Introducing the climate component into landslide susceptibility mapping

Intense rainfall and snowmelt are commonly recognized predisposing and triggering factors for shallow landslides, especially in mountainous environments. Due to climate change, their frequency and magnitude could vary, thus modifying soils response. Therefore, their inclusion in planning instruments becomes fundamental. The aim of our study was to derive a susceptibility model adaptable to climate changes, through the inclusion of variables summarizing intense rainfall and snowmelt processes. We selected the territory of the Mont-Emilius and Mont-Cervin Mountain Communities (northern Italy) as study area. To define the summary variables, we investigated the relationships between landslide occurrences and meteorological events (reference period 1991-2020). For landslide susceptibility mapping, we set up a Generalized Additive Model. For model training, we extracted from the local inventory 298 dated landslide points and we selected 300 random non-landslide points. We defined a reference model analyzing the statistiscal significance of the variables (relief, NDVI, land cover and geology predictors). Similarly, we optimized a model including the climate variables, checking also their smooth functions to ensure physical plausibility. Finally, we validated the optimized model through a k-fold cross-validation. We evaluated all models through contingency tables and area under the receiver operating characteristic curve (AUROC). Also, we investigated variable importance through the decrease in explained variance. The climate variables that resulted statistically and physically significant are the effective annual number of rainfall events with intensity-duration characteristics above a defined threshold (EATean) and the average number of melting events occurring in a hydrological year (MEn). In the optimized model, together they accounted for 5% of the model deviance. The optimized model showed a true positives rate and an AUROC higher than the reference model (2.4% and 0.8%, respectively). The introduction of the meteorological variables caused a transition of vulnerability class in 11.0% of the modelling area. The k-fold validation confirmed the statistical and physical significance of the meteorological variables in 74% (EATean) and 93% (MEn) of the fitted models. Our findings stress the utility of these variables in improving the performance of susceptibility models and making them adaptable to climate changes.