This study presents a region-based dual machine learning framework for flood hazard mapping, combining flood-prone area classification and flood depth estimation. Flood inventory points were generated from Sentinel-1 Synthetic Aperture Radar (SAR) imagery processed via Google Earth Engine, serving as the primary input for model training. A set of flood conditioning factors—including topographic, hydrologic, and land surface variables—were compiled using Digital Elevation Model (DEM) data, land use/land cover, soil type, slope, drainage density, and proximity to rivers. These inputs were used to train and evaluate multiple classification models, with the bagged decision tree algorithm yielding the highest accuracy (93.2%) with Area under ROC curve (AUC) 0.933 for flood-prone area identification. The feature importance analysis highlighted slope, elevation, and distance to river as key predictors. For flood depth estimation, a regression modeling approach was adopted using the same database, incorporating flood inventory depth values extracted from SAR backscatter calibration and validated against in situ or ancillary datasets. Among the several tested models, the bagged regression tree outperformed others in terms of RMSE (0.99 m) and R 2 (0.6), demonstrating its suitability for spatially continuous flood depth mapping. The results were integrated into a GIS-based flood hazard mapping framework, enabling visualization of both flood extents and their associated depths. This machine learning-driven approach facilitates high-resolution (30 m grid size) flood risk assessment, particularly useful in data-scarce regions, and provides a valuable decision-support tool for disaster preparedness and infrastructure planning.
Sadhwani et al. (Wed,) studied this question.