You are on page 1of 8

CHINESE JOURNAL OF GEOPHYSICS Vol.51, No.

1, 2008, pp: 168175

RECONSTRUCTION OF SEISMIC DATA WITH LEAST SQUARES INVERSION BASED ON NONUNIFORM FAST FOURIER TRANSFORM
MENG Xiao-Hong1,2 , GUO Liang-Hui1,2 , ZHANG Zhi-Fu1,2 , LI Shu-Ling1,2 , ZHOU Jian-Jun2
1 State Key Laboratory of Geological Processes and Mineral Resources, China University of Geosciences, Beijing 100083, China 2 Key Laboratory of Geo-detection (China University of Geosciences, Beijing), Ministry of Education, Beijing 100083, China

Abstract Reconstruction of uneven seismic data is a crucial step of the analyzing and processing of the data. In this paper, a least squares inversion based on non-uniform fast Fourier transform is proposed, which can reconstruct both uneven and alias seismic data. The algorithm of non-uniform fast Fourier transform is used to do Fourier transform of uneven data. An adaptive DFT weighted norm regularization constraint is used during the reconstruction. The inverse problem is solved by the preconditioned conjugate gradient method, which makes the solutions stable and convergence quick. Experimental results on theoretical model and raw uneven seismic data show that the proposed method is fast, ecient and applicable. Key words Non-uniform fast Fourier transform, Reconstruction of uneven seismic data, Least squares inversion

1 INTRODUCTION In many cases, such as when there are barriers, fenced-o areas or bad traces in disks, or limited by economic conditions, the sampled seismic data are usually uneven or sparsely distributed. There are several ways to process these data, such as to generate neighbouring traces by copy or linear interpolation, to ignore blank traces (or preserve bad traces), and stack all traces of 3-D data onto a bin despite uneven sampling points. Such relatively simple processing would result in serious alias and aect application of most of routine techniques. Therefore it is needed to perform reconstruction using proper mathematical methods to these unevenly sampled data. On the reconstruction of seismic data, there have been many detailed studies and various solution methods, which can be classied into ve sorts below: (1) Interpolations in the space domain, such as SINC function interpolation based on sampling theory after suppressing alias by NMO[1] , maximum coherent dip interpolation[2] , trace gather weighted slant-stack interpolation[[3] , event attribute interpolation[4] , dierential power spectrum slant coherent interpolation[5] , and t-x domain prediction error lter (PEF) interpolation[6] . (2) Interpolations in the frequency domain, including f -x domain trace interpolation[7] , f -k domain trace interpolation[8] , f -k domain prediction-error lter interpolation[9,10] , and f -k domain projection lter interpolation[11] . (3) Interpolations based on Radon transform, which utilize slant stack and parabolic Radon transform or hyperbolic Radon transform to construct interpolation operators for reconstruction of uneven data[12,13] . (4) Interpolations based on wave equation, e.g. data regularization by alternating NMO and anti-DMO[14] , and the AMO method[15] . (5) Interpolations by combination of dierent methods, such as wave eld decomposition interpolation by picking dips in f -x domain[16,17] . The common feature of the ve kinds of methods above is that they use a denite base function to represent seismic signal and perform corresponding processing in the representation domain. Then the processed result is inversely transformed into the signal domain, accomplishing reconstruction of signal and data recovery. In
E-mail: mxh@cugb.edu.cn

Meng X H et al.: Reconstruction of Seismic Data with Least Squares Inversion Based on Nonuniform Fast 169 theory these methods should yield relatively precise reconstructions. In practise there are, however, still many issues remaining to be studied in detail due to complicated real cases, such as alias, stability of algorithms, and practicality of methods. Liu[18] and Wang[19] proposed a sparsely constrained least squares inversion based on Fourier transform, which introduces sparse constraint based on Cauchy criterion[20] into the model space. On the basis of their work, we use non-uniform Fourier transform[21] to replace the sparse constraint operator for the frequency spectrum of reconstructed data. In the solution process, the preconditioned conjugate gradient method[2224] is utilized to modify the frequency spectrum values from non-uniform Fourier transform in a repeated iteration way, so that increasingly to approximate the true one and to remove the spatial alias. The reconstructed data can facilitate further data processing such as elimination of multiples, AVO analysis, and migration imaging. 2 NON-UNIFORM FAST FOURIER TRANSFORM (NFFT) In a 1-D case, non-uniform discrete Fourier transform is dened as
N

Y (m) =
n=1

yn ejmkxn , m =

M M ,, 1, 2 2

(1)

where xn is the position of the sampled point in space, yn is the signal value of xn , k is sampling interval in the frequency domain, N is the number of sampled points which are uneven, and M is the component number in the frequency domain. Suppose that X = 2/k , for any l we have e
2 jmk xn +l k

= ej(mkxn +2ml) = ejmkxn ,

(2)

thus xn [0, X ). Let kxn = 2 (xn /X ), using xn to replace xn /X , Eq.(1) is rewritten as


N

Y (m) =
n=1

yn ej2mxn xn , m =

M M ,, 1, 2 2

(3)

where x [0, 1). Further, Eq.(3) is rewritten as


N

Y (m) =
n=1

yn (x xn )ej2mx dx.

