Integrating SAR remote sensing and machine learning for comprehensive flood risk assessment in Kogi state, Nigeria

Flooding in Nigeria's Niger-Benue confluence threatens millions, yet existing assessments rely on optical imagery vulnerable to cloud cover and lack critical depth information for damage assessment. This study integrates SAR and machine learning to characterize flood dynamics in Kogi State across three major events (2018, 2020, 2022) with four objectives: (1) map flood extent using cloud-penetrating Sentinel-1 SAR, (2) estimate flood depths through cost-distance interpolation, (3) assess land use-specific exposure via SAR-optical data integration, and (4) predict susceptibility using ensemble machine learning. Sentinel-1 GRD imagery, Sentinel-2 multispectral data, NASA DEM, and CHIRPS precipitation records were processed through Google Earth Engine and ArcGIS Pro. Flood depths were modeled using cost-distance interpolation from corrected DEMs, while Random Forest and eXtreme Gradient Boosting algorithms were trained on 290 spatial points using nine environmental conditioning factors (elevation, slope, NDVI, river proximity, topographic wetness index, flow accumulation, soil type, land cover, and precipitation). Flood extent ranged from 161 km 2 (2020) to 405 km 2 (2022), with maximum depths reaching 3.82 m in riparian zones. Cultivated land consistently comprised 75-85% of inundated areas, while built-up exposure doubled from 5 km 2 to 10 km 2 , reflecting unplanned urban expansion into high-risk zones. XGB achieved 93.1% accuracy (AUC = 0.94), with elevation (>50%), river proximity (∼22%), and slope (∼9%) as dominant predictors. This framework demonstrates that SAR-machine learning integration provides reliable flood risk assessment in data-limited tropical regions where traditional monitoring proves inadequate, with full transferability to vulnerable sub-Saharan African basins using freely available data and open-source tools.