Item

Abstract

One of useful geochemical observations for evaluating the volcanic activity is measuring soil gas emission, which may indicate deep-sited changes in the volcanic system. At Mt. Azuma, Japan, a continuous observation station has been deployed since October 2017 to monitor soil gas fluxes. Mt. Azuma exhibited unrest in 2018, and the soil CO₂ flux increased drastically in August 2018 (e.g., Fukui et al., 2019). However, the CO₂ flux fluctuated largely over the entire observation period because of the perturbation of external environment parameters (such as temperature and humidity), which made it difficult to find smaller flux changes from a deep volcanic origin. In this study, we employed a machine learning approach to detect such small, unclear flux changes. First, we analyzed our CO₂ emission data with the multiple regression analysis (e.g., Morita et al., 2019). The model was trained using flux data from October 2017 to July 2018 (just before the significant flux increase) to represent typical background gas emission. Then we calculated flux residuals of the trained model in 2018. The residuals showed that the flux change occurred not only in August 2018 but also in July 2018, indicating that the model might detect the volcanic status change earlier than August 2018. We also applied a deep learning model with a long short-term memory (LSTM) layer to the same dataset and found that the LSTM model also supported the flux change in July. For further investigation, we will test other machine learning techniques and use seismic observation data.