Adaptive Multiple Subtraction with a Pattern-based Technique

Jiuying Guo*, Veritas DGC Inc

Summary A pattern-based adaptive multiple subtraction technique is presented in this paper. It jointly uses the primary prediction error filter(PEF) and its projection signal filter that is signal preserved as the constraints to calculate the adaptive matching filter in f-x domain. Both synthetic and real data tests show its effectiveness. Introduction Surface-related multiple elimination(SRME) (Berkhout, 1982; Verschuur et al, 1992) has been an effective algorithm to remove multiples that are difficult to eliminate using other conventional algorithms. SRME has two steps. The first step is multiple prediction, and the second step is multiple adaptive subtraction. In the first step, the multiple model is constructed by 2D convolutions of the seismic data without knowledge of the subsurface structures, and it can handle complex structures. In reality, however, variations in the acquisition wavelet, cable feathering, dip in crossline direction, boundary effect, limited offset range can introduce time shifts or amplitude artifacts into the predicted multiples. Consequently, there are discrepancies between the real multiples and the predicted multiples. The second step takes them into account. Conventional method to do this is to apply a matching filter in t-x domain to the predicted multiples, and then subtract the match-filtered multiples from the seismic data. This matching filter is calculated by a least squares algorithm. Unfortunately, when multiples strongly interfere with primaries, these methods give biased primaries after subtraction. Spitz(1999) presented a pattern-based algorithm, which is based on the popular assumption that the primaries and multiples are predictable in f-x domain. It uses the prediction error filter(PEF) of the primaries as the constraints of minimization to reduce the freedom of the subtraction. This algorithm has proved to be particularly efficient when attenuating the multiples in the most complex structure areas where the multiples strongly interfere with the primaries. Guitton et al(2001) introduced a similar algorithm in t-x domain. Abma et al (2002) made comparisons of different adaptive subtraction techniques. However, in reality, the effectiveness of Spitzs approach is still somewhat limited when multiples are not perfectly predictable and strong random noises exist in the seismic data. In this paper, a more effective approach is presented. It calculates the matching filter by iteratively applying both primary PEF and projection signal filter(Soubaras, 1994) as the pattern constraints to the LSQR. Both synthetic and real data tests show the results are encouraging. Theory Surface-related multiple elimination

The first step of SRME is to predict the multiples based on the feedback model. Its matrix notation is

P = P PBP + (PB

)2 P (PB )3 P

+ ......

(1) (2)

B = S 1 R D 1

where P is the input data, P is the multiple-free data, and S is the seismic wavelet vector, D is the receiver matrix, R is the downward reflection matrix of the surface. The matrix multiplications are 2D convolution of the data with itself in t-x domain. Due to various limitations, there are discrepancies between the predicted multiples and the real multiples as previously discussed. A matching filter has to be used to adaptively subtract the multiples. Unfortunately, when primaries strongly interfere with multiples, conventional adaptive subtraction algorithms give biased primaries. In order to solve the problem, an effective algorithm is presented in this paper. Adaptive multiple subtraction using a pattern-based technique Linear events in f-x domain are predictable. A locally linear event can be described by a local plane differential equation

P P + = 0 x t

(3)

where P ( x , t ) is the wavefield, and is local plane wave slope. If is constant, the Fourier transform of (3) is

dP = 0 + i P dx
The general solution of equation (4) is

(4)

(x ) = P (0 )e i x P

(5)

Adaptive multiple subtraction with a pattern-based technique

is the Fourier transform of P . From (5), we can where P easily see the predictability of the local linear event in f-x

domain.

The

prediction

corresponding PEF is 1, e . For multi-events with different slopes, the PEF is obtained by cascading several two-term filters, and it can be represented by a polynomial. The coefficients of the polynomial are the coefficients of the data PEF

filter
ix

is

e ix ,

and

its

subtraction, Spitz proposed a method that uses the predicted multiple model as the input pattern. Then, everything that resembles this pattern within a given subset is subtracted from the data. This approach can be formulated as

A p (Mf P ) 0
= MTAT A M f p p

(10) (11)

Ad .

The Z transform of
ld

Ad

has the form (6)

M T A Tp A p P

Ad ( z ) =

(z z )
d,j j =1

where A

is the matrix of the primary PEFs.

Ad ( z ) can be factored to the product of primary PEF

A p ( z ) and multiple PEF Am ( z ) . We can calculate

Ad ( z )

from the data, and

Am ( z )

from the predicted

multiples.
A d (z ) = A p (z )A m (z )

(7)

where A ( z ) = p

(z z )
lp p, j j =1

The real data is always contaminated by additive random noises. In order to attenuate these noises, a signalpreserving algorithm based on a projection filter was introduced(Soubaras,1994). In addition, multiples are not always perfectly predictable in f-x domain. In order to solve these problems, an algorithm that iteratively and jointly uses the primary PEF and the projection signal filter as the constraints in the minimization is proposed here. Iteratively calculating the primary PEF can handle the problem when the multiples are not perfectly predictable, and using the projection signal filter can reduce the effects of the additive random noises while preserving signals. This process can be formulated as

Am (z ) =

(z z )
lm m, j j =1

A p B p (Mf P ) 0
= M T BT AT A B M f p p p p

(12) (13) is

M T BT p A p A pB pP

ld = lp + lm

where the matrix of projection signal filters B

A m (z ) .

