An Efficient Transfer Learning-Based OBS Seismic Phase Picker (OBSPD) Trained on Cascadia Subduction Zone Earthquake Dataset

Abstract

Abstract Earthquake monitoring and many seismological studies depend on seismic phase arrivals. Thus, detecting seismic events and picking the phase arrival times are fundamentally important. In the recent years, seismic phase picking models based on deep learning approaches have been widely developed. These deep learning models can achieve better performances than traditional phase picking methods and improve the quality of phase picking for land-based earthquake monitoring. However, these models might not perform well on data from ocean-bottom seismometers (OBSs), because they are trained exclusively using onshore seismic data and have limited out-of-distribution generalization ability. Nevertheless, there are insufficient labeled OBS phase arrivals dataset to train a deep learning model from scratch. In this study, we developed an automatic phase detection model for OBS data (OBS phase detection [OBSPD]) using the transfer learning approach based on an existing U-GPD model with pretrained weights from a generalized phase detection model feature extraction system. We developed OBSPD with a limited amount of training data (2784 three-component event waveforms) from the Cascadia subduction zone (CSZ) OBS deployments. Our results show that transfer learning can achieve lower model loss with less overfitting compared to when training a model from scratch. Our new OBSPD model outperforms four existing deep learning pickers in terms of phase detection accuracy with smaller arrival time residuals on a test OBS dataset at CSZ, especially for P phases.