Hybrid explainable ensemble machine learning framework for flood susceptibility prediction in Somalia

Flooding remains one of the most severe hydro-meteorological hazards in Somalia, particularly along the Jubba River system, where recurring inundation threatens settlements, agricultural land, infrastructure, and livelihoods. This study developed a hybrid explainable artificial intelligence framework for high-resolution flood susceptibility mapping in Luuq District, Gedo Region, Somalia, using Sentinel-1 Synthetic Aperture Radar flood validation, geospatial conditioning factors, ensemble machine learning, and SHAP-based model interpretation. A balanced flood inventory of 800 samples, consisting of 400 validated flood points and 400 spatially constrained non-flood points, was constructed from UNOSAT flood extent data and independently verified using Sentinel-1 SAR backscatter changes in Google Earth Engine. Fourteen flood-conditioning factors representing topographic, hydrological, climatic, environmental, and anthropogenic controls were used to train and evaluate seven predictive models: Random Forest, ExtraTrees, XGBoost, LightGBM, CatBoost, Soft Voting Ensemble, and Stacking Ensemble. The models achieved excellent predictive performance on the independent test dataset, with AUC values ranging from 0.9930 to 0.9960. CatBoost produced the highest individual AUC of 0.9960, while the Soft Voting Ensemble achieved the best overall classification accuracy of 97.9%, F1-score of 0.9793, recall of 0.9833, and Kappa coefficient of 0.9583. SHAP analysis identified land use and land cover, elevation, distance to river, HAND, and TWI as the dominant predictive correlates of flood susceptibility, consistent with physically meaningful floodplain processes. The final susceptibility map showed that Very High flood susceptibility covered 407.77 km 2 , representing 4.93% of the district, while the Reliable Risk zone covered 540.66 km 2 , representing 6.53% of the study area. Spatial validation demonstrated that 97.8% of documented flood points fell within the Very High susceptibility class, and 98.3% occurred within full multi-model consensus zones. These findings confirm the robustness, interpretability, and operational value of the proposed framework for flood risk reduction, land-use planning, and disaster preparedness in data-scarce arid riverine environments.