Y015 Complementary Data-driven Methods for Interbed Demultiple of Land Data S. Sonika* (WesternGeco), A. Zarkhidze (WesternGeco), J. Heim (WesternGeco) & B. Dragoset (WesternGeco) SUMMARY Interbed multiples occur in all types of seismic datasets and are difficult to address. These multiples can appear very similar to primaries due to comparable velocities, and may be difficult to differentiate from primary energy and/or distort the amplitude of primary reflections. Hence, traditional methods based on velocity, dip discrimination, or periodicity, generally may not be capable of attacking these multiples but data-driven techniques such as SRME and IMP have provided better results. In this paper, complementary data-driven techniques, namely DID and XIMP, for predicting interbed demultiples have been discussed. These robust algorithms are suitable for all acquisition geometries, whether 2D or 3D, including land orthogonal geometries and provide an effective attenuation of interbed multiples.
Introduction Interbed multiples occur in all types of seismic datasets and are difficult to address. These multiples can appear very similar to primaries due to comparable velocities, and may be difficult to differentiate from primary energy and/or distort the amplitude of primary reflections. Traditional methods based on velocity, dip discrimination, or periodicity, generally may not be effective in attacking these multiples. Interbed multiple prediction (IMP)-based on the 1D earth assumption and applied to postmigration gathers has been partially successful in removing these complex multiples (El-Emam et al. 2005), but cannot be considered a full solution as it is applied late in the seismic data processing sequence, limiting the derivation of an accurate velocity, which is critical for imaging, time or depth, and for further reservoir characterization. This problem is further compounded in onshore seismic data. These, in general, have wide acquisition geometries and suffer from poor sampling, especially for shallow reflectors, which adds to the challenges of multiple attenuation. In this paper, two complimentary data-driven multiple attenuation techniques are discussed; these address the various challenges posed by the complexity of acquired data and the generated interbed multiples. Their success is illustrated by data examples. Theory In this paper, two methods to attenuate interbed multiples are discussed. 1. Deterministic interbed demultiple (DID). 2. Extended interbed multiple prediction (XIMP). DID is a deterministic method of predicting the internal multiples that are created by a single layer in the earth s subsurface. Originally designed to handle shallow water-column-related multiples (Moore and Bisley, 2006), it was later adapted to attenuate interbed multiples.the prediction is accomplished by a wavefield extrapolation. The method predicts all multiples created by a targeted layer as shown in Figure 1. DID requires a model of the layer, including estimated reflection coefficients at top and bottom. The main assumption of the method is that the layer has a moderate complexity within the 3D gather. In other words, DID relies on a local 1D assumption for the layer. Nevertheless, it does account properly for the non-periodic moveout of multiples in traces at non-zero offsets. Because the layer is assumed to be flat, there are no azimuthal effects accounted for by the multiple prediction. DID works with 3D gathers and outputs multiple models for each trace of the input data. Overall, it can form a reliable 3D model as long as geological complexity is moderate. In practice, adaptive subtraction of the multiples predicted by DID can correct for small errors in the predictions. This means that the assumption of a locally flat layer does not have to be strictly true. Small deviations from that assumption, as well as small errors in the estimated reflection coefficients are allowed. DID is typically used to predict multiples caused by a shallow layer. The value of DID is its simplicity, efficiency, and accuracy when the field data meet its requirements. DID works very well in the presence of moderately complex geology and helps avoid more difficult data preconditioning, such as massive interpolation/regularization of the data. Figure 1 DID predicts all multiples created by a single layer, such as that between the two red horizons shown. The thicker red horizon is the multiple generator. XIMP is a true-azimuth interbed prediction algorithm that assumes input data are free of surfacerelated multiples. The concept is relatively simple; but the actual implementation is quite complicated due to the millions of traces that are acquired and must be processed to generate the multiple model.
