Regional Fine-Tuning of Pretrained Phase Pickers for Deep-Focus Earthquake Phase Picking in Northeast China

Abstract Pretrained deep-learning phase pickers are now widely used in earthquake monitoring, but their transferability to deep-focus earthquakes recorded by regional seismic networks remains poorly constrained. We evaluate three SeisBench pretrained pickers, PhaseNet, EQTransformer, and generalized phase detection (GPD), using manually picked Northeast China deep-focus earthquake records. The fine-tuning set contains 1638 manually picked arrivals (875 P and 763 S arrivals) from 29 events during 2008–2022, 400 noise windows, and data augmentation applied to S-phase waveforms. The blind test uses 660 analyst-reviewed arrivals from four 2023–2025 events recorded by 135 stations, including 86 stations not used during fine-tuning. At validation-selected thresholds and a ± 1.0 s matching tolerance, fine-tuned PhaseNet, EQTransformer, and GPD yield P-phase F1 scores of 0.893, 0.673, and 0.350, respectively, and S-phase F1 scores of 0.315, 0.377, and 0.252. At a stricter ±0.5 s tolerance, the corresponding P-phase F1 scores are 0.758, 0.601, and 0.195, and the S-phase F1 scores are 0.243, 0.283, and 0.154. Post hoc blind-test F1-optimal thresholds are retained only as upper-bound estimates. Record-level bootstrap intervals, matching-tolerance tests, and leave-one-event-out analyses show that PhaseNet P-phase picking improves most clearly. In contrast, S-phase picking remains the primary empirical limitation. One event located near the edge of the source–station geometry strongly affects the pooled performance metrics for S-phase picking. We therefore interpret this result as an illustrative case suggesting possible sensitivity to source–station geometry, rather than as evidence for a general event-level pattern.