Physics-based deep learning for non-stationary adaptive multiple suppression

Abstract

Surface-related multiple elimination (SRME) is an indispensable processing step in marine seismic exploration, comprising two main components: multiple prediction and adaptive subtraction. While the multiple prediction is theoretically data-driven and relies on the data itself, the adaptive subtraction step requires a fairly subjective manual control. Our approach introduces a purely data-driven and physics-based deep learning (PDL) algorithm for improving the adaptive multiple subtraction to significantly enhanced performance. There are two main innovations in the proposed deep-learning (DL) multiple subtraction approaches: first, being data-driven and self-supervised, which removes the dependence on inflexible label preparation and model training processes; secondly, being governed by the adaptive matching filtering (AMF) equation to guide the training process. The DL module serves as a strong regression filter analogous to the traditionally known AMF. We segment a moderately sized seismic profile and an initial multiple model into a large number of patches using a patching scheme for self-contained training. The number of patches is controlled by the patch size and the shift size to enable a sufficient number of training samples required by mainstream DL platforms. The patching-based self-training and regression approach, PDL, is inherently a non-stationary matching approach that aims to obtain a globally optimal matching from the initial multiple models to the input full wavefield data through iterative patching-based training. It is also closely connected with the non-stationary regression approach, which was developed almost two decades ago. However, the PDL approach achieves even stronger regression performance due to the backpropagation-based optimization mechanism in DL, and has shown significantly improved multiple subtraction performance over a number of examples.