Identifying Radon hazard areas: Machine learning-driven Geogenic Radon Potential mapping in Hessen

The health impacts of the radioactive Radon are well-documented by the World Health Organization (WHO) and numerous studies. Geogenic Radon Potential (GRP) refers to the natural production of Radon by the Earth, independent of anthropogenic influences. GRP has been a focal point of research aimed at understanding the factors influencing radon variability and its spatial distribution. However, the limited availability of systematic soil-gas radon concentration measurements, along with other constraints, often leads to coarse-resolution modeling of GRP. With the availability of adequate and quality data, regional studies can be promising in investigating these influencing factors, and modelling of GRP hazards at finer spatial scales.

 

This study uses GRP survey data provided by the Hessian Agency for Nature Conservation, Environment and Geology (HLNUG) to develop machine learning models for predicting the spatial distribution of GRP in the state of Hessen, Germany, and to produce a high-resolution GRP hazard map. The models employed include Random Forest Regressor (RF), Support Vector Regressor (SVR), Gradient Boosting Regressor (GBR), and Multi-Layer Perceptron Regressor (MLPR). The dataset comprises 1,509 GRP sampling points for an area of about 21.000 km², and 37 potential predictors related to geology, soil characteristics, and climatic variables—key factors known to influence radon levels. Sequential Feature Selection (SFS) and a 5-fold spatial cross-validation strategy were employed to mitigate autocorrelation effects and enhance model generalization. Model performance was evaluated using multiple metrics and compared against ground-truth values and local geology.

 

Results revealed that the RF and GBR models outperformed others, achieving R² scores of 0.69 and 0.65 on the validation dataset, respectively, while the SVR and MLPR models underperformed. Predicted GRP values ranged from 8.9 to 178.2 for RF and 1.7 to 268.4 for GBR. Geological and soil properties emerged as the dominant predictors of GRP variability in Hessen, with predicted maps highlighting a strong dependence on local geological features. High-risk areas were effectively identified by the RF model. The study also highlights the need for additional measurements in data-scarce regions and the exploration of hybrid physics-based models that integrate domain-specific knowledge into spatial predictions.