Automatic P-wave Detection, Phase Picking, and Polarity Determination based on Deep Learning

P-wave arrival time and first-motion polarity are fundamental observations in seismology, which are used to determine hypocenter locations and focal mechanisms of earthquakes. In this study, we develop three convolutional neural network (CNN) models that perform P-wave event detection (E-Taro), phase picking (P-Jiro), and first-motion polarity determination (F-Saburo). In training and testing the CNN models, we use about 130 thousand 250 Hz and about 40 thousand 100 Hz waveform data observed in western Japan. For the 250 Hz (100 Hz) waveform data, E-Taro has the accuracy of 98.1% (97.3%); the difference between the arrival times determined by human experts and PJiro is -0.005 s (-0.012 s) in average with a standard deviation of 0.038 s (0.077 s); F-Saburo has the accuracy of 97.9% (95.4%). Finally, by applying the three CNN models to continuous waveform data, we showed the arrival time and first-motion polarity can be obtained without helps of human experts. The results of the CNN models are in good agreement with human experts