(4)

It can be noted that Y (m) in Eq.(3) is the Fourier transform of the pulse sequence
N

y (x) =
n=1

yn (x xn ).

(5)

Now Eq.(4) is simplied as Y (m) =

y (x)ej2mx dx,

(6)

which means that Y (m) and y (x) are a pair of Fourier transform. Its fast algorithm is as follows (Fig. 1): (1) Make convolution of non-uniformly sampled data y (x) with a proper short lter factor g (x): y1 (x) = y (x) g (x), so that its frequency spectrum space is band-limited; (2) Sample y1 (x) uniformly, yielding y2 (x); (3) Perform periodic continuation of y2 (x) resulting in y3 (x), then implement fast Fourier transform to y3 (x) yielding Y3 (m);

170

Chinese J. Geophys.

Vol.51, No.1

(4) Make de-convolution of Y3 (m) with the frequency spectrum G(m) of a lter operator g (x), resulting in the frequency spectrum of non-uniformly sampled data y (x): (m) = Y3 (m) . Y G(m) (7)

Two parameters are required for the fast calculation above. One is the number q of discretizing (sampling) the short lter g (x). The other is the over sampling factor f , that determines the number Nc of regular sampling on y1 (x) = y (x) g (x), i.e. f = Nc /M . In this work the used short lter g (x) is a cone Gauss pulse function: g (x) = e
x 1 1 bh bx2 2 + 2 cos x c

xc < x < xc Others

(8)

0,

Fig. 1 Sketch of non-uniform fast Fourier transform (NFFT) algorithm

2 /q, x is where xc = (q/2)x, b = (2 1/f )Nc discrete sampling interval for g (x), bh is the correction factor, q is the number of discrete sampling on g (x), and Nc is the number of regular sampling on y1 (x) = y (x) g (x) (see literature [25]).

3 RECONSTRUCTION OF SEISMIC DATA BY THE LEAST SQUARES INVERSION BASED ON NFFT In this paper we present the algorithm for 1-D interpolation along x direction in the f -x domain, which can be easily extended to multiple-dimension cases. First, the real seismic data set in the t-x domain is transformed into the f -x domain. Then regularized interpolation is performed to every instant frequency f along x direction in space. Suppose that x denotes the data on a regular grid of length M , x = [x1 , x2 , , xM ]T , y is the known uneven measured data, y = [y1 , y2 , , yN ]T . Dene a sampling matrix A, of which the elements are Aij = n(i),j . The relation between the data on the regular grid and the uneven measured data can be described by a linear system: y = Ax. (9) Assume the complete regular data is M = 9, or x = [x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 , x9 ], and the measured data set is N = 2, 3, 5, 7, 9 or y = [x2 , x3 , x5 , x7 , x9 ]T , then Eq.(9) becomes x1 x2 x2 0 1 0 0 0 0 0 0 0 x3 x 0 0 1 0 0 0 0 0 0x 3 4 x5 (10) x5 = 0 0 0 0 1 0 0 0 0 , x7 0 0 0 0 0 0 1 0 0 x6 0 0 0 0 0 0 0 0 1 x7 x9 x8 x9

Meng X H et al.: Reconstruction of Seismic Data with Least Squares Inversion Based on Nonuniform Fast 171 where AAT = I N , I N is a unit matrix of N N order. Suppose that the data x is band-limited signal with band-width , and dene x = Bx, where the band-limited operator B is B= 1 H F F , M (12) (11)

where F is the operator matrix for discrete Fourier transform, F H is its complex conjugate transpose, is an 1, k . M M diagonal matrix, i.e. = diag(k ), k = 0, k / Thus there is 1 1 (13) y = ABx = A F HF x = A F HX , M M where X = F x, x = F H X . The interpolation problem from Eq.(9) is an undetermined system (the regularized data set is greater than the measured data) and without unique solution. Usually a priori information is invoked to constrain the solution range. The basic idea is to search a least solution in the space of all possible solutions, which minimizes the model norm. If error is taken into account, the inverse problem is simplied into a constrained minimum one for the following solutions 2 2 J (x) = ABx y 2 (14) 2 + Lx 2 , where Lx 2 is the regularized constraint, L is the roughness factor, and is trade-o factor. A Cauchy distributed constraint is introduced as the regularization term, and the following functional is minimized: 2 2 J (x) = ABx y 2 (15) 2 + x p, where x
2 p

