Predicting the thickness of shallow landslides in Switzerland using machine learning

Abstract. Landslide thickness is a key parameter in various types of models used to simulate landslide susceptibility. In this study, we developed a model providing improved predictions of potential shallow landslide thickness in Switzerland. We tested three machine learning models based on random forests, generalized additive model, and linear regression and compared the results to three existing models that link soil thickness to slope and elevation. The models were calibrated using data from two field inventories in Switzerland ("HMDB" with 709 records and "KtBE" with 517 records). We explored 37 different covariates including metrics on terrain, geomorphology, vegetation height, and lithology at three different cell sizes. To train the machine learning models, 21 variables were chosen based on the variable importance derived from random forest models and expert judgement. Our results show that the machine learning models consistently outperformed the existing models by reducing the mean absolute error by at least 17 %. The random forests models produced a mean absolute error of 0.25 m for the HMDB and 0.19 m for the KtBE data. Models based on machine learning substantially improve the prediction of landslide thickness, offering refined input for enhancing the performance of slope stability simulations.