Complex Near-surface Velocity Modeling via U-net

Abstract

Summary Accurate near-surface velocity structure is the key to improve the precision of statics and seismic imaging. We propose a novel method for complex near-surface velocity modeling based on a modified U-net from pre-stack seismic data. The method makes use of waveform information rather than travel time only. We design a number of complex near-surface velocity models and simulate shot gathers using the finite difference scheme. During the forward stage, the network develops a nonlinear relationship between the multi-shot seismic data and the corresponding velocity models. During the inversion stage, the trained network can be used to predict velocity models from the new shot gathers in a few minutes. Supported by numerical experiments on synthetic models, this method achieve a promising performance in complex near-surface velocity inversion.