Geographical reaction profiles of flood occurrence: insights from statistical, machine-learning, and ensemble models

Flooding remains one of the most destructive natural hazards that threaten human lives, infrastructure, and sustainable land management. Despite advances in machine learning, limited research has systematically integrated multiple metaheuristic optimization algorithms with artificial neural networks (ANNs) to improve prediction accuracy and robustness, especially in less studied regions such as Golestan Province, Iran. Addressing this gap, this study develops and compares four hybrid multilayer perceptron (MLP) models that combine GIS and remote sensing data with ANNs optimized using the cuckoo optimization algorithm (COA), harmony search (HS), training-learning-based optimization (TLBO), and whale optimization algorithm (WOA). Based on 412 recorded flood events, the results show that TLBO-MLP achieved the highest and most consistent prediction performance (AUC = 0.9739), demonstrating strong generalization ability across training and test datasets. COA-MLP improved the AUC value from 0.9561 to 0.9646, while HS-MLP and WOA-MLP achieved AUC values ​​of 0.9427 and 0.9563, respectively. The findings highlight the superior robustness of TLBO-MLP and the different optimization efficiencies among the algorithms. This study provides a new comparative framework for meta-heuristic optimized ANN models and provides practical guidance for accurate flood susceptibility mapping in data-poor areas.