SegPhase: Development of Arrival Time Picking Models for Japan’s Seismic Network Using the Hierarchical Vision Transformer

We present SegPhase, a novel seismic arrival time picking model designed to efficiently process large-scale seismic data recorded by Japan’s dense seismic networks. SegPhase is implemented in three configurations to accommodate various observation settings. 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. SegPhase improved the match rate of arrival time picks with human analysts by approximately 11% in the test dataset. Additionally, when applied to continuous waveforms, it detected approximately 15% more seismic events than the most commonly used deep learning model, PhaseNet, particularly enhancing the detection of small events (M0.6 or less). Benchmark evaluations demonstrated that SegPhase achieved high classification performance, almost 100% distinguishing between seismic signals and noise and accurately identifying different seismic wave phases. Additionally, arrival time picking exhibited low residuals, with Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) values of 0.03 and 0.06 for P-waves and 0.06 and 0.10 for S-waves, respectively. We also examined the threshold of output probability values when applying SegPhase to continuous waveforms for which the optimal threshold is not known. By lowering the threshold to 0.1, we observed an increase in detected events without significant changes in hypocenter location error and observed-calculated (O-C) discrepancies. Furthermore, arrival times with high output probability were more effectively utilized. Based on these results, we recommend setting the threshold to 0.1 to enhance event detection while maintaining robust arrival time accuracy. Our findings demonstrate that SegPhase effectively picks seismic arrivals across diverse datasets, enhances event detection, and supports high-resolution seismic monitoring.