Comparing three feedback internal multiple elimination methods
Song, Jiawen*; Verschuur, Eric; Chen, Xiaohong
Abstract
Multiple reflections have posed a great challenge for current seismic imaging and inversion methods. Compared to surface multiples, internal multiples are more difficult to remove due to poorer move-out discrimination with primaries and we are left with wave equation-based prediction and subtraction methods. In this paper, we focus on the comparison of three data-driven internal multiple elimination (IME) methods based on the feedback model, where two are well established prediction-and-subtraction methods using back-propagated data and surface data, referred to as CFP-based method and surface-based method, respectively, and the third one, an inversion-based method, has been recently extended from estimation of primaries by sparse inversion (EPSI). All these three methods are based on the separation of events from above and below a certain level, after which internal multiples are predicted by convolutions and correlations.
We begin with theory review of layer-related feedback IME methods, where implementation steps for each method are discussed, and involved event separation are further analyzed. Then, recursive application of the three IME methods is demonstrated on synthetic data and field data. It shows that the two well established prediction-and-subtraction methods provide similar primary estimation results, with most of the internal multiples being removed while multiple leakage and primary distortion have been observed where primaries and internal multiples interfere. In contrast generalized EPSI provides reduced multiple leakage and better primary restoration which is of great value for current seismic amplitude-preserved processing. As a main conclusion, with adaptive subtraction avoided, the inversion-based method is more effective than the prediction-and-subtraction methods for internal multiple elimination when primaries and internal multiples overlap. However, the inversion-based method is quite computationally intensive, and more researches on efficient implementation need to be done in future.