Soil air radon as a seismic precursor: Disentangling meteorological and tectonic influences using machine learning

In this study, 78,721 soil air radon measurements (15-min intervals, December 2006-March 2008) were measured at a depth of ∼80 cm using an AlphaMETER-611 detector at the Gebze monitoring station, located near the North Anatolian Fault Zone (NAFZ)- Türkiye, and analyzed together with meteorological parameters and regional seismicity. Statistical methods, including correlation, cross-correlation, and hierarchical clustering, were first applied to distinguish environmental from seismic influences, revealing strong positive correlations with air and soil temperatures (r = 0.63-0.66, p < 0.001), moderate negative correlation with relative humidity (r = -0.40, p < 0.001), and weaker effects from precipitation (r = -0.18, p < 0.05). Subsequently, three machine learning models, Long Short-Term Memory (LSTM), Support Vector Regression (SVR), and Random Forest (RF), were applied to identify possible anomalous deviations. Several positive anomalies preceded shallow earthquakes (M_L4.8-5.7) by 22, 18, 15, 8, 7, and 1 days. Cross-model agreement highlights reproducible radon anomalies temporally associated with seismic activity. The LSTM model achieved the highest reconstruction performance (R² = 0.86), and its anomaly amplitudes showed an exponential association with earthquake magnitude (R² = 0.998) in post-event analysis.