11:45 〜 12:00
[SCG45-47] Anomaly Detection Using a Probabilistic Model Describing Spatiotemporal Patterns of Tectonic Tremors: Application to Western Japan
キーワード：スロー地震、テクトニック微動、統計モデル、異常検知
Changes in slow earthquake patterns are thought to be related to the preparation phase of huge earthquakes (e.g., Matsuzawa et al., 2010). Modeling the standard behavior of slow earthquakes may lead to an assessment of the risk of major earthquake occurrence. Tectonic tremors have higher detection frequency and hypocenter determination accuracy, providing valuable clues for investigating the detailed evolution of slow earthquakes. While previous studies have focused on the temporal information (e.g., Lengliné et al., 2017; Ide & Nomura, 2019), considering the spatial characteristics, such as inhomogeneous distributions and migrations, as well may improve the model expression capability.
In this study, we first confirmed the importance of spatial information. We considered four factors contributing to the tremor occurrence: 1. steady background seismicity, 2. self-triggering effect acting in a narrow space, 3. adjacent-triggering effect from neighboring areas, and 4. long-term recurrence effect. As combinations of these, we constructed three types of temporal models (1 only, 1+2, 1+2+4), and two types of spatiotemporal models (1+2+3, 1+2+3+4). We applied and compared these five models to the tremor catalog in western Japan, which was constructed using the method of Yano and Ide (2024). The analysis period was from April 2004 to December 2023. When comparing the models using Akaike Information Criterion, the multivariate Hawkes model consisting of background seismicity, self-triggering, and adjacent-triggering (1+2+3) was found to be statistically superior. This model captured tremor migration patterns that move distances of about 5 to 10 km on time scales ranging from sub-hour to half a day; however, the direction of such short-term migration does not necessarily align with the direction of migration in longer time window.
We also investigated the contributions of each effect in the multivariate Hawkes process. The model revealed that when the grid cell size is 5 km in diameter, self-triggering accounts for an average of about 25%, reaching as high as 50%, followed by adjacent-triggering at around 11%. As the cell size increases, the contribution of self-triggering rises to an average of about 50%, with a maximum of approximately 80%, while adjacent-triggering decreases to about 7%. This cell-size dependency suggests that spatial structure of tremors has finiteness and much of the tremor activity consists of fine structures. The characteristic spatial and temporal scales of the small structure are around 5 km and 1 hour, respectively. The contribution of background seismicity was less than 5% of all events, regardless of grid cell size. Additionally, the effective time for self-triggering is about an order of magnitude smaller than that for adjacent-triggering, which is consistent with previous studies suggesting that tremor propagation is controlled by diffusion processes (e.g., Ide et al., 2010; Ando et al., 2012).
We attempted anomaly detection based on the Kolmogorov-Smirnov test. In many areas, no anomalies were detected over the 20-year period, indicating that the multivariate Hawkes model well described the standard state of tremor activities. However, there were cases where periods of quiescence or increased activity were observed at times when rare SSE occurred in regions where SSEs hardly occurred throughout the entire 20-year period. The detected deviations from the model could potentially indicate changes in local slip distributions or stress states. Our methodology, with improved temporal resolution, may provide a quantitative framework for discussing the changes in the number and migration pattern of tremors before large earthquake (Nadeau & Guilhem, 2009; Shelly, 2009), and it could be valuable for detecting hidden SSEs and monitoring stress changes at plate boundaries through tremor activity analysis.
In this study, we first confirmed the importance of spatial information. We considered four factors contributing to the tremor occurrence: 1. steady background seismicity, 2. self-triggering effect acting in a narrow space, 3. adjacent-triggering effect from neighboring areas, and 4. long-term recurrence effect. As combinations of these, we constructed three types of temporal models (1 only, 1+2, 1+2+4), and two types of spatiotemporal models (1+2+3, 1+2+3+4). We applied and compared these five models to the tremor catalog in western Japan, which was constructed using the method of Yano and Ide (2024). The analysis period was from April 2004 to December 2023. When comparing the models using Akaike Information Criterion, the multivariate Hawkes model consisting of background seismicity, self-triggering, and adjacent-triggering (1+2+3) was found to be statistically superior. This model captured tremor migration patterns that move distances of about 5 to 10 km on time scales ranging from sub-hour to half a day; however, the direction of such short-term migration does not necessarily align with the direction of migration in longer time window.
We also investigated the contributions of each effect in the multivariate Hawkes process. The model revealed that when the grid cell size is 5 km in diameter, self-triggering accounts for an average of about 25%, reaching as high as 50%, followed by adjacent-triggering at around 11%. As the cell size increases, the contribution of self-triggering rises to an average of about 50%, with a maximum of approximately 80%, while adjacent-triggering decreases to about 7%. This cell-size dependency suggests that spatial structure of tremors has finiteness and much of the tremor activity consists of fine structures. The characteristic spatial and temporal scales of the small structure are around 5 km and 1 hour, respectively. The contribution of background seismicity was less than 5% of all events, regardless of grid cell size. Additionally, the effective time for self-triggering is about an order of magnitude smaller than that for adjacent-triggering, which is consistent with previous studies suggesting that tremor propagation is controlled by diffusion processes (e.g., Ide et al., 2010; Ando et al., 2012).
We attempted anomaly detection based on the Kolmogorov-Smirnov test. In many areas, no anomalies were detected over the 20-year period, indicating that the multivariate Hawkes model well described the standard state of tremor activities. However, there were cases where periods of quiescence or increased activity were observed at times when rare SSE occurred in regions where SSEs hardly occurred throughout the entire 20-year period. The detected deviations from the model could potentially indicate changes in local slip distributions or stress states. Our methodology, with improved temporal resolution, may provide a quantitative framework for discussing the changes in the number and migration pattern of tremors before large earthquake (Nadeau & Guilhem, 2009; Shelly, 2009), and it could be valuable for detecting hidden SSEs and monitoring stress changes at plate boundaries through tremor activity analysis.