Abstract: We applied PhaseNet, a deep learning seismic phase detection and arrival time picking algorithm to the data from the Sichuan Network and the Capital Region Network and tested its generalization ability. Based on 126761 event waveforms corresponding to 29328 events (M_L0~M_L4) recorded by 199 seismic stations in the Capital Region Network from January 2010 to October 2018, and 120233 event waveforms corresponding to 16595 events (M_L0~M_L6) recorded by 227 seismic stations in Sichuan and neighboring provinces from April to September 2019, we established SC and CA test datasets, and used the pre-trained PhaseNet model for automatically detecting P、 S phases and picking up arrival time. The test shows that PhaseNet has good seismic phase detection capability, the F₁ value of P and S detection exceeds 0.75 (threshold value is 0.5 s), and has a relatively stable and accurate ability to pick up P arrival time (threshold value 0.1 s), where the F₁ value exceed 0.6 on both datasets, and the F₁ values for S phase picking are 0.33 (SC) and 0.53 (CA). We further analyzed the relationship between the test results and the epicentral distance, magnitude, signal-to-noise ratio, and the location of the station, which not only defined the scope of the current PhaseNet model, but also indicated the direction for the next step to continue training more optimized models. The above results show that the PhaseNet algorithm have great application potential in the automatic seismic phases detection of regional networks and can currently provide assistance for the automatic cataloging work.

Key words: PhaseNet, Generalization ability, Arrival time picking, Phase detection