Estimate near-surface velocity with reversals using deep learning and full-waveform inversion: A field-data-driven method

Abstract

Estimating shallow velocity from land seismic data is challenging due to complexities in the near surface. One such complexity arises from velocity reversals, which can be caused by anomalous high/low-velocity layer intrusions into the otherwise increasing velocity-depth trend. Velocity reversal imposes significant challenges for conventional velocity model building methods to accurately estimate near surface models. Turning ray tomography or tomostatics relies on accurate picking of first breaks, which becomes infeasible when velocity reversals generate shingling events. In this paper, we proposed a field-data-driven deep learning framework, which combines field-specific training, deep learning and full-waveform inversion, to estimate the near surface velocity model from land seismic data. Instead of training over purely random synthetic data, we propose to incorporate a-priori knowledge about the field under study to generate field-specific data set for training the proposed deep neural network. In addition, we combine the deep learning model with full-waveform inversion to further improve the near-surface velocity model. Synthetic and field data examples show that the proposed framework can successfully handle the velocity reversal complexity in land seismic data and achieve promising performance on estimating near surface velocity models.