[S-CG50] Driving Solid Earth Science through Machine Learning

[SCG50-P14] Preseismic anomaly detection of atmospheric radon concentration using Random Forest analysis

*Mayu Tsuchiya¹, Hiroyuki Nagahama¹, Jun Muto¹, Mitsuhiro Hirano¹, Yumi Yasuoka² (1.Department of Earth Sciences, Tohoku University, 2.Kobe Pharmaceutical University)

Keywords:Machine learning, Random Forest analysis, Radon concentration in the atmosphere, The 2011 Tohoku-oki earthquake, The 1995 Kobe earthquake

Currently, various anomalies occurring before earthquakes are being studied for earthquake prediction, and one of these is the radioactive element radon (²²²Rn). It is known that radon concentrations in soil, water, and the atmosphere fluctuate in response to crustal movements. In recent research, anomalies have been detected by statistically analyzing the fluctuations in radon concentrations before earthquakes, and quantitative evaluations of radon have been conducted. However, there was a problem with the method used to determine the parameters to be used in the analysis, so this study aimed for a more objective analysis. Radon concentrations in the atmosphere during the period when there were no major earthquakes were learned as normal period, and the appearance of variations and differences were compared with the data observed during preseismic period by Random Forest analysis. The standard deviation between the predicted and observed values was used to obtain a statistical threshold for an abnormality, and an abnormality was detected by assuming that an anomaly was detected when the value exceeded the threshold. At Kobe Pharmaceutical University (N34.733063, E135.283268), radon concentration data in the air from an exhaust monitor were used. Radon concentrations in the atmosphere were predicted for the period 1984-1988 as the normal period (teacher data) and for the period 1990-1995 as the preseismic period and compared with actual observation data. The results revealed that from the end of 1994, the observed values consistently exceeded the predicted values. Around 25 days before the 1995 Kobe earthquake, concentrations exceeding three times the standard deviation calculated from the difference between predicted and observed values were observed. This suggests an abnormal increase in atmospheric radon concentration before the occurrence of the 1995 Kobe earthquake. Additionally, during 1990 to 1991, a decrease in radon concentration was observed, believed to be due to a period of calm before the seismic activity. At Fukushima Medical University(N37.690326, E140.470872), ionizing current values observed by an exhaust monitor were used as an indicator corresponding to atmospheric radon concentration because of their linear relationship with radon concentration. Ionizing current values for preseismic period (2008 -2011) were predicted, using the years 2002-2007 as the normal period (teacher data). As a result, concentrations exceeding three times the standard deviation calculated from the difference between predicted and observed values were observed by the end of October 2010. Additionally, while radon concentrations typically exhibit seasonal variations with high peaks in winter, in 2010, the peak occurred earlier than usual, showing a high peak from autumn to winter. These anomalies suggest an abnormal increase in atmospheric radon before the 2011 Tohoku-oki earthquake. Furthermore, similar observations of concentrations exceeding three times the standard deviation were noted one week before the Iwate-Miyagi Inland earthquake in 2008, similar to the pattern observed before the 2011 Tohoku-oki earthquake. These results indicate the potential to identify abnormalities in atmospheric radon concentration before earthquake occurrences, suggesting the possibility of capturing earthquake precursors. In addition, in the previous anomaly detection of radon concentration in the atmosphere, analysis was conducted using data from which seasonal variations were removed. However, as shown in the results obtained at Fukushima Medical University in this study, anomalies could be detected at the same time as in the previous study using the observed ionizing current values as they were, contributing to simplification of the procedure required for anomaly detection.