Spatially Validated Machine Learning Flood Susceptibility Assessment in the Volta River Basin Using Extreme Rainfall and Multi-Source Geospatial Data

Abstract

Flood susceptibility mapping in large transboundary basins is often constrained by sparse observations, heterogeneous environmental conditions, and validation strategies that overstate model skill when spatial dependence is ignored. This study develops a machine learning framework for the Volta River Basin integrating terrain, drainage, rainfall extremes, vegetation, land cover and soil properties. Fifteen predictors derived from SRTM, MERIT Hydro, CHIRPS, MODIS NDVI, ESA WorldCover and OpenLandMap were harmonized at a nominal 1 km scale. Flood occurrence was obtained from a satellite-based historical flood inventory, with persistent water excluded using JRC Global Surface Water. A balanced dataset of 10,000 samples comprised 5000 flood and 5000 background observations, with background samples located beyond a 3 km buffer around mapped floods. Samples were grouped into 0.5° spatial blocks, producing 127 training and 43 independent holdout groups. Random Forest outperformed XGBoost, achieving 95.26% accuracy, 95.06% balanced accuracy, 97.46% precision, 92.20% sensitivity, 97.92% specificity, F1 = 0.9476, ROC-AUC = 0.9923 and average precision = 0.9906. A three-predictor terrain/drainage baseline using only elevation, distance to river and drainage density achieved 90.29% accuracy and ROC-AUC = 0.9676 on the identical spatial holdout, showing measurable incremental discrimination from the broader predictor set. Variable importance and SHAP assigned their largest model-attributed contributions to elevation, distance to river and drainage density. Very-low susceptibility dominated basin-wide, while high and very-high susceptibility followed low-lying, drainage-connected corridors. Because balanced sampling was used without probability calibration, the mapped values are relative susceptibility scores rather than calibrated flood probabilities. The map supports regional screening, prioritization and planning rather than event-specific hydraulic inundation prediction.