[S-CG50] Driving Solid Earth Science through Machine Learning

[SCG50-07] Detection of slow slips from seismic wave records using random forest

*Kazuki Ohtake¹, Aitaro Kato¹, Yutaro Okada², Takuya NISHIMURA³ (1.Earthquake Research Institute, The University of Tokyo, 2.Graduate School of Science, Kyoto University, 3.Disaster Prevention Research Institute, Kyoto University)

Keywords:Machine Learning, Nankai Trough, SSE, GNSS

At the plate boundary of the Nankai Trough, slow earthquakes occur in the transition zone between locked and deep stable sliding zone. Slow earthquakes consist of slow slip events (SSEs) observed in the geodetic time scale and low-frequency earthquakes (LFEs) and tremors observed in the seismic time scale, which have been associated with huge earthquakes (Obara and Kato, 2016). In recent years, short-term SSEs (S-SSEs) with durations of days have been detected using GNSS data (Nishimura et al., 2013, Okada et al., 2022), but S-SSEs, especially small events, are difficult to detect and might be overlooked because GNSS data are less sensitive than tiltmeters and strainmeters. On the other hand, the synchronization between SSEs and tremors was discovered in the Cascadia subduction zone (Rogers and Dragert, 2003) and in southwest Japan (Obara et al., 2004). Rouet-Leduc et al. (2019) showed that machine learning can be applied to continuous seismic waveform records in the Cascadia subduction zone to estimate temporal changes of the displacement rate of GNSS stations synchronized with SSE events. In this study, our goal is to estimate the temporal changes of displacement rate of GNSS stations in western Shikoku by machine learning using continuous seismic waveform records. For seismic waveform records, we calculated statistical features using data from Hi-net stations in western Shikoku from April 2004 to March 2020 for each day. For geodetic data, we calculated GNSS displacement rate for GEONET stations around western Shikoku using the method of Okada et al. (2022). To enhance the GNSS signal, we computed the sum of inner products of the observed displacement rate with the theoretical displacement during a stacked SSE (Kano et al., 2019), according to Bletery and Nocquet (2023). Using random forest, an ensemble of decision trees, we constructed a model that estimates the temporal changes of GNSS displacement rate from statistical features of continuous seismic waveform records. The results show that the GNSS displacement rate estimation of our model is highly correlated with the actual GNSS displacement rate. Our model detected responses of displacement rate considered as S-SSEs form changes of seismic waveform records, including events that are not in the S-SSE catalog of Okada et al. (2022). This study indicates that we may detect SSEs in the Nankai Trough more accurately by machine learning using continuous seismic waveform records.