Abstract

Reliable seismic phase picking is essential to earthquake monitoring, as it fundamentally affects earthquake location and source inversion. In challenging field conditions, nonstationary waveforms, diverse morphological features and intense background noise all hinder the detection of weak phases. Seismic time series also exhibit inherent spatiotemporal asymmetry. Nevertheless, mainstream networks rely on symmetric and uniform feature extraction strategies. They overlook asymmetric properties of waveforms and introduce additional picking errors. We therefore present SymPhase, a symmetry-aware collaborative attention network, to achieve precise and robust P- and S-phase picking. Using a 1D encoder–decoder backbone, the model combines global enhancement and local refinement. It captures both long-range dependencies and local features, reducing missed weak-phase detections and minimizing arrival-time bias. Extensive tests are conducted on the CEED and DiTing datasets. The results demonstrate that SymPhase outperforms both TCN and PhaseNet. On the CEED dataset, the F1 scores for P and S phases are 0.9797 and 0.9006, with mean absolute errors of 0.0761 s and 0.1003 s. On the difficult DiTing dataset, the S-phase F1 score reaches 0.4724 with a corresponding error of 0.7386 s. These results validate its superior performance for weak signal recognition. With strong accuracy and noise robustness, SymPhase provides a viable solution for automated earthquake monitoring systems.

IPC Classification

Keywords