Integrating machine learning and multi-criteria decision analysis for flood susceptibility mapping in the Hadejia–Jama’are River Basin

Abstract

Flooding remains a recurrent environmental hazard in semi-arid northern Nigeria, particularly within the Hadejia–Jama’are River Basin, where the interaction of topographic, hydrological, and anthropogenic factors intensifies vulnerability. This study developed an integrated geospatial modelling framework combining machine learning techniques with Multi-Criteria Decision Analysis (MCDA) to produce accurate and spatially explicit flood susceptibility maps. A hybrid approach integrating the Analytical Hierarchy Process (AHP) with Support Vector Machine (SVM), Artificial Neural Network (ANN), and Random Forest (RF) models was applied using ten conditioning factors, including elevation, slope, rainfall, drainage density, Topographic Wetness Index, land use/land cover, soil texture, flow accumulation, and river proximity, standardized to 30 m resolution. Multicollinearity diagnostics (VIF < 5; TOL > 0.2) ensured model reliability, while performance was evaluated using ROC–AUC metrics. Results revealed that low-lying northern zones (316–435 m), areas with high drainage density (27.2–34.2 km/km²), and regions receiving 1180–1440 mm rainfall exhibited the highest flood susceptibility. The AHP model classified 83.7% of the basin as moderate risk, whereas SVM showed more spatial differentiation with 29.2% moderate, 27.3% high, and 11.5% very high susceptibility (6,323.11 km²). ANN and RF models also indicated dominance of moderate susceptibility (76.1% and 73.3%), with high-risk areas reaching 18.5% and 22.7%, respectively. AHP achieved the highest predictive accuracy (AUC = 0.91), followed by SVM (0.83), RF (0.81), and ANN (0.79), demonstrating that while AHP ensures robustness, machine learning models provide better spatial discrimination of flood-prone zones.