Landslide susceptibility mapping using a physically-based model: the May 2023 Emilia-Romagna meteorological event

Rainfall-induced shallow landslides are an increasingly frequent hazard, with a direct impact on the land and on the people who live on it. Physically-based models assess their susceptibility by coupling a mechanical stability analysis with a hydrological description of the slope. The difficulty lies in the choice of the input parameters. The geotechnical and hydrological properties these models require cannot be measured at catchment scale: they have to be assumed within plausible intervals and calibrated on the observed landslides. Two major rainfall events struck the Emilia-Romagna region of Italy in May 2023. They triggered thousands of shallow landslides in the Apennine sector, mostly debris slides and debris flows, on slopes where no instability had been recorded before. This thesis applies the Fast Shallow Landslide Assessment Model (FSLAM) to the municipal territory of Casola Valsenio (84 km², 100 to 960 m a.s.l.), calibrating and evaluating it on the inventory compiled immediately afterwards. The purpose is not the best performance the model can give, but a measure of what each modelling choice is worth. Six configurations were compared on a common basis, fixed in advance: the validation dataset, the search space of the parameters and the selection rule. They differ in the calibration strategy and in the structure of the model. Across the six, accuracy ranges from 0.653 to 0.665 and the area under the ROC curve from 0.703 to 0.740, while the parameter sets that produce them differ by orders of magnitude. What the choices do change is the operating point of the model, the plausibility of the calibrated parameters and the consistency of the state of the slopes before the event. Accuracy alone therefore cannot select the parameters. Physical plausibility has to be set as a criterion before the calibration, not read off the results afterwards. What limits the model here is not its formulation, but the spatial information it can draw on.