An Attention-Based Deep Learning Method for Continuous-Interval Seismic Phase Picking

Seismic phase picking can degrade under low signal-to-noise ratio (SNR) conditions and when waveform components are unavailable. We modify PhaseNet by adding long short-term memory units to the encoder, parallel attention to the skip connections, and a convolutional block attention module to the decoder. We also use interval labels: samples from the P arrival to the S arrival are assigned to the P segment, and samples from the S arrival to P + 3(S − P) are assigned to the S segment. The model was trained on the Stanford Earthquake Dataset (STEAD) and evaluated on STEAD, the Italian seismic dataset for machine learning (INSTANCE), and the Texas Earthquake Dataset. At a 0.5 s tolerance on STEAD, the model obtained an S-wave F1 score of 0.980, compared with 0.869 for the retrained PhaseNet baseline, 0.969 for SEA, and 0.974 for EQTransformer. On INSTANCE, the model obtained P/S F1 scores of 0.909/0.803, compared with 0.863/0.741 for PhaseNet. Tests with missing components show that the model retains useful performance when one component is removed, but performance decreases when two components are removed, especially when the vertical component is absent. These results indicate improved phase picking relative to PhaseNet, while also showing sensitivity to cross-dataset shifts and severe channel loss.