Identification of Seismic Wave First Arrivals from Earthquake Records via Deep Learning

For seismic location and tomography, it is important to pick P- and S-wave first arrivals. However, traditional methods mainly determine P- and S-wave first arrivals separately from a signal processing perspective, which requires the extraction of waveform attributes and tuning parameters manually. Also, traditional methods suffer from noise as they are operated on the whole earthquake record. In this paper, we propose a deep neural network framework to enhance picking P- and S-wave first arrivals from a sequential perspective. Specifically, we first transform the picking first arrival problem as a sequence labelling problem. Then, the rough ranges for P- and S-wave first arrivals are determined simultaneously through the proposed deep neural network model. Based on these rough ranges, the performance of existing picking methods can be greatly enhanced. Experimental results on two real-world datasets demonstrate the effectiveness of the proposed framework.