Flood susceptibility mapping using hybrid models optimized with Artificial Bee Colony

Floods are the most common type of natural hazard causing economic and human losses. Mapping the susceptibility to flooding is essential for the effective management of flood risk. This paper aims to propose and evaluate four new hybrid models. The Spercheios river basin in Greece was chosen as a case study. The ensemble Random Forest (RF) and Extreme Gradient Boosting (XGBoost) algorithms were combined with the statistical methods of Frequency Ratio (FR) and Weight of Evidence (WoE). The models optimized by the metaheuristic Artificial Bee Colony (ABC) method. Flood inventory of the study (564 locations) developed by Synthetic Aperture Radar (SAR) imagery and flood archive. Flood locations associated with twelve conditioning factors and the dataset randomly split into training and testing sets (70%-30%). Feature selection was performed using the Information Gain Ratio (IGR) method. Input data for the training of the models were the results of FR and WoE. Receiver Operating Characteristic (ROC) curve and statistical metrics were carried out for the validation of the models. It was proved that all models had an AUC value>0.95 and the most precise model was WoE-RF-ABC (AUC = 0.9675). Variable importance of factors showed agricultural lowlands and artificial areas are more likely to be flooded. This study showed that these four optimized hybrid models could accurately predict flood areas prone to flooding and help the decision-makers.

Automated summary

Floods are a common natural hazard that cause significant economic and human losses. This paper proposes and evaluates four new hybrid models for mapping flood susceptibility using the Spercheios river basin in Greece as a case study. The models combine ensemble algorithms with statistical methods and are optimized using the Artificial Bee Colony (ABC) method. The results show that these models accurately predict flood-prone areas and can assist decision-makers.