Multiscale landslide susceptibility analysis and risk assessment adaptable to linear infrastructure design using innovative methodologies

Landslides in complex geological settings are widespread hazards that, under climate change, increasingly threaten infrastructure and communities. Linear infrastructure such as railways is particularly vulnerable because of rigid construction and its critical transport function; susceptibility mapping can therefore support route planning and mitigation. This PhD project, developed with ITALFERR s.r.l., improves susceptibility workflows by separating landslide types and producing decision-ready outputs. First, along the Fabriano–Castel Planio railway (Marche, Italy), five type-specific susceptibility maps were produced and evaluated with AUROC. They were combined via complementary probability to obtain an overall map and quantify each type’s contribution. Susceptibility classes were defined with an ensemble reclassification (six methods) and class boundaries set by the statistical mode. Weight of Evidence (WoE) and a boosted Generalized Additive Model (GAMBoost) were compared. InSAR ground motion showed weak correspondence, indicating a temporal mismatch between long-term susceptibility and short-term deformation. Second, the May 2023 Emilia-Romagna extreme rainfall sequence (two events 14 days apart; combined return period >500 years) triggered thousands of landslides and required rapid mapping. A multi-institutional inventory of ~80,000 polygons from high-resolution aerial imagery provided ground truth for landslide-type-specific probability modelling. Rainfall was treated as a continuous time series and embedded in a Transformer Neural Network coupled with a Dense Neural Network for static predictors. Explainability was ensured via SHAP-based Expected Gradients, quantifying the temporal and spatial influence of rainfall. Finally, a transferable, low-cost desk-study framework is proposed to screen debris-flow hazard where inventories are scarce, combining morphometrics, empirical volume estimates, and exposure/vulnerability data, demonstrated in Central Val Camonica (Italian Alps). Overall, the thesis advances susceptibility analysis by integrating landslide typing, transparent classification, and interpretable spatiotemporal modelling.