Evaluation of Machine Learning Approaches and their Extrapolation for Flood Hazard Mapping

This research has two primary aims: first, to assess the capability of machine learning (ML) techniques generally used for flood susceptibility mapping in replicating flood hazard maps derived from hydraulic modeling, and second, to apply the trained ML models to unseen catchments using a similar set of parameters. The study focuses on two catchments—Argens and Gapeau—located in South-Eastern France. Reference flood hazard maps were generated using the FLOODOS 2D hydraulic model at a 5-meter resolution, simulating water depths for a 1,000-year return period across 1,163.1 km of rivers. From these maps, a balanced dataset of flood and non-flood points was created and split into training and validation subsets (70:30) via random sampling. The analysis employed several geo-environmental factors as explanatory variables, including a 5-meter resolution Digital Terrain Model, Height Above Nearest Drainage, river slope, and river discharge data used in hydraulic modeling. Three advanced ML models—artificial neural networks, random forests, and extreme gradient boosting—were trained on this dataset. These trained models were then tested on the Gapeau region to evaluate their robustness and effectiveness in replicating flood hazards. Model performance was assessed using metrics such as the Area Under the Receiver Operating Characteristic Curve (AUROC) and the Critical Success Index, which measure prediction accuracy for flood extents. Results indicated that ML models effectively mapped flood hazards in complex geo-topographic regions like the Argens basin. Notably, two models achieved AUROC scores exceeding 0.9 when applied to the untrained Gapeau region, demonstrating good transferability and predictive accuracy.