Data-driven flood susceptibility assessment using hybrid machine learning and optimization techniques: case of the Sedrata Watershed, NE Algeria

Abstract

Flooding is one of the most disastrous natural hazards around the globe, causing enormous ecological and socio-economic losses; therefore, reliable assessment tools are required for informed risk management. This research proposes a hybrid flood susceptibility modeling framework that incorporates the Random Forest Regressor (RFR) model coupled with three metaheuristic optimization algorithms, namely Grasshopper Optimization Algorithm (GOA), Salp Swarm Algorithm (SSA), and Ant Colony Optimization (ACO). The proposed approach is applied to the Sedrata Watershed, Northeastern Algeria, using satellite data along with twelve physiographic and environmental conditioning factors, including slope, rainfall, land use, drainage density, curvature, convexity, and aspect. A total of 317 flood locations detected from Sentinel-1 imagery were used for training and validation of the hybrid models. In this respect, the Weight of Evidence and Geographically Weighted Regression methods have been applied to investigate the influence of each conditioning factor. These hybrid models significantly enhanced predictive performance when compared to the standalone RFR model, showing Area Under Curve (AUC) values of 0.928, 0.925, and 0.920 for RFR-ACO, RFR-SSA, and RFR-GOA models respectively, against 0.904 obtained by the baseline model. The spatial mapping showed that about 26% of the study area has high to very high flood susceptibility. These findings confirm the strong potential of combining Machine Learning and metaheuristic optimization in view of enhancing the predictive performance offlood susceptibility maps. The proposed framework offers an interpretable decision-support tool for sustainable flood risk management and land use planning in the Sedrata region.