Deep Learning Ensemble for Seismic First-Break Event Picking

Abstract

Summary In this study, we investigate the use of an ensemble of deep learning models to improve the quality and efficiency of seismic first break event picking. In traditional workflows, we often perform first break picking iteratively and spend extensive quality control and refining efforts before arriving at the final result. If we ask more than one expert to perform the picking, we expect the picks to be consistent between these experts at most locations, whereas the discrepancy only occurs at very challenging locations. Therefore, by taking only the picks that are consistent across multiple experts can provide us the high confidence picks. Following a similar strategy, we use multiple deep learning models that act as independently experts, then take only the first break picks that are consistent among these models. We can greatly reduce the uncertainty involved in the picks and reduce the amount of quality control needed. Compared with traditional workflow, the proposed strategy provides more valid picks that benefit the near surface velocity model building process, at a fraction of the cost.