Improving Deep Learning-Based Seismic Phase Picking by Addressing Label Imbalance

Deep learning-based phase-picking models have significantly improved seismic waveform analyses efficiency. However, detection performance degrades when multiple seismic phases occur within a single waveform trace, particularly during intense seismic sequences such as aftershock activity. This may be attributed to pronounced label imbalance in semantic-segmentation-based phase picking, where narrow P- and S-phase label regions are overwhelmed by background label. To address this, we introduce an area-weighted soft cross-entropy loss that emphasizes phase-arrival regions during training without modifying model architecture. The proposed loss function was evaluated using two phase-picking architectures, SegPhase and PhaseNet. Experiments on the test dataset demonstrated that the weighted loss function reduced the scatter of arrival time residuals and improved the temporal consistency of predicted picks. Weighted-loss models also exhibited more recall-oriented behavior, reducing missed arrivals, though at the cost of increased false positives. Application to continuous waveform data from the 2019 Ridgecrest earthquake sequence showed that the weighted-loss models captured the main spatial and temporal characteristics of the seismicity while detecting more events than the original-loss models. These results demonstrate that loss function design constitutes an effective approach for improving the sensitivity of deep learning-based phase picking and enhancing event recovery in continuous waveform analysis.