A p ( z ) can be obtained by deconvolving the A d ( z ) with

2 1 , is a pre-whitening constant. B p = 2I A p AT p + I

Examples The second step of SRME is adaptive multiple subtraction. Conventional way calculates the matching filter by minimizing the energy of the primaries. It can be formulated as Both synthetic datasets (Courtesy of BP) and real datasets have been successfully tested using this new algorithm. Fig. 1 shows one flat primary event and two dipping multiple events with different amplitudes. The amplitude of the steepest event is twice of that of the other dipping event. Fig. 2 shows the two dipping events that corresponding to the multiples to be eliminated from the data. Fig. 3 shows the result of the standard adaptive subtraction, which tells us the multiples are only partially eliminated. Fig. 4 shows the result of the new algorithm, which indicates that the two dipping events have been successfully removed. Fig. 5 shows the events subtracted by the new algorithm. Fig. 6 shows the second synthetic data that consists of a horizontal primary event, and a curved-multiple event. Fig. 7 shows the predicted multiple that is slightly shifted with respect to the curved event in Fig. 6. Fig. 8 is the result of

Mf P 0
= M f

(8)
M
T

(9)

where M is the matrix of multiples, f is the matrix of is the LSQR solution, and P is the matching filters, f matrix of input data, T denotes the conjugate transpose. When primaries strongly interfere with multiples, this algorithm gives biased primaries after subtraction. In order to solve the problem linked with the standard LSQR

Adaptive multiple subtraction with a pattern-based technique

Spitzs algorithm. It indicates that the primary is also attenuated due to that the curved event is not perfectly predictable. Fig. 9 shows the events subtracted from the data by Spitzs algorithm. Its obviously not what we expected. Fig. 10 shows the result of the new algorithm. Clearly the multiples have been almost completely removed. Fig. 11 shows the events subtracted by the new algorithm. Fig. 12 shows a common channel section with multiples. Fig. 13 shows the predicted multiples. Fig. 14 shows the result of standard adaptive multiple subtraction in t-x domain. We can see some residual multiples still exist. Fig. 15 shows the result of the new algorithm. There are no noticeable residual multiples. Fig. 16 shows the difference between Fig.12 and Fig.15. Its obvious that the multiples are almost perfectly removed.

Fig. 5: Events subtracted by the new algorithm

Fig. 6: Data

Fig. 1: Data

Fig. 7: Model(multiple)

Fig. 2: Model(multiples)

Fig. 8: Result of Spitzs algorithm

Fig. 3: Result of standard adaptive subtraction

Fig. 9: Events subtracted by Spitzs algorithm

Fig. 4: Result of the new algorithm

Adaptive multiple subtraction with a pattern-based technique

Fig. 10: Result of the new algorithm Fig. 15: Result of the new adaptive subtraction algorithm

Fig. 11: Event subtracted by the new algorithm

Fig. 16: Difference between Fig. 12 and Fig. 15 Conclusions A new algorithm of adaptive multiple subtraction with a pattern-based technique is introduced in this paper. It calculates the matching filter by jointly and iteratively using the signal PEF and the projection signal filter as constraints to the LSQR problem. Both synthetic and real data tests show its effectiveness. The extension of the algorithm to 3D is straightforward. References Abma, R., Kabir, N., Simon, A. S., McLain B., and Michell S., 2002, Comparisons of adaptive techniques for multiple attenuation. SEG expanded abstracts. Berkhout, A. J., 1982, Seismic migration, imaging of acoustic energy by wave field extrapolation, A theoretical aspects: Elsevier. Guitton, A., Brown, M. Rickett, J., and Clapp, R., 2001, Multiple attenuation using t-x pattern-based subtraction method, 71st Ann. Internat. Mtg: Soc. of Expl. Geophys., 1305-1308. Soubaras, R., 1994, Signal-preserving random noise attenuation by the F-X projection: 64th Ann. Internat. Mtg, Soc. Expl. Geophys., Expanded Abstract, 1576-1579. Spitz, S., 1999, Pattern recognition, spatial predictability, and subtraction of multiple events: The Leading Edge, 18, no.1, 55-58. Verschuur, D. J., Berkhout, A. J., and Wapenaar, C. P. A., 1992, Adaptive surface-related multiple attenuation: Geophysics, 57, 1166-1177.

Fig. 12: Common channel section

Fig. 13: Predicted multiples

Fig. 14: Standard adaptive multiple subtraction in t-x domain