Speaker
Description
Variations in radon concentration in underground environments have long been investigated as potential indicators of geodynamic processes. Changes in tectonic stress may open new migration pathways for subsurface gases, leading to measurable anomalies in radon concentration. Such anomalies have been reported prior to several seismic events, suggesting that radon monitoring could contribute to the understanding of processes preceding earthquakes.
This study presents an analysis of long-term radon monitoring conducted in cave systems in Central Europe. Resulting datasets capture both short-term fluctuations and long-term seasonal variability of radon concentration. A key challenge in interpreting radon time series lies in separating anomalies potentially related to endogenous geophysical processes from variations caused by exogenous influences such as temperature, atmospheric pressure, humidity, etc.
In addition to classical time-series approaches, we explore several machine-learning techniques for anomaly detection and prediction. These include Empirical Mode Decomposition combined with Support Vector Regression (EMD–SVR), which decomposes the signal into intrinsic mode functions that are subsequently modeled using SVR to detect deviations from normal behavior, and neural-network approaches based on Long Short-Term Memory (LSTM) architectures. In particular, an LSTM auto-encoder is applied to learn typical temporal patterns in radon variability and detect statistically unusual deviations from these patterns.
Detected anomalies are subsequently compared with seismic catalogues to investigate possible temporal correlations between radon anomalies and seismic events, and potentially volcanic activity. The aim of this study is to evaluate the capability of machine-learning-based anomaly detection methods for identifying radon signals that may be associated with geodynamic processes.