Abstract
Rainfall-induced landslides pose significant challenges in Italy due to their small scale, fragmented distribution, and the severe damage they cause to populations, structures, and infrastructures. These challenges are amplified by the increasing frequency of intense precipitation events and the complex geomorphological settings of the Italian landscape. Therefore, the identification of such phenomena remains of critical importance. Despite extensive research on semi- and automatic landslide detection, many regional landscapes remain underexplored with advanced deep learning models. This study employs high-resolution (3 m) RGB Planet imagery to evaluate the performance of UNet and SegFormer for small landslide segmentation in these scarcely investigated regions. Unlike UNet's reliance on localized convolutional operations, SegFormer leverages a transformer-based architecture with attention mechanisms, enabling enhanced capture of both local and global features crucial for detecting small and spatially scattered landslides. Experimental results on Emilia-Romagna (Central Italy, affected by two major thunderstorms in May 2023) show that the optimal SegFormer configuration (patch size 64, batch size 16) achieves an F1-score of 73.42%, Recall of 71.72%, and mIoU of 63.87%, outperforming UNet in segmentation precision and reliability. Further cross-region testing on Tuscany (Central Italy, affected by abundant rainfall in March 2025) confirms SegFormer's potential for robust generalization, achieving an F1-score of 50.01%, Recall of 51.77%, and mIoU of 66.35%. These findings underscore SegFormer's potential for operational-scale mapping of rainfall-induced small landslides and the pressing need for automated mapping solutions to support disaster preparedness in precipitation-prone contexts.