12:00 PM - 12:15 PM
[STT43-06] Site selection based on Bayesian sensitivity
for earthquake early warning
Keywords:Earthquake early warning, Bayesian method
An Earthquake Early Warning (EEW) system detects seismic waves in the early stages of an earthquake and issues warnings before strong shaking reaches a specific location. In Japan, the Japan Meteorological Agency (JMA) provides EEW alerts by analyzing real-time seismic data collected from a nationwide network of seismic stations. To enhance the accuracy and reliability of EEW, various methods have been introduced, including the Integrated Particle Filter (IPF) method (Liu and Yamada, 2014; Tamaribuchi et al., 2014; Wu et al., 2014) and the Extended Integrated Particle Filter (IPFx; Yamada et al., 2021), which significantly improves source estimation and earthquake discrimination.
IPF and IPFx are Bayesian source estimation-based EEW methods that enhance the accuracy of earthquake source estimation and distinguish multiple simultaneous earthquakes. IPFx, an improved version of IPF, processes continuous waveforms and integrates all Japanese real-time seismic networks into a unified framework. To balance computational efficiency and accuracy, IPFx employs a seismic station selection strategy in source estimation: it selects 20 seismic stations near the first triggered station, 20 stations within the Voronoi cell containing the first triggered station, and 10 additional stations to ensure sufficient coverage.
This study aims to improve the site selection scheme in IPFx to achieve a better balance between computational efficiency and accuracy. We propose a selection method based on the Bayesian sensitivity formula (Perez et al., 2006; Millar and Stewart, 2007; Giordano, Broderick, and Jordan, 2018; Iba and Yano, 2022), which is closely related to the Widely-Applicable Information Criterion (WAIC; Watanabe, 2010). Our selection scheme seeks to maintain the posterior distribution even when observation sites are selectively chosen and assigns an easily computed importance value to each site, that is, the posterior variance of the likelihood at each site.
We conducted numerical experiments to assess the temporal variation and stability of observation site selection. By leveraging a time-varying likelihood function, we can track changes over time, using the same likelihood function as in the IPFx method. When the wave has not yet arrived at a site, its values are almost random. Observation sites where the wave has arrived gain greater importance, and the most important sites tend to be positioned to ensure better coverage. Further details will be reported.
IPF and IPFx are Bayesian source estimation-based EEW methods that enhance the accuracy of earthquake source estimation and distinguish multiple simultaneous earthquakes. IPFx, an improved version of IPF, processes continuous waveforms and integrates all Japanese real-time seismic networks into a unified framework. To balance computational efficiency and accuracy, IPFx employs a seismic station selection strategy in source estimation: it selects 20 seismic stations near the first triggered station, 20 stations within the Voronoi cell containing the first triggered station, and 10 additional stations to ensure sufficient coverage.
This study aims to improve the site selection scheme in IPFx to achieve a better balance between computational efficiency and accuracy. We propose a selection method based on the Bayesian sensitivity formula (Perez et al., 2006; Millar and Stewart, 2007; Giordano, Broderick, and Jordan, 2018; Iba and Yano, 2022), which is closely related to the Widely-Applicable Information Criterion (WAIC; Watanabe, 2010). Our selection scheme seeks to maintain the posterior distribution even when observation sites are selectively chosen and assigns an easily computed importance value to each site, that is, the posterior variance of the likelihood at each site.
We conducted numerical experiments to assess the temporal variation and stability of observation site selection. By leveraging a time-varying likelihood function, we can track changes over time, using the same likelihood function as in the IPFx method. When the wave has not yet arrived at a site, its values are almost random. Observation sites where the wave has arrived gain greater importance, and the most important sites tend to be positioned to ensure better coverage. Further details will be reported.