Landslide susceptibility methodology for railway planning: a comparative analysis of statistical and machine learning methods in a case study of Marche region, Italy

Landslides pose serious risks to infrastructure, particularly railways, due to their rigid construction and essential transport role. Susceptibility mapping is a valuable tool during the feasibility phase of railway projects, helping identify high-risk areas and inform mitigation strategies. However, effective application requires both reliable classification of landslide types and robust reclassification methods for clear communication with stakeholders. This study presents a comprehensive workflow for landslide susceptibility mapping, combining Weight of Evidence (WoE) and a Generalized Additive Model with boosting (GAMB). We generated separate susceptibility maps for five landslide types and evaluated them using AUROC metrics. The maps were merged into an overall susceptibility map using a complementary probability approach, which also allowed assessment of each type's sensitivity to the overall susceptibility. To improve threshold reliability, we implemented an ensemble reclassification method using six approaches and applied the statistical mode to define more objective class boundaries. Visualizations of susceptibility along the railway route and its adjacent sides were developed for practical application. The methodology was applied to a 22 km planned railway section in the Marche region (Italy). Results revealed high spatial variability: rockfall types showed the highest accuracy (AUC = 0.94 WoE, 0.98 GAMB), while slides performed poorest. GAMB consistently outperformed WoE in reliability and smoothness of results. Finally, a comparison with EGMS ground motion data showed no significant correlation (R2 ≈ 0.1), underscoring the temporal disconnect between long-term susceptibility and short-term ground deformation.