| 03 Dec 2025
Rapid Landslide Mapping During the 2023 Emilia-Romagna Disaster: Assessing Automated Approaches with Limited Training Data
Abstract. The catastrophic rainfall events of May 2023 in the Emilia-Romagna region, Italy triggered more than 80.000 landslides (https://doi.org/10.5281/zenodo.13742643, Pizziolo et al., 2024) and placed extraordinary demands on emergency response systems. One of the most critical tasks during the emergency was mapping the landslides caused by the event, a process carried out manually that required substantial time and effort (Berti et al., 2024).This study explores the potential of automated landslide mapping to support rapid disaster response, evaluating two deep learning models, U-Net and SegFormer, under realistic emergency constraints, including minimal training data. The models were trained exclusively in one affected municipality area (Casola Valsenio, 84 km2) and tested on three additional municipalities (Predappio, 91 km2; Modigliana, 101 km2; Brisighella, 194 km2) with varying geological settings. To reflect a range of operational scenarios, we tested seven combinations of input layers, progressively increasing in complexity from post-event Sentinel-2 imagery alone to full integration of high-resolution aerial imagery, NDVI change maps, and slope data.
Results show that both models achieved comparable segmentation performance, with SegFormer displaying greater robustness to variations in input layers and geological conditions, while U-Net was more sensitive but occasionally more accurate with rich inputs. Both models successfully identified landslides, but encountered difficulties in shadowed zones, cultivated fields, and geologically distinct terrains. A key limitation emerged in Brisighella FAA, a sector dominated by Blue Clay formations, where poor generalization was traced to the lack of lithological diversity in the training data, highlighting the need for geologically balanced training datasets.
Despite these challenges, the study confirms the operational value of automated mapping as a first-pass tool. Both models delivered accurate baseline maps that could support manual validation and prioritization in time-critical scenarios. The findings suggest that AI-based mapping could become a key component of emergency management protocols. While manual revision remains essential, these tools offer a scalable, time-efficient solution to the increasing demand for rapid and spatially detailed hazard information.
The study employed deep learning techniques based on convolutional neural network architectures (U-Net and SegFormer) to segment landslides across regions with distinct geological characteristics. The primary objective was to assess the effectiveness and limitations of automated landslide mapping in practical scenarios and to evaluate whether deep-learning approaches can reliably replace manual mapping. The results highlight several important aspects, including a comparison of models trained using different architectures and input layers. The discussion further examines the impact of incorporating a lithological layer into the model and presents an analysis of buildings potentially affected by landslide hazards. Despite the interesting approach to landslide segmentation, the manuscript presents structural and organizational weaknesses that at times hinder readability. A revision is therefore recommended to improve the clarity of the methodological framework and the presentation of results.