Enhancing radon-deficit technique efficacy: machine learning applications for environmental variable analysis in soil gas monitoring

Soil contamination remains a critical environmental concern, necessitating efficient techniques for site characterization and remediation. The radon-deficit technique (RDT) offers a non-invasive approach to identifying organic contamination, relying on the behavior of radon-222 (222Rn) as a tracer. However, RDT results are influenced by environmental variables such as soil moisture, temperature, and atmospheric pressure, potentially leading to uncertainties. This study evaluates the application of machine learning (ML) models—including linear regression (LR), random forest (RF), artificial neural network (ANN), and gradient boosting machine (GBM)—to predict 222Rn activity in soil gas based on environmental parameters. A year-long dataset of continuous measurements was collected from an uncontaminated granite-based site in Madrid, encompassing variables such as soil moisture, ambient and soil temperatures, and atmospheric conditions. ANN and RF models exhibited superior performance in predicting 222Rn variability, identifying soil moisture and ambient temperature as the most influential predictors. The findings demonstrate that ML can significantly enhance the reliability of RDT by accounting for environmental variability, enabling more accurate identification of contamination hotspots. While the application of these models requires substantial datasets, they offer a promising tool for improving the efficacy of contamination screening and long-term remediation monitoring. Further studies are recommended to explore ML’s predictive capacity in contaminated sites and expand the approach to diverse geological contexts.