LANDSLIDE SUSCEPTIBILITY MAP BASED ON MACHINE LEARNING: A VALIDATION BASED ON THE HEAVY RAINFALL EVENT OF MAY 2023 IN EMILIA ROMAGNA, ITALY

This study develops and validates landslide susceptibility models using machine learning techniques, specifically the Random Forest (RF) algorithm implemented on the Google Earth Engine (GEE) platform. The models are applied at both national and regional scales in Italy, targeting two landslide categories: "All" (All types) and " Rainfall-induced" (Rapid). A national-level susceptibility map is generated using input data from the all country, while the regional-level map is obtained using only a subset of the data. The case of the extreme rainfall event in May 2023 in the Emilia Romagna region of Italy is examined to test the performance of the models. In this event, more than 65,000 landslides occurred across 72 km2. The results are critically analysed to understand how scale and data volume affect accuracy. The national models provide a broad overview but may miss localized variations captured by the provincial models. The provincial "All" model excels, with 83.45% of landslides in the highest susceptibility categories, benefiting from larger training data. However, the provincial "Rapid" model underperforms due to limited training data. The study highlights the importance of scalability and data quality. National models offer a useful starting point, while provincial refinement with localized data can significantly enhance predictions. A multi-scale approach combining national coverage with provincial specificity is recommended for future analyses to improve landslide risk assessment and management strategies.