Radon hazard mapping: usability of environmental predictors including atmospheric radon and radon flux and knowledge transfer between regions (Belgium and Germany)

Abstract

Radon is a naturally occurring radioactive noble gas. When it accumulates indoors it can be a health hazard. Radon hazard mapping assigns areas to a geogenic radon potential, that reflects the availability and spatial distribution of radon in soil. The possible knowledge transfer from one region to another and the usability of predictors for radon hazard mapping were analysed. Included in the set of predictors were “atmospheric radon” and “radon flux”. A machine learning workflow is outlined using a random forest model to predict the geogenic radon potential in Belgium and Germany. The German data was used as training data and the model performance was evaluated on spatially separated validation data sets in both regions. It was possible to predict the geogenic radon potential for Belgium only using training data from Germany. The evaluation of the model performance on the Belgian validation data set was essential to find this model. The model showing the highest model performance in Belgium differs in main characteristics as number, selection and importance of predictors from the predictive model working best in Germany. The predictions of the geogenic radon potential of these models were accurately in their country but not in the other. The models used different predictors, except the predictor “soil moisture”, which was present in both models. The performance increase for single predictors in Germany is in the range of a few percent, whereas in Belgium a single predictor (“coarse fragments”) can improve the model by over 100%. Among the 30 candidate predictors “radon flux” was present in the best model for Belgium.