Flood Susceptibility Mapping with GFI 2.0 and Artificial Intelligence Models

Floods are among the most damaging natural hazards, motivating the development of rapid and scalable tools for floodplain mapping across multiple return periods and for post-event assessment. The Geomorphic Flood Index (GFI) is widely used to identify flood-prone areas using topographic information, but it can exhibit reduced reliability under complex hydraulic conditions—particularly near confluences where backwater controls water levels—and it may systematically overestimate inundation extents when used as a binary classifier.

This study advances the GFI framework by explicitly accounting for backwater effects at river confluences and along tributary junctions. In parallel, to reduce the intrinsic overestimation of GFI-derived floodplains, we test a suite of Artificial Intelligence (AI) classifiers—Random Forest, XGBoost, and Neural Networks—trained through a multi-parametric formulation that combines GFI with auxiliary predictors, including precipitation, lithology, land use, and slope. The approach is evaluated across multiple Italian catchments, using satellite-derived inundation and hydrodynamic simulations as independent benchmarks. Model performance is quantified against the baseline GFI approach using a standard threshold-based binary classification using an optimal cutoff.

The proposed framework aims to improve post-event flood delineation under observational constraints (e.g., satellite data gaps due to cloud cover, vegetation, or imaging limitations) and to provide a computationally efficient surrogate for extending hydrodynamic information to additional return periods or large basins where full numerical modelling is impractical. Preliminary results indicate that Random Forest provides the most robust performance across study sites. Incorporating backwater effects yields clear gains at confluences, primarily by reducing omission errors and improving the representation of hydraulically controlled inundation patterns. Moreover, the AI-based correction substantially mitigates the overestimation typically associated with standard GFI mapping, resulting in floodplain delineations that are more consistent with complex hydrodynamic processes and suitable for scalable flood hazard applications.