Abstract:We use the seismic waveform of Hubei seismic network as the data set and the most widely used PhaseNet and EQTransformer models to pick the seismic phase; its performance and generalization ability are tested and evaluated. The results show that for P waves, when the phase probability threshold is 0.1 or 0.3, PhaseNet has a better recall rate than EQTransformer and can detect more microseismic events. Although the recall rate of EQTransformer is slightly lower, the precision rate is higher. The pickup effect of S wave is poorer than that of the P wave. Although the precision rate of PhaseNet is lower than that of EQTransformer, its recall rate is significantly higher, and the F1 value can also be maintained at about 0.8, which has a relatively stable picking performance. We further analyze the relationship between the picking results of the two models and the epicentral distance, signal-to-noise ratio and magnitude of the event. The results show that the phase-picking effect of PhaseNet has a strong correlation with the epicentral distance and the signal-to-noise ratio, but little relationship with the magnitude. The influence of EQTransformer on the signal-to-noise ratio is relatively strong; the higher the signal-to-noise ratio, the better the pickup effect, and the lower the influence on the epicentral distance and magnitude.