Michael Maerker, Samanta Pelacani, Boris Schroeder

• In this paper we evaluate different methods to predict soil erosion processes. We derived different layers of predictor variables for the study area in the Northern Chianti, Italy, describing the soil-lithologic complex, land use, and topographic characteristics. For a subcatchment of the Orme River, we mapped erosion processes by interpreting aerial photographs and field observations. These were classified as erosional response units (ERU), i.e. spatial areas of homogeneous erosion processes. The ERU were used as the response variable in the soil erosion modelling process. We applied two models i) bootstrap aggregation (Random Forest: RF), and ii) stochastic gradient boosting (TreeNet: TN) to predict the potential spatial distribution of erosion processes for the entire Orme River catchment. The models are statistically evaluated using training data and a set of performance parameters such as the area under the receiver operating characteristic curve (AUC), Cohen's Kappa, and pseudo R2. Variable importance and response curves provideIn this paper we evaluate different methods to predict soil erosion processes. We derived different layers of predictor variables for the study area in the Northern Chianti, Italy, describing the soil-lithologic complex, land use, and topographic characteristics. For a subcatchment of the Orme River, we mapped erosion processes by interpreting aerial photographs and field observations. These were classified as erosional response units (ERU), i.e. spatial areas of homogeneous erosion processes. The ERU were used as the response variable in the soil erosion modelling process. We applied two models i) bootstrap aggregation (Random Forest: RF), and ii) stochastic gradient boosting (TreeNet: TN) to predict the potential spatial distribution of erosion processes for the entire Orme River catchment. The models are statistically evaluated using training data and a set of performance parameters such as the area under the receiver operating characteristic curve (AUC), Cohen's Kappa, and pseudo R2. Variable importance and response curves provide further insight into controlling factors of erosion. Both models provided good performance in terms of classification and calibration; however, TN outperformed RF. Similar classes such as active and inactive landslides can be discriminated and well interpreted by considering response curves and relative variable importance. The spatial distribution of the predicted erosion susceptibilities generally follows topographic constraints and is similar for both models. Hence, the model-based delineation of ERU on the basis of soil and terrain information is a valuable tool in geomorphology; it provides insights into factors controlling erosion processes and may allow the extrapolation and prediction of erosion processes in unsurveyed areas.…