Metaheuristic-optimized ANN for flood susceptibility mapping: a case study in the Khiavchai Watershed, Iran

Abstract This study evaluates and compares four hybrid artificial neural network (ANN) models optimized with metaheuristic algorithms—namely APO-ANN, RUN-ANN, HHO-ANN, and WOA-ANN—for flood susceptibility mapping in the Khiavchai watershed, Ardabil Province, Iran. Input data included flood conditioning factors such as elevation (DEM), slope, aspect, Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Topographic Wetness Index (TWI), distance to rivers, land use, hydrological soil groups, and maximum daily precipitation, derived from Sentinel-1/2, Landsat 8 imagery, and ALOS PALSAR DEM. The flood inventory was prepared using Sentinel-1 SAR imagery for the period 2016–2024. Results indicated that the APO-ANN model exhibited the best performance, achieving the highest accuracy (76.9% on the test set), AUC (0.738), F1-score, and balanced precision-recall, along with the fastest convergence and lowest final error. This superiority was attributed to a more balanced variable importance (integrating topographic, hydrological, and anthropogenic factors) and higher generalization capability. The RUN-ANN and HHO-ANN models showed moderate performance, whereas WOA-ANN performed the weakest due to overfitting. The final flood susceptibility map, generated using the APO-ANN model, identified high-risk areas primarily along main drainage channels, low-lying regions, and the southern part of the watershed. This study presents the first application of the novel APO and RUN algorithms in hydrological modeling and recommends APO-ANN as a superior approach for flood susceptibility mapping in complex watersheds.