Automated hypocenter determination of tectonic tremors in the Nankai subduction zone using convolutional neural networks combined with semblance analysis

Abstract Recent advances in deep learning have enhanced our ability to analyze seismic waveforms. Here, we developed and evaluated a convolutional neural network (CNN) model to classify tectonic tremors, earthquakes, and noise in seismic waveform data recorded by a seismic array in the Nankai subduction zone. The trained CNN model achieved high accuracy, with both precision and recall exceeding 97%, and correctly detected 96% of distant earthquakes. The probability of tectonic tremor as a function of the signal-to-noise ratio (SNR) increased steeply from 10 to 90% at an SNR of 4. We highlighted tectonic tremor waveforms using the integrated gradients (IG) method for interpreting CNN models. IG filter averaging over the stations of an array outperforms bandpass filters and other interpretation methods for CNN models in locating tectonic tremors by semblance analysis, providing the largest number of tectonic tremor sources. As reported previously, located sources of tectonic tremor during episodic tremor and slip events migrate along the strike of the subducting plate. The source location error increases significantly at epicentral distances greater than 30 km because of low SNRs. The technique developed in this study equips CNN models with a high ability to distinguish tectonic tremors and earthquakes from noise and to locate tectonic tremors with sources that are not far from seismic stations. Graphical abstract