = Ex , (4)
where  
D2 −D1
 D3 −D1 
−D1
 
 D4 
 .. ..


 . . 

A= D3 −D2 (5)
 


 D4 −D2 

 .
.. .. 

 . 

 DJ −DJ−2 
DJ −DJ−1
78th EAGE Conference & Exhibition 2016
Vienna, Austria, 30 May – 2 June 2016
 30 May – 2 June 2016 | Reed Messe Wien

Trace number Trace number

(a) 10 20
(b) 10 20
0 0

0.1 0.1
Time (s)

Time (s)
0.2 0.2

0.3 0.3

(c) (d)
Trace number
10 20
0 0

0.1
Time (s)

Time (s)

0.05
0.2

0.3

0.10
True wavelet Estimated wavelet

Figure 1 Performance of the modified SMBD method using synthetic data with SNR=3.5. a) True syn-
thetic reflectivity sequences. b) Seismic traces with SNR=3.5. b) Estimated sparse reflectivity sequences.
d) True and estimated wavelets.

and
x = [r1 , r2 , r3 , . . . , rJ ]T . (6)

Matrix E has the same structure as matrix A but constructed from the noise time series only. We will
assume that noise term is white and Gaussian, the reflectivity is sparse and the source wavelet is a smooth
function. Therefore, we propose to find the signals x and w that minimize the cost function

{x̂, ŵ} = argmin ||Ax||22 + λx ||x||1 + λn ||Wx − d||22 + λw ||w||22 (7)

x,w

where W is a block diagonal matrix made of the convolutional matrices of the stationary source wavelet.
The cost function is non-linear and we will solve it by an alternating minimization technique. By fixing
the source wavelet the problem can be solved for the reflectivity using any L2 − L1 solvers. By fixing
the reflectivity, using the updated version of it, the estimation of the source wavelet can be cast as an
L2 − L2 problem which has a closed form solution. We repeat the alternating process till convergence.
There are three regularization parameters one needs to choose. The optimum λn is the one that satisfies
||Ax(λn )||22 = ||Ex(λn )||22 for any non-zero sparse x where the noise term is calculated via N(λn ) =

78th EAGE Conference & Exhibition 2016

Vienna, Austria, 30 May – 2 June 2016
 30 May – 2 June 2016 | Reed Messe Wien

(a) CMP (b) CMP (c) Trace number

200 400 600 800 200 400 600 800
1.6 1.6 0

2.1 2.1

0.05
2.6 2.6
Time (s)

Time (s)

Time (s)
3.1 3.1
0.10

3.6 3.6

0.15
4.1 4.1

4.6 4.6
0.20
Estimated wavelet
Figure 2 Performance of the modified SMBD method using Gulf of Mexico dataset. a) Near offset section
of data set from the Gulf of Mexico. b) Estimated sparse reflectivity. c) Estimated source wavelet.

Wx − d. Optimum values for λx and λw can be chosen based on L-curve methods or the generalized
cross validation approach. In next section we will show the efficiency of the method on synthetic and
two real data examples. Comparing to our previously proposed method, SMBD, the current algorithm is
more suitable to process noisy data, hence we will call this technique as modified SMBD. In essence, in
this paper, we have incorporated the noise term λn ||Wx − d||22 that serves to control the fitting of noisy
data. In the absence of this term, the method in only reliable for blind deconvolution scenarios with a
high SNR.

Examples

In following the examples, we use a delta time series as the initial estimate of source function and the data
as an initial estimate of reflectivity series. Choosing the data as an initial solution for reflectivity series
accelerates the convergence of our our improved SMBD method. To test the algorithm, we generated a
synthetic example with realistic signal to noise ratio (SNR = 3.5). The data and reflectivity series are
shown in Figures 1a and b. The estimated reflectivity series is represented in Figure 1c. The estimated
source wavelet is compared with the true one in Figure 1d. The algorithm is ran for 9 alternating passes
with 1000 iterations for the reflectivity estimation part. The results are in good accordance with the true
reflectivity and source wavelet time series. The quality of estimations are better than our previously
proposed SMBD method.

Next, we applied our method to a Gulf of Mexico near offset section. We ran the algorithm for 7
passes with 100 iterations for the reflectivity estimation part. Figure 2 shows the results of Gulf of
Mexico dataset. As it is clear from the figure, the estimated reflectivity series looks coherent in the offset
direction thank to the application of multichannel algorithm which takes advantage of the statistical
properties of the signal of interest. The estimated source function is also interesting. We were able to
estimate the main phase and the bubble effect of airgun source function. The deconvolved data show
higher resolution that the original data set. We also applied our method on a 2D stack section of the
Alaska North slope dataset (line 31-81). The data are already processed, so we expect to see a little
bit of improvement in the resolution of the dataset and to estimate a zero phase residual wavelet as the
common source function after application of the improved SMBD. Figure 3 shows the results for this

78th EAGE Conference & Exhibition 2016

Vienna, Austria, 30 May – 2 June 2016
 30 May – 2 June 2016 | Reed Messe Wien

(a) CMP (b) CMP (c) Trace number

100 200 300 400 500 100 200 300 400 500
0 0 0

0.5 0.5

0.05

1.0 1.0
Time (s)

Time (s)

Time (s)
1.5 1.5 0.10

2.0 2.0

0.15

2.5 2.5

0.20
Estimated wavelet
Figure 3 Performance of the modified SMBD method using Alaska North slope stack section. a) 2D stack
section of Alaska North slope dataset. b) Estimated sparse reflectivity. c) Estimated source wavelet.

land data example. The estimated reflectivity and source function series are in good accordance with our
expectations.

Conclusions
We have presented an improved version of the previously proposed SMBD method. The improved
SMBD algorithm alleviates some of the problems encountered in SMBD. For instance, the new SMBD
models the noise, hence it is more applicable to field data. Moreover, the regularization parameters can
be automatically defined. We used alternating minimization technique to solve for both reflectivity and
source function time series. Starting with constant time series as an initial guess for source function
and data as an initial estimate for reflectivity series always results in satisfactory performance of the
technique. The method is successfully tested on a realistic synthetic example and on marine and land
real datasets.

Acknowledgment
We thank the sponsors of the Signal Analysis and Imaging Group (SAIG) at the University of Alberta.
We also thank WesternGeco for providing the Gulf of Mexico, Mississippi Canyon dataset and the USGS
for the Alaska North slope dataset and the SEG for facilitating access to the data via http://wiki.seg.org.

References

Kazemi, N., Bongajum, E. and Sacchi, M. [in press] Surface-Consistent Sparse Multichannel Blind
Deconvolution of Seismic Signals. IEEE Trans. Geoscience and Remote Sensing.
Kazemi, N. and Sacchi, M. [2014] Sparse multichannel blind deconvolution. Geophysics, 79(5), V143–
V152.
Rietsch, E. [1997] Euclid and the art of wavelet estimation, Part II: Robust algorithm and field-data
examples. Geophysics, 62(6), 1939–1946.
Sroubek, F. and Milanfar, P. [2012] Robust Multichannel Blind Deconvolution via Fast Alternating Min-
imization. IEEE Trans. Image Processing, 21, 1687 – 1700.
Xu, G., Liu, H., Tong, L. and Kailath, T. [1995] A least-squares approach to blind channel identification.
IEEE Trans. Signal Processing, 43(12), 2982 –2993.

78th EAGE Conference & Exhibition 2016

Vienna, Austria, 30 May – 2 June 2016

View publication stats