2:30 PM - 2:45 PM
[SCG45-28] Detection of slow slip events from continuous seismic waveforms in western Shikoku based on random forest model
Keywords:Nankai Trough, SSE, Seismic Waveform, Machine Learning
Along the plate boundary of the Nankai subduction zone, slow earthquakes occur in transitional zones between locked and deep stable sliding regimes. Although previous studies using GNSS data (Nishimura et al., 2013; Okada et al., 2022) have attempted to construct a comprehensive catalog of slow slip events (SSEs), detecting small SSEs remains challenging due to the lower sensitivity of GNSS data compared to tilt-meters and strain-meters. SSEs and tremors have been observed to synchronize in the Cascadia and Nankai subduction zones (e.g., Rogers and Dragert, 2003; Obara et al., 2004). Leveraging this synchronization, Rouet-Leduc et al. (2019) demonstrated that machine learning can predict the temporal change in the displacement rate of GNSS stations associated with SSEs using continuous seismic waveform records along the Cascadia subduction zone. However, this approach has yet to be tested in different tectonic settings or validated for extracting slip evolution along plate boundary faults. Here, we aim to detect SSEs in western Shikoku, Japan, by applying the machine learning approach to continuous seismic waveform records.
We calculated a time series of statistical features per day using seismic waveform records obtained from Hi-net stations in western Shikoku from April 2004 to December 2023. For geodetic data, we calculated GNSS displacement rate for GEONET stations around western Shikoku using the method of Okada and Nishimura (2023) and Okada (2024, doctoral thesis). To enhance the GNSS signal, we computed the sum of the inner products of the observed displacement rate with the theoretical displacement during stacked SSEs (Kano et al., 2019), according to Bletery and Nocquet (2023). Using random forest, an ensemble of decision trees, we predicted the temporal change in the GNSS displacement rate from statistical features of continuous seismic waveform records. We defined an SSE as an event where the predicted stacked GNSS displacement rate exceeds a threshold. For each detected SSE, surface displacement was computed by fitting the time series of each horizontal GNSS component with sum of a ramp function and a linear function. We then applied a nonlinear inversion method (Matsu’ura and Hasegawa, 1987, Nishimura et al., 2013) to the derived surface displacement field to estimate the finite fault model of each SSE.
During training, our model's predictions of GNSS displacement rate time series showed a strong correlation with observed data. By analyzing the temporal change in seismic waveform records, the model successfully detected displacement rate variations associated with SSEs. While most detected events correspond to those listed in the existing SSE catalog (Okada et al., 2022), a few previously unlisted events were also identified. These findings highlight the potential for more accurate SSE detection and a more comprehensive characterization of the slip spectrum along the Nankai subduction zone.
We calculated a time series of statistical features per day using seismic waveform records obtained from Hi-net stations in western Shikoku from April 2004 to December 2023. For geodetic data, we calculated GNSS displacement rate for GEONET stations around western Shikoku using the method of Okada and Nishimura (2023) and Okada (2024, doctoral thesis). To enhance the GNSS signal, we computed the sum of the inner products of the observed displacement rate with the theoretical displacement during stacked SSEs (Kano et al., 2019), according to Bletery and Nocquet (2023). Using random forest, an ensemble of decision trees, we predicted the temporal change in the GNSS displacement rate from statistical features of continuous seismic waveform records. We defined an SSE as an event where the predicted stacked GNSS displacement rate exceeds a threshold. For each detected SSE, surface displacement was computed by fitting the time series of each horizontal GNSS component with sum of a ramp function and a linear function. We then applied a nonlinear inversion method (Matsu’ura and Hasegawa, 1987, Nishimura et al., 2013) to the derived surface displacement field to estimate the finite fault model of each SSE.
During training, our model's predictions of GNSS displacement rate time series showed a strong correlation with observed data. By analyzing the temporal change in seismic waveform records, the model successfully detected displacement rate variations associated with SSEs. While most detected events correspond to those listed in the existing SSE catalog (Okada et al., 2022), a few previously unlisted events were also identified. These findings highlight the potential for more accurate SSE detection and a more comprehensive characterization of the slip spectrum along the Nankai subduction zone.