4:30 PM - 4:45 PM
[SCG60-09] Methodological development of fault slip detection from GNSS using deep learning and application to western Shikoku
Fault slip phenomena in subduction zones are crucial for understanding the processes of stress accumulation and release at plate boundaries. In particular, slow slip events (SSEs) have been observed preceding some large earthquakes (e.g., Voss et al., 2018), necessitating a detailed understanding of their spatiotemporal characteristics to investigate interactions with fault slip associated with major earthquakes. In this study, we develop a novel deep learning-based method to detect subtle fault slips from noisy GNSS data and verify its effectiveness through application to GNSS observations in western Shikoku, Japan.
The proposed method consists of two convolutional neural networks (CNNs). The first CNN, referred to as the interpolator, estimates a high-resolution noise-reducted smoothed horizontal displacement field from static GNSS displacement data. The second CNN, referred to as the estimator, estimates the slip distribution on the plate interface based on the displacement field obtained from the interpolator. To enhance the detection capability and robustness of the model, we trained the CNN models using synthetic data that simulate various patterns of slip on the plate interface and various noise patterns.
We applied the proposed method to the GNSS data during the periods of short-term SSEs reported in previous studies (Sekine et al., 2010; Okada et al., 2022). As a result, the method estimated slip distributions consistent with prior studies for 42 out of 47 cataloged events (89%). Furthermore, by applying a moving-window approach to daily GNSS coordinate time series from 1996 to 2023, we conducted a spatiotemporal analysis of slip evolution. We successfully detected multiple previously unreported short-term SSEs, many of which exhibited spatiotemporal synchronization or adjacency with tremor activity, similar to previously known events. These findings suggest that the proposed method has a high capability for accurately detecting actual SSE occurrences.
Acknowledgements
In this study, we utilized daily coordinate data from GEONET observation stations of the Geospatial Information Authority of Japan (GSI), which were processed from RINEX data by Professor Takuya Nishimura of the Disaster Prevention Research Institute, Kyoto University.
The proposed method consists of two convolutional neural networks (CNNs). The first CNN, referred to as the interpolator, estimates a high-resolution noise-reducted smoothed horizontal displacement field from static GNSS displacement data. The second CNN, referred to as the estimator, estimates the slip distribution on the plate interface based on the displacement field obtained from the interpolator. To enhance the detection capability and robustness of the model, we trained the CNN models using synthetic data that simulate various patterns of slip on the plate interface and various noise patterns.
We applied the proposed method to the GNSS data during the periods of short-term SSEs reported in previous studies (Sekine et al., 2010; Okada et al., 2022). As a result, the method estimated slip distributions consistent with prior studies for 42 out of 47 cataloged events (89%). Furthermore, by applying a moving-window approach to daily GNSS coordinate time series from 1996 to 2023, we conducted a spatiotemporal analysis of slip evolution. We successfully detected multiple previously unreported short-term SSEs, many of which exhibited spatiotemporal synchronization or adjacency with tremor activity, similar to previously known events. These findings suggest that the proposed method has a high capability for accurately detecting actual SSE occurrences.
Acknowledgements
In this study, we utilized daily coordinate data from GEONET observation stations of the Geospatial Information Authority of Japan (GSI), which were processed from RINEX data by Professor Takuya Nishimura of the Disaster Prevention Research Institute, Kyoto University.