Multimodal machine learning for urban flood susceptibility in data-scarce Sahelian cities: Ouagadougou, Burkina Faso

Sahelian cities face increasingly damaging floods, yet the susceptibility maps guiding local planning are usually produced by expert-weighted multicriteria methods whose weights are subjective and rarely validated against observed floods. We develop a multimodal Geospatial Artificial Intelligence (GeoAI) framework for Ouagadougou, Burkina Faso, fusing fourteen factors from Copernicus and MERIT Hydro terrain, Sentinel-1 and Sentinel-2 imagery, built-up surfaces, OpenStreetMap drainage, soil and CHIRPS rainfall in Google Earth Engine. Random Forest and XGBoost classifiers were trained on a hybrid inventory of 1,593 points, pairing Sentinel-1 detections from five rainy-season events (2016–2022) with documented black-spots, and evaluated under spatial block cross-validation. Two design choices known to inflate accuracy were tested: discrimination falls from AUC 0.904 to 0.874 when the absence-exclusion buffer is removed, and a coordinates-only control reaches 0.805, leaving the predictors a stable gain of +0.068. The buffer-free model significantly outperforms a conventional Analytic Hierarchy Process baseline parameterized from regional literature (0.771; p = 0.018). Mapped extent proves far more assumption-dependent than the pixel ranking (Spearman r = 0.960), the two highest classes spanning 1.9–15.9% of the commune. Because every metric rests on this single inventory, with no independent external ground truth, the map is an operational ranking pending field validation.