Abstract. Our study is aimed at estimating the added value provided by Numerical Weather Prediction (NWP) data for the modelling and prediction of rainfall-induced shallow landslides. We implemented a quantitative indirect statistical modelling of such phenomena by using, as input predictors, both geomorphological, geological, climatological information and numerical data obtained by running a limited-area weather model. Two standard statistical techniques are used to combine the predictor variables: a generalized linear model and Breiman's random forests. We tested these models for two rainfall events that occurred in 2011 and 2013 in Tuscany region (central Italy). Modelling results are compared with field data and the forecasting skill is evaluated by mean of sensitivity–specificity receiver operating characteristic (ROC) analysis. In the 2011 rainfall event, the random forests technique performs slightly better than generalized linear model with area under the ROC curve (AUC) values around 0.91 vs. 0.84. In the 2013 rainfall event, both models provide AUC values around 0.7.

Using the variable importance output provided by the random forests algorithm, we assess the added value carried by numerical weather forecast. The main results are as follows: (i) for the rainfall event that occurred in 2011 most of the NWP data, and in particular hourly rainfall intensities, are classified as "important" and (ii) for the rainfall event that occurred in 2013 only NWP soil moisture data in the first centimetres below ground is found to be relevant for landslide assessment. In the discussions we argue how these results are connected to the type of precipitation observed in the two events.