Optimized earthquake analysis procedures based on deep learning phase picker

Recently, the number of seismic stations has increased, and a dense temporary seismic network has been installed to monitor seismicity in the area of interest. As a result, a large amount of seismic data was generated. To analyze these, several seismic processing systems based on deep learning have recently been developed and are reported to be more effective than traditional system. This is due to the use of various deep learning based phase pickers and event associator using various techniques. However, deep learning model can be very sensitive to small perturbations in the data, which are also seen in deep learning phase pickers. It provides different results depending on the selection of the window in seismograms. Also in the event association, if there is a lot of noises in the detected phases, it can lead to falsely associate events. Therefore, in order to analyze seismic data effectively, these problems should be solved. In this study, we have developed earthquake analysis procedures. First of the procedures, we detect phases using PhaseNet (Zhu and Beroza, 2019). Second, we associate the phase with binder_max (Sheen and Friberg, 2021) using maximum likelihood technology based on grid search. Then event location was used hypoinverse (Klein, 2002). For PhaseNet optimization, we used more overlapping input data than original PhaseNet. Using the 2020 Haenam earthquake sequence in South Korea and the 2019 Ridgecrest earthquake sequence in Southern California, United States, we compared associators binder_max and GaMMA (Zhu et al., 2021).