Researchers built a balanced flood inventory containing 400 validated flood locations and 400 spatially constrained non-flood locations. Flood records came from UNOSAT flood-extent data and were independently verified using changes in Sentinel-1 Synthetic Aperture Radar backscatter in Google Earth Engine. The models used 14 flood-conditioning factors representing topographic, hydrological, climatic, environmental and anthropogenic controls. Seven approaches were trained and evaluated against an independent test dataset: Random Forest, ExtraTrees, XGBoost, LightGBM, CatBoost, a Soft Voting Ensemble and a Stacking Ensemble.

CatBoost produced the highest individual-model discrimination, with an area under the curve of 0.9960. The Soft Voting Ensemble achieved the best overall classification performance, recording 97.9% accuracy, an F1 score of 0.9793, recall of 0.9833 and a Kappa coefficient of 0.9583. These results describe how accurately the models classified locations in the study’s test dataset; they do not mean that flooding will occur at every location identified as susceptible. The mapped classes are model-based estimates of flood susceptibility based on the selected variables, inventory and spatial sampling.

SHAP-based model interpretation identified land use and land cover, elevation, distance to river, the Height Above Nearest Drainage metric and the Topographic Wetness Index as the dominant predictive correlates. Their prominence was consistent with physically meaningful floodplain processes. In addition to the very high-susceptibility class, the final map established a “Reliable Risk” zone covering 540.66 square kilometers, or 6.53% of the district. Spatial validation showed that 98.3% of documented flood points fell within areas where all seven models agreed on the susceptibility classification.

The research was a district-scale susceptibility study rather than a forecast of flood frequency, depth or damage, and its results depend on the quality and representativeness of the flood inventory, radar-derived validation data and selected conditioning factors. Within those limits, the framework combines ensemble prediction with model interpretation to identify areas that may warrant attention in flood-risk reduction, land-use planning and disaster preparedness. The authors present the approach as a potentially useful operational tool for mapping flood susceptibility in the study area.

Subject of Research: Explainable ensemble machine learning for flood-susceptibility mapping in Luuq District, Somalia

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

Article References: Dhaqane, A. I., & Al Sayeem, M. A. (2026). Hybrid explainable ensemble machine learning framework for flood susceptibility prediction in Somalia. Discover Sustainability. https://doi.org/10.1007/s43621-026-04629-0

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04629-0

Keywords: flood susceptibility, machine learning, Somalia, ensemble models, Sentinel-1 radar, SHAP analysis, Jubba River, satellite mapping, disaster preparedness, Hybrid, explainable, ensemble