SegPhase: development of arrival time picking models for Japan’s seismic network using the hierarchical vision transformer

Seismic phase picking is a fundamental task in seismology that is crucial for event detection and earthquake cataloging; however, manual analysis is impractical given the scale of modern seismic networks. We present SegPhase, a novel seismic arrival time picking model designed to efficiently process large-scale seismic data recorded by dense seismic networks in Japan. In contrast to conventional convolution-based models, SegPhase employs a hierarchical vision transformer structure that utilizes multi-head self-attention to dynamically focus on important waveform features, such as P- and S-wave onsets, noise, and coda waves. Compared to PhaseNet, the most widely used deep learning model, SegPhase improved arrival time match rates by ~ 11% and detected ~ 15% more events in continuous waveform tests, particularly enhancing the detection of small-magnitude events. Benchmark evaluations demonstrated that SegPhase achieved high classification performance in identifying P- and S-waves. We also examined the threshold of the output probability values when applying SegPhase to continuous waveforms for which the optimal threshold was unknown. By lowering the threshold to 0.1, we observed an increase in the number of detected events without noticeable changes in the hypocenter location error and observed–calculated discrepancies. This was achieved by more effectively utilizing high-probability picks, which further improved phase association. Based on these results, we recommend a threshold of 0.1 to enhance event detection while maintaining accurate arrival times. Our findings demonstrate that SegPhase enables robust arrival picking across diverse datasets and supports high-resolution seismic monitoring.

Graphical Abstract