Seismic detection based on unsupervised station-wise phase picks using deep learning

Recently, deep learning has gained much attention for seismic detection, replacing the conventional STA/LTA method based on abrupt change of amplitude. Due to the powerful representation capability of neural networks, a deep learning model is remarkably flexible for fitting to waveform data, which allows for high-performed seismic detection. On top of that, a further extension is to apply deep learning model to multi-station framework. Using full information of waveforms observed at various stations, the multi-station framework can largely enhance EQ detection capability. Typically, seismic detection is performed in a station-wise manner, which is in turn combined to yield a network-based detection. However, a conventional approach involves complicated tuning of relevant parameters, which is not straightforward to optimize. In the present study, we propose a simple but effective method for a network-based detection, which requires neither labeled seismic data nor complicated parameter tuning. We considered to combine station-wise phase picks (P-phase and S-phase) yielded by a pre-trained deep learning model. Within a time-window, we defined a propensity score for earthquake based on maximum values of station-wise phases. For the propensity score, both the number of stations and cutoff value were determined in an unsupervised manner. We applied the proposed method to one-week continuous waveforms from Metropolitan Seismic Observation network (MeSO-net), which was obtained in the bustling Tokyo metropolitan area. It was demonstrated that the proposed method most effectively performed for such noise-contaminated waveforms.