Both the sustainable use of our resources and the prevention of geohazards requires reliable information about the spatial distribution of soil and geological properties. Since direct measurements are costly, artificial intelligence (AI) methods are used to estimate these attributes, leveraging a machine learning algorithm which relates laboratory measurements or expert class information to environmental covariates derived, e.g. from relief, geology and climate data. This study evaluates random forest (RF) as an AI technique to predict the occurrence of debris on slopes of the entire Black Forest in 10 m resolution. It also examines whether RF models can be applied to measured geogenic radon potential (GRP) for assessing the risk of possibly harmful radon concentrations inside buildings in Baden-Wuerttemberg.
A suite of 6770 expert class labels indicating whether hillside debris of at least 1 m thickness occurs or not, were associated with main geological classes and various terrain attributes obtained from a LiDAR-based digital elevation model. RF classification showed very good results with an accuracy rate of 86 %. GRP mapping is currently based on 580 radon measurements in soil gas at 1 m depth and a set of covariates comprising soil attributes, climate variables and geological data such as uranium concentrations. Preliminary results indicate that the GRP map generated by using a state-specific RF model is highly useful in identifying municipalities as vulnerable areas for which action is needed to mitigate this particular threat to human health.