Rapid landslide mapping during the 2023 Emilia-Romagna disaster: assessing automated approaches with limited training data

The catastrophic rainfall events of May 2023 in the Emilia-Romagna region, Italy, triggered more than 80 000 landslides, as documented in a publicly available inventory (https://doi.org/10.5281/zenodo.13742643, Pizziolo et al., 2024), and placed extraordinary demands on emergency response systems. One of the most critical emergency tasks was landslide mapping, which was carried out manually and required substantial time and effort. This study investigates the potential of automated landslide mapping to support rapid disaster response by evaluating two deep learning models, U-Net and SegFormer, under realistic emergency constraints, including limited training data. The models were trained using data from one affected municipality, Casola Valsenio (84 km 2 ), and tested on three additional municipalities, Predappio (91 km 2 ), Modigliana (101 km 2 ), and Brisighella (194 km 2 ), characterized by different geological settings. To represent a range of operational scenarios, we tested seven combinations of input data, from post-event Sentinel-2 imagery alone to the integration of high-resolution aerial imagery, vegetation change maps, and slope data. Both models achieved comparable segmentation performance, with SegFormer showing greater robustness to variations in input data and geological conditions, while U-Net was more sensitive but occasionally more accurate when richer inputs were available. Both models successfully identified landslides, but showed limitations in shadowed areas, cultivated fields, and geologically distinct terrains. A major limitation emerged in the Brisighella area, where poor generalization was associated with the dominance of Blue Clay formations and the limited lithological diversity of the training data. These results highlight the importance of geologically balanced datasets for improving model transferability. Overall, the study confirms the operational value of automated landslide mapping as a first-pass tool for emergency response. Although manual revision remains necessary, both models produced reliable baseline maps that can support validation and help prioritize interventions in time-critical situations, offering a scalable and time-efficient approach to the growing need for rapid and spatially detailed hazard information.