FloodGuard: An AI-Powered Tool for Flood Risk and Vulnerability Mapping in Ungauged Basins.

Floods are among the most impactful natural phenomena affecting society. The associated risks are influenced by factors such as urban expansion into high-risk areas, modifications to rivers and watersheds (e.g., artificial channels, flow redirection, and structures like dams and dikes), and the effects of climate change, including more frequent and severe events. In recent years, the application of Artificial Intelligence (AI) in climate and weather risk assessment has gained increased attention due to its ability to handle numerical and categorical variables, uncover nonlinear relationships, and achieve high performance.
In this study, we introduce FloodGuard, an AI-powered tool for flood vulnerability and risk mapping. FloodGuard employs the concept of regionalization in ungauged basins and leverages a flood inventory derived from satellite imagery (e.g., Copernicus Emergency Mapping Service) over extensive areas (e.g., national or continental scales). The methodology selects the most relevant historical flood events and transfers this information to train a Random Forest machine learning model for estimating flood extent and producing a flood exposure map. Inputs to the model include the Geomorphic Flood Index (GFI), the Elevation, the Horizontal Distance to the Nearest River, Precipitation, the NDVI, and information on Land Use and Lithology. Flood prediction map is evaluated using maps generated from hydrological and hydraulic models. To assess vulnerability, we apply a geomorphic approach proposed by Manfreda and Samela (2019). This approach estimates flood depth, which is useful for estimating fast vulnerability levels. Finally flood risk is estimates with a GIS-based model.
The primary objective of this study is to provide a preliminary simple tool to estimate a flood risk and provide risk maps. At the same time, this study evaluates evaluate the transferability of machine learning models from regions with flood records to ungauged areas using satellite observations. Limitations include uncertainties inherent to machine learning models and the lack of association with specific return periods. Preliminary results across Italy demonstrate that the Random Forest model achieves high performance (AUC > 0.9) and exhibits robust generalization capabilities (e.g., combined error (rfp + (1-rtp)) of 0.58).

Keywords: Artificial Intelligence, Machine Learning, Flood risk, Flood vulnerability, GFI. 

This study was carried out within the RETURN Extended Partnership and received funding from the European Union Next-Generation EU (National Recovery and Resilience Plan -NRRP, Mission 4, Component 2, Investment 1.3 -D. D. 1243 2/8/2022, PE0000005).