15:30 〜 15:45
[SCG45-06] Comparison and improvement of statistical models for activity of low-frequency earthquakes
キーワード：スロー地震、低周波地震、統計モデル、南海トラフ
Slow earthquakes are a general term for fault slip events that have much longer durations than regular fast earthquakes of comparable seismic moments (Ide et al., 2007). They have been actively studied for more than two decades. Previous studies have revealed detailed spatiotemporal distributions of slow earthquakes in several subduction zones (Beroza & Ide, 2011; Obara & Kato, 2016; Nishikawa et al., 2023). However, statistical or physical models that enable forecasts of slow earthquake activity have not yet been fully developed.
Attempts to construct statistical models for low-frequency earthquakes (LFEs), a type of slow earthquake, have just begun. Lengliné et al. (2017) and Tan & Marsan (2020) modeled LFE activity as random occurrences at a constant rate and LFE-to-LFE triggering. This model is very similar to the epidemic-type aftershock-sequence (ETAS) model (Ogata, 1988), which is a standard statistical model of regular seismicity. In contrast, Ide & Nomura (2022) expressed interevent times of tectonic tremors (i.e., swarms of LFEs) as a mixture distribution of a log-normal distribution and a Brownian passage time distribution. This approach resembles long-term forecasts of large-earthquake recurrence on active faults (Earthquake Research Committee, 2022). Hereafter, I refer to this model as the IN model.
I compared the above statistical models and evaluated their performance using the Akaike information criterion. I applied the ETAS model and IN model to LFEs that occurred below Shikoku Island in the Nankai Trough from 2004 to 2015. I did not use the models proposed by Lengliné et al. (2017) and Tan & Marsan (2020) because these models have far more model parameters (over a hundred) than the ETAS and IN models and therefore are not suitable for the direct comparison. Both the ETAS and IN models have only five parameters. I used the LFE catalog created by Kato & Nakagawa (2020). I determined the LFE minimum magnitude for my analyses using the Goodness-of-Fit method (Wiemer & Wyss, 2000) and set it to M 0.4.
The results show that the LFE activity has an Omori-Utsu’s p-value (1.3 or larger) significantly higher than regular seismicity (usually 1.0) and an almost zero alpha-value, suggesting that the LFE activity decays very rapidly and has no magnitude dependence on its clustering. The high p-value is consistent with the results of the previous studies ( Lengliné et al., 2017; Tan & Marsan, 2020). Furthermore, the IN model generally outperforms the ETAS model, implying that considering not only the short-term clustering of LFEs but also the long-term recurrence of LFE clusters is important to better forecast the LFE activity.
However, there is still room for improvement in both models. I am now improving the models by modeling the interevent times of the LFEs using more complex distributions and by considering the influences of nearby slow slip events.
Attempts to construct statistical models for low-frequency earthquakes (LFEs), a type of slow earthquake, have just begun. Lengliné et al. (2017) and Tan & Marsan (2020) modeled LFE activity as random occurrences at a constant rate and LFE-to-LFE triggering. This model is very similar to the epidemic-type aftershock-sequence (ETAS) model (Ogata, 1988), which is a standard statistical model of regular seismicity. In contrast, Ide & Nomura (2022) expressed interevent times of tectonic tremors (i.e., swarms of LFEs) as a mixture distribution of a log-normal distribution and a Brownian passage time distribution. This approach resembles long-term forecasts of large-earthquake recurrence on active faults (Earthquake Research Committee, 2022). Hereafter, I refer to this model as the IN model.
I compared the above statistical models and evaluated their performance using the Akaike information criterion. I applied the ETAS model and IN model to LFEs that occurred below Shikoku Island in the Nankai Trough from 2004 to 2015. I did not use the models proposed by Lengliné et al. (2017) and Tan & Marsan (2020) because these models have far more model parameters (over a hundred) than the ETAS and IN models and therefore are not suitable for the direct comparison. Both the ETAS and IN models have only five parameters. I used the LFE catalog created by Kato & Nakagawa (2020). I determined the LFE minimum magnitude for my analyses using the Goodness-of-Fit method (Wiemer & Wyss, 2000) and set it to M 0.4.
The results show that the LFE activity has an Omori-Utsu’s p-value (1.3 or larger) significantly higher than regular seismicity (usually 1.0) and an almost zero alpha-value, suggesting that the LFE activity decays very rapidly and has no magnitude dependence on its clustering. The high p-value is consistent with the results of the previous studies ( Lengliné et al., 2017; Tan & Marsan, 2020). Furthermore, the IN model generally outperforms the ETAS model, implying that considering not only the short-term clustering of LFEs but also the long-term recurrence of LFE clusters is important to better forecast the LFE activity.
However, there is still room for improvement in both models. I am now improving the models by modeling the interevent times of the LFEs using more complex distributions and by considering the influences of nearby slow slip events.