[S-CG60] Driving Solid Earth Science through Machine Learning

[SCG60-04] Trend Extraction from Time Series Data Using the State Space Model and Its Application to Anomaly Detection

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

Keywords:Machine learning, Random Forest analysis, State Space Model, Radon concentration in the atmosphere, The 2011 Tohoku-oki earthquake, Anomaly detection

Currently, various anomalies that occur before earthquakes are being studied to predict seismic events, and one such anomaly is the radioactive element radon (²²²Rn). It is known that radon concentrations in soil, water, and the atmosphere fluctuate in response to crustal deformation. Recent studies have statistically analyzed fluctuations in radon concentrations before earthquakes to detect anomalies and quantitatively evaluate radon behavior. Among these studies, atmospheric radon concentration measurements widely utilize a device called AlphaGUARD. AlphaGUARD has the advantage of being a compact and inexpensive instrument, eliminating the need for large-scale observation facilities and allowing measurements in various locations. However, due to its small size, the device has a limited internal air volume, making it prone to overestimation or underestimation of concentrations compared to large gas flow ionization chambers. This results in significant variability in measurement data. To address this issue, we applied a State Space Model that extracts latent components from time-series data to atmospheric radon concentration measurements from AlphaGUARD and extracted fluctuation trends. The extracted trend exhibited long-period fluctuations compared to the highly variable, short-wavelength fluctuations observed in AlphaGUARD raw data. Furthermore, the extracted trend was similar to the fluctuations in atmospheric radon concentrations measured simultaneously at the same location using a large gas flow ionization chamber. This method was also applied to atmospheric radon concentration measurements taken by AlphaGUARD at the Kozumi station of the Miyagi Prefectural Nuclear Center, located on the Oshika Peninsula, Miyagi Prefecture. As a result, we successfully extracted trends from the atmospheric radon concentration data observed at Kozumi station using the State Space Model. The timing of the extracted trend increases coincided with the anomaly detection timing found in a previous study that analyzed atmospheric radon concentrations measured with a large gas flow ionization chamber at Fukushima Medical University. Additionally, we applied Random Forest analysis—a type of machine learning—to detect anomalies using the extracted trend. By calculating the difference between the extracted trend and the predicted values obtained through Random Forest, we found that atmospheric radon concentrations exceeding three times the standard deviation of the difference were observed before the Tohoku-oki earthquake. Furthermore, when comparing the extracted trend and prediction residuals with GPS data from the Oshika Peninsula, significant GPS displacements were observed near the timing when the sign of the residuals changed. This suggests that the atmospheric radon concentration fluctuations captured by our approach were associated with crustal deformation. We also conducted Random Forest analysis on AlphaGUARD raw data. While it successfully detected pre-seismic anomalies, frequent anomalies were also observed outside the seismic periods. This indicates the necessity of establishing new criteria for defining anomalies when using AlphaGUARD raw data for anomaly detection. These findings suggest that the State Space Model is an effective approach for facilitating the interpretation of highly variable data when conducting atmospheric radon concentration anomaly detection using conventional methods.