A New Insight Into Automatic First-Arrival Picking Based on Reinforcement Learning

First-arrival picking is critical for static corrections and near-surface model building. We propose a fully automatic first-arrival picking methodology based on reinforcement learning. Unlike conventional methods, which usually limit the picking to a trace-by-trace procedure followed by a quality control step, we use the energy ratio of the whole shot gather as a reward map and formulate the picking procedure as an Markov decision process (MDP) in which the learning agent aims to find the first-arrival onsets of the shot gather by maximizing the total cumulative reward. Outliers and mis-picks associated with bad or dead traces can be avoided automatically in the proposed method. Both synthetic and real data examples demonstrate that this new technology enables geoscientists to automatically track first arrivals without ever using any prior knowledge, even in the presence of coherent noise and discontinuous first arrivals.