is the adaptive DFT weighted norm dened by x


2 p

Xk |Pk |2 Xk =

1 H H 2 1 x F P Fx = PFx M M

,
2

(16)

where P is M M matrix, in which when |Pk = 0(k ) the diagonal element is 1/|Pk |, and when |Pk | = 0(k / ), the diagonal element is 0. Since the shape of |Pk |2 is like the power spectrum of the reconstructed complete regular data set, which can be approximated by the periodic chart of x, while it is unknown, the follow formula can be made out by the modied periodic chart: 1 (n1) (n1) (n)2 Pk = X (X ) . (17) M The initial estimate |Pk |(0) is obtained by non-uniform fast Fourier transform of uneven measured data. To ensure stable solution, fast convergence and high precision, the preconditioned conjugate gradient 1 method is used to solve Eq.(15). The applied preconditioned equation is Q = P F x, then we have M M AF H P 1 y Q= , (18) 0 where P 1 is the inverse of P , which is also an M M diagonal matrix. Solving (18), the desired reconstructed data can be obtained by x = M F H P 1 Q. (19)

172

Chinese J. Geophys.

Vol.51, No.1

As A in Eq.(18) is a coecient matrix operator, AF H can constitute a non-uniform Fourier transform operator which can be calculated by the non-uniform fast Fourier transform algorithm. Let T = AF H , then Eq.(18) is rewritten as (20)

M T P 1

Q=

y 0

(21)

Thus, the reconstruction of seismic data by least squares inversion based on non-uniform fast Fourier transform can be described by the following process: (1) Perform non-uniform fast Fourier transform of uneven measured data, yielding the initial estimate |Pk |(0) of the power spectrum; (2) Using the Eq.(20) of non-uniform fast Fourier transform, obtain matrix T ; (3) Solve Eq.(21) by the preconditioned conjugate gradient algorithm, and calculate x(0) from Eq.(19); (4) Calculate the weighted function |Pk |(n) using the modied periodic chart Eq.(17); x ( n) x (n1) (5) When < is satised, let iteration end, otherwise return to step (3) to continue. ( n x ) 4 TESTS ON THEORETICAL MODELS Based on the theoretical method and calculation procedures presented in this paper, a trial computation is rstly performed on a simple theoretical model (single high-frequency linear event). This model contains 40 traces of uneven seismic data with sampling time interval 4 ms, where traces of varied amounts are missing at 4 places along the theoretical event (Fig. 2a). Using the approach of this work to reconstruct seismic data, 65 traces are generated (Fig. 2b). It demonstrates that the eective signals have been roughly recovered.

Fig. 2 The uneven seismic data model with single high-frequency linear event (a), and its reconstructed seismic data using least squares inversion based on non-uniform fast Fourier transform

Figure 3a shows another theoretical data model which has 4 low-frequency curved events, and 64 traces, of which 28 traces are blank, and the sampling time interval is 4 ms. Using the method presented in this paper, the data of this model is reconstructed through 10 times of iteration in the preconditioned conjugate gradient algorithm. As shown in Fig. 3b, in the resulting data the events are well reconstructed.

Meng X H et al.: Reconstruction of Seismic Data with Least Squares Inversion Based on Nonuniform Fast 173

Fig. 3 The uneven seismic data model with four low-frequency curved events (a), and its reconstructed seismic data using least squares inversion based on non-uniform fast Fourier transform

5 TEST ON REAL SEISMIC DATA A real 2-D seismic data from an oil eld is chosen for the test, in which seismic traces are taken arbitrarily for 1-2 seconds, totally 74 uneven-distributed traces (Fig. 4a). Using the method presented in this paper to reconstruct these uneven data yields 110 traces (Fig. 4b). And Fig. 4c displays the reconstruction of the same data by the routine least squares inversion. Comparing the seismic proles before and after reconstruction using the least squares inversion based on non-uniform fast Fourier transform, we can see the good continuity, and the original structural features in the new images are more obvious with more natural variations. While the reconstruction of the data by the routine least squares inversion cannot recover the information of the events very well. 6 CONCLUSIONS In this paper we propose a new least squares inversion based on non-uniform fast Fourier transform. This method assumes that reconstructed seismic data is band-limited in space. In the process of inversion construction, a regularization constraint of adaptive DFT weighted norm is introduced to make the reconstruction subject to band-limit, spectral shape and brace are bounded. The non-uniform fast Fourier transform is used on uneven data to obtain the initial estimate of regularization constraint. Then adaptive regularization constraint is approximated by the periodic chart of x. The preconditioned conjugate gradient algorithm is utilized to ensure stability and convergence speed of the solution. Tests on theoretical models and real seismic data indicate that this method is stable, fast, precise and practical with respect to the routine least squares inversion.

