Repairing GSMP estimated multiples under coarse sampling using deep learning

Abstract

The data-driven surface-related multiple elimination (SRME)-type approach requires fully sampled sources and receivers during the multidimensional convolution process. Otherwise, the estimated multiples will be aliased. Compared to expensive reconstruction processes before prediction, dealiasing on the estimated multiples from limited sources might provide a potential easier solution in a 2D scenario, where deep learning (DL)-based methods suit well for this highly non-linear problem. Unfortunately, DL-based multiple dealising will not function well for 3D data due to extremely coarse sampling in either source or receiver side. Thus, data interpolation/reconstruction is the only option, though the performance might not be desired. Generalized surface multiple prediction (GSMP) is the most used on-the-fly interpolation approach in 3D. Still, GSMP accuracy heavily relies on the existing traces. When fed with coarsely sampled recorded data only, GSMP tends to generate multiples with low amplitude and distorted phase, especially for small offsets. We propose a U-Net framework to repair GSMP estimated multiples such that the amplitude loss and distorted phase can be restored. In this way, the strong non-linear mapping power from DL can help repair the GSMP estimated multiples.