XIMP is based on the extension of the surface-related multiple attenuation technique, as proposed by Jakubowicz (1998), that decomposes an internal multiple wavefield into three components. The concept is illustrated in Figure 2; the internal multiples can be predicted by convolving two primaries X s,x 1 and X r,x 2 and correlating with X 1,X 2. The downward reflection point (DRP) occurs at p2, the multiple-generating horizon. The position of X 1, X2 is not known, a-priori. The solution to this is shown in Figure 3 which depicts the same case in plan view. For a single trace, SA and RB are convolved and then correlated with trace AB for all possible grid locations, within the aperture, to form a multiple-contributing gather (MCG). The MCG is then stacked, and the constructive and destructive interference of the traces, gives the multiple model for the trace, for a particular horizon. All of the multiple-generating horizons must be identified and a multiple model must be created for each of them. The extent and sampling of the MCGs determine the quality of the predicted model. The algorithm based on this concept is suitable for all acquisition geometries whether 2D or 3D, including land orthogonal geometries, and was designed to predict multiple models for more than one generator simultaneously. Figure 2 Schematic for interbed multiple prediction. Figure 3 Plan view of a DRP for target trace SR. To compute the internal multiple model for source X s and receiver X r, (Figure 4), - one source-side trace X s, X 1,i is convolved with all possible receiver-side traces X r,x 2,j, X r X 2,j+1, and so on and correlated with traces X 1,i, X 2,j ; X 1,i,X 2,j+1... respectively, to form an inner MCG. The next sourceside trace X s, X 1,i+1 is chosen and the process of convolving and correlating is repeated with this trace and subsequent traces ( X s, X 1, i+2 ) to form the full MCG (Figure 5a). The unstacked MCG can be visualized as saddle shaped. Stacking inner MCGs forms the outer MCG (Figure 5b). Stacking the outer MCG generates the multiple model for the target trace. Although, the multiple prediction process is a very flexible data-driven technique, it can be adversely affected by poor signal-to-noise and poor spatial sampling especially for shallow multiple generators. Signal to noise ratio can be enhanced by effective data preconditioning, and poor spatial sampling can be addressed by acquiring dense full-azimuth geometries, that fully sample the potential multiple raypath. The multiple generators must be identified using well or VSP data and then interpreted over the survey area. If a process, such as XIMP follows application of DID, then the generating horizon used for the DID prediction is not reused in the later process. Data example A stacked result from synthetic 3D data is shown in Figure 6. The multiple model from XIMP, using horizon 1 as generator shows good correlation with the input multiple. More details of this synthetic example are given by Wu et al., 2011.Our field data example is from a 115-km 2 survey acquired with full-azimuth, square-patch geometry in the northwest Raudhatain region of Kuwait using 3D-point receiver methodology. The full-azimuth square-patch geometry with a shot and receiver line interval of 200m and a station interval of 25 m yielded a maximum fold of over 800.
Figure 4 Plan view of a DRP for target trace SR. Figure 5 a) Unstacked MCG b) Outer MCG Figure 6 Synthetic data stack with multiple-generating horizons and horizon 1 multiple models. This area was chosen to test demultiple techniques, as there is a clear structure and the effects of the demultiple techniques can be clearly seen. In addition, analysis of well logs and modelling of the multiple generators indicated the presence of the strong interbed multiple generators both in shallow and deeper section. The demultiple process was started with application of a 3D true-azimuth surface multiple prediction algorithm (Dragoset et al. 2008). XIMP was used to generate multiples models for all the horizons except the shallowest horizon, which suffered from poor sampling and low signal-tonoise ratio. For the shallowest horizon, DID was used to predict the multiple model. Predicted interbed multiple models were simultaneously and adaptively matched to the input seismic data using least-squares filters and subtracted (El-Emam et al. 2011). Figure 7 shows an example gather and semblance plot that demonstrates that slow as well as fast multiple were removed by the multiple attenuation process. Well ties are an important QC to determine the quality of multiple attenuation. Figure 8 shows an improved well tie with clear multiple attenuation in the reservoir area. a b c d e Figure 7 CMP gather before demultiple(a), after demultiple(b) and difference(c), velocity semblance before(d) and after(e) interbed demultiple.
Figure 8 Well tie before (left) and after (right) interbed demultiple. Conclusions In this paper, we demonstrated complementary approaches to address the challenges of interbed multiple attenuation and showed a case study from a land dataset. In the general area of the data example, it was shown in the past that data have significant challenges due to the presence of multiples. As there is little or no discrimination in velocity or dip, these cannot be easily attenuated using conventional methods based on periodicity and velocity or dip discrimination. It was shown that those multiples respond well to the data-driven techniques such as SRME and IMP. It was proven by various case studies that there is no simple solution and no silver bullet. To address the specific challenges of each survey, a careful analysis of the data is required and usually a combination of methods and approaches is the best solution. DID and XIMP are methods that naturally complement each other, allowing compensation for the limitations of land geometries and addressing the full interval of the multiple generators from top to bottom of the section. Acknowledgements The authors thank WesternGeco for permission to publish this work and Kuwait Oil Company for permission to show the data examples. References Dragoset, B., Moore, I., Yu, M. and Zhao, W. [2008] 3D General surface multiple prediction: An algorithm for all surveys. 79th SEG Annual International Meeting, Expanded Abstracts, 2426-2430. El-Emam, A., Moore, I. and Shabrawi, A. [2005] Interbed multiple prediction and attenuation: Case history from Kuwait. 75th SEG Annual International Meeting, Expanded Abstracts, 448-451. El-Emam, A., Al-Deen, K.S., Zarkhidze, A. and Walz, A. [2011] Advances in interbed multiples prediction and attenuation: Case study from onshore Kuwait. 82nd SEG Annual International Meeting, Expanded Abstracts, 3546-3550. Moore, I. and Bisley, R. [2006] Multiple attenuation in shallow-water situations. 68 th Conference & Exhibition, Extended Abstracts, F018. EAGE Jakubowicz, H. [1998] Wave equation prediction and removal of interbred multiples. 68 th Annual International Meeting, Expanded Abstracts, 1527-1530. SEG Wu, Z.J., Sonika, Dragoset, B. [2011] Robust internal multiple prediction algorithm. 82 nd SEG Annual International Meeting, Expanded Abstracts, 3541-3545.