PSI - Issue 84

Francesca Ceccato et al. / Procedia Structural Integrity 84 (2026) 599–606

604

primarily associated with hilly terrain and transitional zones between mountainous and plain areas. This distribution reflects the gradual and progressive nature of slow-moving landslides, which are often controlled by soil properties, hydrological conditions, and long-term slope deformation rather than extreme topographic gradients. Feature importance derived from the RF models provides insight into the dominant conditioning factors controlling different landslide processes (Figure 3a). For rapid landslides, terrain-related variables dominate the feature importance ranking. Slope and Gaussian curvature emerge as one of the most influential factors, followed by soil density, lithology, and clay content. This indicates that rapid landslides are strongly controlled by local terrain steepness, slope geometry, and material strength. Precipitation also shows relatively high importance, underscoring its role as a key triggering factor for fast-moving failures. In contrast, slow landslides exhibit a more balanced contribution from geomorphological, soil, and hydrological factors. While slope, soil density and HAND remain important, variables such as clay content, sand content, lithology, and precipitation show comparatively higher relative influence than in the rapid landslide model. Notably, vegetation-related indices (e.g., NDVI) and land use/land cover (LULC) display moderate importance for both landslide types, suggesting an indirect but non-negligible influence through root reinforcement, surface runoff regulation, and anthropogenic modification of slopes. The predictive performance of the RF models was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC–ROC) (Figure 3b). The rapid landslide model achieved an AUC of 0.93, while the slow landslide model achieved an AUC of 0.96, indicating excellent and very good discriminatory power, respectively, both models demonstrate strong predictive capability and robustness, confirming the suitability of the RF approach for large-scale landslide susceptibility assessment using multi-source geospatial data.

Fig. 4. Map showing bridge risk under slow-moving hazard conditions, with a regional overview on the right and a zoomed-in cluster view on the left. Bridges are classified into five risk levels from low to very high, highlighting localized hazard for the bridges within the study area.

Table 2. Percent (%) of bridges fall in different susceptibility classes for Rapid and Slow Landslide. Susceptibility Class Percent (%) of bridges fall in different susceptibility class Rapid Slow Low 70.23 83.72 Moderate 18.14 9.30 Moderate-High 8.84 4.65 High 1.40 1.40 Very High 1.40 0.93