Mapping flood susceptibility using Random Forest exploiting satellite observations and geomorphic features.

Flood events are among the most destructive natural hazards, requiring comprehensive risk management strategies to mitigate their impact on society and the environment. This study uses the potential of the Random Forest (RF) model to assess the flood susceptibility in Italy, evaluating 26 potential flood conditioning factors (FCF). A holistic strategy called Average Merit of Information (AMI) was employed to maximize the information contained within FCFs. At the same time, correlation issues were addressed using the Pearson correlation index and the Variance Inflation Factor (VIF). Satellite observations and regional records of historical flood events were adopted to calibrate the model and represent the maximum flood extension. Eleven sets of factors (SoF) were evaluated using a validation set and compared with official flood hazard maps. The RF model trained with SoF-1 (mean maximum daily precipitation (MMDP), the Geomorphic Flood Index (GFI), distance from the nearest river (DNR), elevation, lithology, soil properties, Normalized Difference Vegetation Index (NDVI), and land cover) demonstrated superior generalisation capacity compared to other SoFs. The inclusion of GFI significantly improved prediction accuracy in most unexplored areas, though challenges persist in flat regions and some areas without information. Ultimately, integrating updated satellite-derived information, complementary datasets, and adequate predictors facilitates the accurate identification of flood-prone areas, streamlining computational processes and providing decision-makers with preliminary analysis.