Machine-learning detection of slow slip events in western Shikoku, Japan, through joint analysis of seismic and geodetic data

Abstract Over the past decade, short-term slow slip events (S-SSEs) have been detected along the entire Nankai Trough using Global Navigation Satellite System (GNSS) data. To enhance the detection of S-SSEs, we focused on the spatial and temporal coincidence of tremors and S-SSEs, a phenomenon known as episodic tremor and slip. We developed a machine learning-based method to detect S-SSEs directly from continuous seismic waveforms and applied it to seismic and geodetic data in western Shikoku, Japan. We trained a random forest regression model using statistical features extracted from continuous seismic waveforms as input variables and GNSS-derived displacement rates as target outputs. We predicted the GNSS displacement rate over a period of ~ 6 years and defined S-SSEs as periods when the predicted GNSS displacement rate increased sharply. We then estimated fault models for each detected S-SSE. The predicted displacement rates were correlated strongly with the observed displacement rates, and we identified a total of 23 S-SSEs, including 5 previously unrecognized events. The results demonstrate the effectiveness of machine learning using continuous seismic waveforms for improving S-SSE detection along the Nankai Trough. Graphical Abstract