11:15 AM - 11:30 AM
[SCG53-03] Development of a real-time source model estimation method for volcanic regions
★Invited Papers
Keywords:real-time, REGARD, MCMC, Uncertainy
It is essential to understand crustal deformation caused by earthquakes and volcanic activities immediately and to model the sources from the viewpoint of understanding the phenomena and disaster prevention. Real-time GNSS is useful to detect crustal deformation immediately. GSI has developed and operated REGARD in cooperation with Tohoku University. In recent years, GSI has been using Markov Chain Monte Carlo methods (hereinafter "MCMC") to quantify the uncertainty of the estimation results due to observation and model errors. RUNE (Ohno et al., 2021), a method for immediately quantifying the uncertainty in estimation using MCMC, has been developed, and its performance is being evaluated in REGARD. While REGARD has been explicitly designed for coseismic events, real-time GNSS has yet to be fully utilized for events with a time band of less than one day, such as volcanic activities. In addition, no system can automatically determine the status of such events from continuous GNSS observation data in Japan. Improving such a system is an urgent issue for volcanic activity monitoring.
In this study, we have developed a program for the inversion of source models as a fundamental technology for real-time volcano monitoring systems using GNSS, including quantitative evaluation of estimation uncertainties. Following Ohno et al. (2021), we utilize MCMC in real-time by speeding up MCMC computation using parallel computation (computation time: about 15 sec), improving the efficiency of search using parallel tempering methods, and adjusting hyperparameters. The Metropolis Hasting method is used as the sampler of MCMC. Four models are supported: the Mogi model (Mogi, 1958), the Dike model (Okada, 1992), the Spheroid model (Cervelli, 2013), and the Rectangular fault model (Okada, 1992), and multiple models can be estimated simultaneously in any combination. The appropriate combination of these models is determined based on the AIC (Akaike, 1973). As an application of the model uncertainty to be quantified, we have also developed a method for creating a "station placement study map" to visualize the uncertainty of calculated displacement of the ground surface. In monitoring local volcanic activities, it is sometimes desirable to have a denser observation network than GEONET, which is set up at about 20 km. This map aims to provide an objective index for judging the expansion of observation points and the areas where observation points should be maintained in continuous activity monitoring.
To verify the performance of the developed method, we applied it to numerical experimental data and actual data from the 2015 Sakurajima expansion case and the 2023 Noto Peninsula earthquake (M6.5). The estimation results show the median values consistent with the correct values in the numerical experiments and previous studies and quantify the uncertainty of the model estimation in terms of the shape of the posterior probability distribution. In addition, we confirmed that the "station placement study map" can visualize the areas where additional observation points should be added to reduce the uncertainty of model estimation efficiently.
In addition to the inversion program developed in this study, it is necessary to incorporate observation data from other organizations and study the analysis flow for generating crustal movement data to construct a real-time volcano monitoring system. We will continue to develop the system based on the developed estimation method while collaborating with other organizations to utilize observation data.
In this study, we have developed a program for the inversion of source models as a fundamental technology for real-time volcano monitoring systems using GNSS, including quantitative evaluation of estimation uncertainties. Following Ohno et al. (2021), we utilize MCMC in real-time by speeding up MCMC computation using parallel computation (computation time: about 15 sec), improving the efficiency of search using parallel tempering methods, and adjusting hyperparameters. The Metropolis Hasting method is used as the sampler of MCMC. Four models are supported: the Mogi model (Mogi, 1958), the Dike model (Okada, 1992), the Spheroid model (Cervelli, 2013), and the Rectangular fault model (Okada, 1992), and multiple models can be estimated simultaneously in any combination. The appropriate combination of these models is determined based on the AIC (Akaike, 1973). As an application of the model uncertainty to be quantified, we have also developed a method for creating a "station placement study map" to visualize the uncertainty of calculated displacement of the ground surface. In monitoring local volcanic activities, it is sometimes desirable to have a denser observation network than GEONET, which is set up at about 20 km. This map aims to provide an objective index for judging the expansion of observation points and the areas where observation points should be maintained in continuous activity monitoring.
To verify the performance of the developed method, we applied it to numerical experimental data and actual data from the 2015 Sakurajima expansion case and the 2023 Noto Peninsula earthquake (M6.5). The estimation results show the median values consistent with the correct values in the numerical experiments and previous studies and quantify the uncertainty of the model estimation in terms of the shape of the posterior probability distribution. In addition, we confirmed that the "station placement study map" can visualize the areas where additional observation points should be added to reduce the uncertainty of model estimation efficiently.
In addition to the inversion program developed in this study, it is necessary to incorporate observation data from other organizations and study the analysis flow for generating crustal movement data to construct a real-time volcano monitoring system. We will continue to develop the system based on the developed estimation method while collaborating with other organizations to utilize observation data.