Statistical modelling of rainfall-induced shallow landsliding using static predictors and numerical weather predictions: preliminary results

We present a quantitative indirect statistical modeling for predicting rainfall-induced shallow landsliding. We consider as input layers both static thematic predictors, such as geomorphological, geological, climatological information, and numerical weather model's forecast. Two different statistical techniques are used to combine together the above mentioned predictors: a Generalized Linear Model and Breiman's Random Forests. We tested these two techniques for two rainfall events that occurred in 2011 and 2013 in Tuscany region (central Italy). Model's evaluation is measured by means of sensitivity-specificity ROC analysis. In the 2011 rainfall event, the Random Forests technique performs slightly better, whereas in the 2013 rainfall event the Generalized Linear Model provides more accurate predictions. This study seeks also to establish whether the rainfall-induced shallow landsliding prediction might substantially benefit from the information provided by the numerical weather model's outputs. Using the variable importance parameter provided by the Random Forests algorithm, we asses the added value carried by numerical weather forecast, in particular in the rainfall event characterized by deep atmospheric convection and heavy precipitations.