Phd in EARTH SCIENCES

Thesis title: Role of wildfires and snow covers as preparatory factors in multi-hazard shallow landslide scenarios through ML emulators

In landslide studies, preparatory factors are dynamic processes that progressively reduce slope stability without directly initiating failure. Although secondary to triggering and predisposing factors, they strongly influence the spatial and temporal probability of landsliding. Wildfires and snow covers are common preparatory processes that dynamically affect slope stability with respect to shallow landslides. From a multi-hazard perspective, the landslide scenario triggered by an earthquake or heavy rainfall can change in response to a wildfire that has recently affected the area or the presence of a snowpack above the ground, as a consequence of the preparatory effects of these processes. Despite their relevance, preparatory factors are not easy to consider in landslide zoning due to data limitations. While recent dynamic Machine Learning (ML) data-driven models could represent suitable frameworks, the absence of reliable landslide inventories prevents their application. To overcome such limitations, this study developed a new methodology based on ML emulators trained on physically-based simulations to study wildfires and snow covers as preparatory factors for shallow landslides. The simulations consisted of landslide scenarios prepared by wildfires and snow covers and triggered by earthquakes or rainfall, generated using two distinct modified versions of the established PARSIFAL model, to account for wildfire and snow preparation. The corresponding emulators, based on Random Forest, were designed to reproduce the deterministic outputs while enhancing computational efficiency, generalisability, and transferability through the use of proxy variables instead of harder-to-find geotechnical parameters. Two case studies were examined: Mt. Epomeo (Ischia Island, Italy) and the Lake Campotosto area (Central Apennines, Italy), for wildfire- and snow-prepared slopes, respectively. The results of both emulators and physically-based models used to train them highlight the different effects of the two preparatory factors, which are far from negligible. Wildfires reduce root reinforcement and consequently soil strength, increasing the likelihood of rainfall-induced shallow landslides for a time frame that generally lasts for some years. Snow covers produce lower effects, modifying stress conditions through loading and melting seasonal cycles, and increasing earthquake-induced instability. In both cases, emulators accurately reproduced physically-based landslide scenarios, resulting in computationally efficient and ready-to-use tools, potentially suitable for decision-making and real-time scenario analysis in multi-hazard contexts.