Improving the accuracy of predicted first-arrival times using deep learning

Recently, the task of first-arrival time picking for seismic shot gathers has been treated as an image segmentation problem, and deep-learning (DL) algorithms have been successfully used to predict first-arrival times. Currently, researchers mainly focus on leveraging cutting-edge DL algorithms to improve the performance of DL in first-arrival picking. There are few publications addressing the quality control of the results predicted by DL. We develop a three-step workflow to improve the accuracy of first-arrival time detection computed using DL algorithms. First, we obtain three predicted results (generation I) by applying the holistically nested U-net (HU-net) to seismic shot gathers, the envelope of seismic shot gathers, and the cosine of the instantaneous phase of seismic shot gathers. Subsequently, we obtain generation II picking by statistically analyzing the predicted generation I picking. Finally, we treat the first-arrival picking task as a constrained path search problem and the generation II picking function as the constraints. The developed workflow is applied to real seismic surveys to demonstrate its effectiveness.