174

Chinese J. Geophys.

Vol.51, No.1

Fig. 4 (a) Real 2-D uneven seismic data from a oil eld, (b) Reconstruction of the data using the least squares inversion based on non-uniform fast Fourier transform, (c) Reconstruction of the data using the routine least squares inversion

ACKNOWLEDGMENTS This work was supported by the National Natural Science Foundation of China (40674066), the National 863 Program (2006AA06Z111, 2006AA06A201-3, 2006AA09A101-0201), and the Focused Subject Program of Beijing (XK104910598). REFERENCES
[1] Jakubowicz H. Waveeld reconstruction. Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts, 1994. 15571560 [2] Larner K, Gibson B, Rothman D. Trace interpolation and the design of seismic surveys. Geophysics, 1981, 46(4): 407409 [3] Lu L. Application of local slant-stack to trace interpolation. Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts, 1985. 560562 [4] King G A. Leong T K, Flinchbaugh B E. The role of interpolation in seismic resolution. Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts. 1984 [5] Monk D, McBeath R G, Wason C B. Interpolation of aliased seismic traces. US Patent, 1993. 5235556

Meng X H et al.: Reconstruction of Seismic Data with Least Squares Inversion Based on Nonuniform Fast 175
[6] Claerbout J, Nichols D. Interpolation beyond aliasing by (t-x) domain in PEFs. 53rd Ann. Internat. Mtg., Eur. Assn. Geosci. Eng. Expanded Abstracts, 1991. 110 [7] Guo J Y, Zhou X Y, Yu S P. Iso-interval trace interpolation in F -X domain. Oil Geophysical Prospecting (in Chinese), 1996, 31(1): 2834 [8] Guo J Y, Zhou X Y. Iso-interval trace interpolation in F -K domain. Oil Geophysical Prospecting (in Chinese), 1996, 31(2): 211218 [9] Spitz S. Seismic trace interpolation in the f -x domain. Geophysics, 1991, 56(6): 785794 [10] Ji J. Interpolation using prediction-error lter simulation (PEFs). Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts, 1993. 11701173 [11] Soubaras R. Spatial interpolation of aliased seismic data. Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts, 1997. 11671190 [12] Thorson J R, Claerbout J F. Velocity-stack and slant stack stochastic inversion. Geophysics, 1985, 50(12): 27272741 [13] Kabir M M N, Verschuur D J. Restoration of missing osets by parabolic Radon transform. Geophysical Prospecting, 1995, 43: 347368 [14] Ronen J. Wave equation trace interpolation. Geophysics, 1987, 52(7): 973984 [15] Chemingui N, Biondi B. Handling the irregular geometry in wide-azimuth surveys. Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts, 1999. 3235 [16] Pieprzak A W, McClean J W. Trace interpolation of severely aliased events. Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts, 1988. 658660 [17] Pieprzak A W, McClean J W. Interpolation of severely ali-ased events: US Patent, 1990. 4922465 [18] Liu B, Sacchi M. Minimum weighted norm interpolation of seismic data with adaptive weights. Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts, 2001. 19211924 [19] Wang Y H. Sparseness-constrained least-squares inversion: application to seismic wave reconstruction. Geophysics, 2003, 68(5): 16331638 [20] Trad D Q, Ulrych T J, Sacchi M D. Accurate interpolation with high-resolution time-variant Radon transforms. Geophysics, 2002, 67(4): 644656 [21] Duijndam A, Schonewille M. Nonuniform Fast Fourier Transform. Soc. Explor. Geophys. Ann. Internat. Mtg., Expanded Abstracts, 1997. 11351138 [22] Meng X H, Hu C S, Liu H Y, et al. The application of unouniform fast fourier transform to interpolation of geophysical data. Journal of Changchun University of Science and Technology (in Chinese), 2000, 30(Suppl.): 150153 [23] Li B, Liu H, Li Y M. 3-D seismic data discrete smooth interpolation using conjugate gradient. Chinese J. Geophys. (in Chinese), 2002, 45(5): 691699 [24] Liu X W, Liu H, Liu B. A study on algorithm for preconstruction of de-alias uneven seismic data. Chinese J. Geophys. (in Chinese), 2004, 47(2): 299305 [25] Duijndam A J W, Schonewille M A. Nonuniform fast Fourier transform. Geophysics, 1999, 64(2): 539551