1:45 PM - 3:15 PM
[SCG55-P11] Grouping of source time functions of crustal earthquakes in Japan using nonlinear dimensionality reduction and hierarchical clustering
Keywords:Source time function, Grouping, Nonlinear dimensionality reduction, Hierarchical clustering, Crustal earthquakes in Japan
The source time functions (STFs) of earthquakes vary from simple with a single peak to complex with multiple peaks. Previous studies based on global earthquake catalogs have suggested that the source location of earthquakes and/or the fault mechanism are related to the complexity of the STFs (e.g., Ye et al. 2018; Yin et al. 2021; Li 2022). In this study, we attempted to group the STFs of approximately 1500 earthquakes in Japan, focusing on their complexity, especially the number of peaks, and investigated the relationship between the complexity of STFs and source information.
The events consist mainly of shallow crustal earthquakes in Japan with magnitudes ranging from 3 to 5. The STFs were estimated as follows. Using the waveform records of an earthquake smaller than the target earthquake as empirical Green's functions, the apparent STF at each observation station was obtained (Yoshida and Kanamori 2023). Then, the combination of multiple point sources was estimated to reproduce the apparent STFs by Kikuchi and Kanamori (1982, 1986, 1991). Finally, the STFs were obtained from the spatiotemporal combination of multiple point sources.
The STFs were grouped as follows. First, Uniform Manifold Approximation and Projection (Mclnnes et al. 2020), a graph-based nonlinear dimensionality reduction technique, was applied to the preprocessed STFs to compress the high-dimensional information of the STFs down to the 2D information. This can produce an 2D map in which earthquakes with similar STFs are embedded in close. Euclidean distance was used as the distance function for the dimensionality reduction. As the pre-processing step for the dimensionality reduction, the STFs were normalized according to theier Mo, were resampled to the 1/3 power of Mo, and were multiplied by the sampling interval, so that the STFs can be compared even among earthquakes of different magnitudes. Hierarchical clustering based on distance on the 2D map was applied to obtain a hierarchical structure of earthquake clusters with similar STFs. Euclidean distance was used as the distance function in the hierarchical clustering, and Ward's method was used as the distance measurement method between clusters.
As a result, we obtained earthquake clusters with similar STFs. Almost half of thel events belong to earthquake groups with simple STFs consisting of a single peak, and the other half belong to earthquake groups with complex STFs with multiple peaks. Among the earthquake groups with complex STFs, we were able to divide them into those with a two-peak type having a large first peak, those with a two-peak type having a large second peak, and others. Moreover, the earthquake groups with the two-peak types can be divided into the groups with the time difference of the peaks.
We investigated the dependence of earthquake source information within each earthquake group, and found no clear relationship between the complexity of the STFs and the earthquake magnitude, the source location, or the fault mechanism. This suggests that the rupture process of crustal earthquakes occurring within the island arc of Japan can take a variety of styles, regardless of the region. We also found that earthquake groups with multiple-peak STFs tended to have larger values of radiated energy enhancement factor (REEF), which is one measure of the STF complexity (Ye et al. 2018).
The events consist mainly of shallow crustal earthquakes in Japan with magnitudes ranging from 3 to 5. The STFs were estimated as follows. Using the waveform records of an earthquake smaller than the target earthquake as empirical Green's functions, the apparent STF at each observation station was obtained (Yoshida and Kanamori 2023). Then, the combination of multiple point sources was estimated to reproduce the apparent STFs by Kikuchi and Kanamori (1982, 1986, 1991). Finally, the STFs were obtained from the spatiotemporal combination of multiple point sources.
The STFs were grouped as follows. First, Uniform Manifold Approximation and Projection (Mclnnes et al. 2020), a graph-based nonlinear dimensionality reduction technique, was applied to the preprocessed STFs to compress the high-dimensional information of the STFs down to the 2D information. This can produce an 2D map in which earthquakes with similar STFs are embedded in close. Euclidean distance was used as the distance function for the dimensionality reduction. As the pre-processing step for the dimensionality reduction, the STFs were normalized according to theier Mo, were resampled to the 1/3 power of Mo, and were multiplied by the sampling interval, so that the STFs can be compared even among earthquakes of different magnitudes. Hierarchical clustering based on distance on the 2D map was applied to obtain a hierarchical structure of earthquake clusters with similar STFs. Euclidean distance was used as the distance function in the hierarchical clustering, and Ward's method was used as the distance measurement method between clusters.
As a result, we obtained earthquake clusters with similar STFs. Almost half of thel events belong to earthquake groups with simple STFs consisting of a single peak, and the other half belong to earthquake groups with complex STFs with multiple peaks. Among the earthquake groups with complex STFs, we were able to divide them into those with a two-peak type having a large first peak, those with a two-peak type having a large second peak, and others. Moreover, the earthquake groups with the two-peak types can be divided into the groups with the time difference of the peaks.
We investigated the dependence of earthquake source information within each earthquake group, and found no clear relationship between the complexity of the STFs and the earthquake magnitude, the source location, or the fault mechanism. This suggests that the rupture process of crustal earthquakes occurring within the island arc of Japan can take a variety of styles, regardless of the region. We also found that earthquake groups with multiple-peak STFs tended to have larger values of radiated energy enhancement factor (REEF), which is one measure of the STF complexity (Ye et al. 2018).