P-wave first-motion polarity determination using deep learning

The P-wave first-motion polarity is used to determine the focal mechanisms of small earthquakes. However, because of the recent increase in the number of seismic stations, a large amount of time is needed to manually determine P-wave first-motion polarity. To solve this problem, a technique for identifying the first-motion polarity of a P-wave based on deep learning has been developed (Ross et al., 2018; Hara et al., 2019; Uchide, 2020). However, the software of the previous study is unavailable and it is not possible to know how well the software performed when the geographical regions for training region data and predicting are different. In this study, we developed models of deep learning to determinate of P-wave first-motion polarity. We implemented VGG16 and ResNet, which have been verified in the field of image recognition, and also a basic convolution neural network structure to compare the performance depending on the model structure. Since the size of our local dataset was small, a base model was trained through the large dataset (about 2.5 million SCSN dataset), and fine tuning was performed using the local dataset, which was 37,973 P-wave first-motion polarities were determined by using earthquakes of a local magnitude of 2.0 or greater in the Korean peninsula between 2017 and 2020. To review the performance of deep learning models, we compared the manually and automatically determined P-wave first-motion polarities of the 2021 Korean Peninsula earthquakes that were not used for training. We also compared the focal mechanisms obtained from P-wave first-motion polarities determined by deep learning with those from published catalogs and previous studies.