Identification of radon anomalies in soil gas using decision trees and neural networks

The time series of radon (222Rn) concentration in soil gas at a fault, together with the environmental parameters, have been analysed applying two machine learning techniques: (i) decision trees and (ii) neural networks, with the aim at identifying radon anomalies caused by seismic events and not simply ascribed to the effect of the environmental parameters. By applying neural networks, 10 radon anomalies were observed for 12 earthquakes, while with decision trees, the anomaly was found for every earthquake, but, undesirably, some anomalies appeared also during periods without earthquakes.