11:45 〜 12:00
[SSS10-09] Spatio-temporal foreshock evolution of crustal earthquakes in Japan and southern California
キーワード：前震活動、地震検出、機械学習
The physical process of earthquake nucleation is unsettled. The nucleation mechanism, the preslip model, the cascade-up model, and their combined model were proposed based on the observations, the numerical simulations, and the laboratory experiments (McLaskey, 2019). In the preslip model, the foreshocks and the mainshock are induced by a background aseismic slip, whose area scales with the mainshock magnitude. In the cascade-up model, since a small-scale rupture incidentally induces a large-scale rupture, it is hard to distinguish them in the early stage. Current studies show observations in support of both the preslip model (e.g. Kato et al., 2016) and the cascade-up model (e.g. Zhu et al., 2022; Ide, 2019). Therefore, investigating what causes these differences is important for understanding earthquake rupture evolution. In this regard, we comprehensively investigated the spatio-temporal foreshock evolutions of the crustal earthquakes in Japan and southern California.
Specifically, the Hi-net stations for Japan and SCSN (Southern California Seismic Network) for Southern California were used in this study. We used the machine-learning method of PhaseNet (Zhu and Beroza, 2019) for the phase picking, which is based on the U-Net convolutional network architecture. We first picked P and S phase arrivals in a week before the mainshock with a magnitude greater than 5.0. Second, we associated the detected phases by the Gaussian Mixture Model Association (GaMMA; Zhu et al., 2022). Third, we estimated the absolute hypocenter locations by the HypoSVI (Smith et al., 2022) solving the Eikonal equation, and the absolute hypocenters were relocated. Finally, we evaluated the spatio-temporal foreshock distributions from a view of clustering and anisotropy (Ross et al., 2022). The results show that the number of detected foreshocks in southern California is similar to or more than the QTM catalog for each mainshock, which was constructed by the template matching method (Ross et al., 2019; Trugman and Ross, 2019). Moreover, for Japanese data, we detected at least twice more foreshocks than the JMA catalog.
Specifically, the Hi-net stations for Japan and SCSN (Southern California Seismic Network) for Southern California were used in this study. We used the machine-learning method of PhaseNet (Zhu and Beroza, 2019) for the phase picking, which is based on the U-Net convolutional network architecture. We first picked P and S phase arrivals in a week before the mainshock with a magnitude greater than 5.0. Second, we associated the detected phases by the Gaussian Mixture Model Association (GaMMA; Zhu et al., 2022). Third, we estimated the absolute hypocenter locations by the HypoSVI (Smith et al., 2022) solving the Eikonal equation, and the absolute hypocenters were relocated. Finally, we evaluated the spatio-temporal foreshock distributions from a view of clustering and anisotropy (Ross et al., 2022). The results show that the number of detected foreshocks in southern California is similar to or more than the QTM catalog for each mainshock, which was constructed by the template matching method (Ross et al., 2019; Trugman and Ross, 2019). Moreover, for Japanese data, we detected at least twice more foreshocks than the JMA catalog.