Intelligent geospatial modelling of flash flood susceptibility using ICO-optimized hybrid ensembles

Small river basins with steep terrain and rapid hydrological response are particularly vulnerable to flash floods, yet reliable and transferable susceptibility models for such catchments remain limited, as most existing studies rely either on classical bivariate statistics or standalone machine learning classifiers without a systematic hyperparameter-optimization step. This study addresses this gap by proposing an integrated framework for flash flood susceptibility assessment that combines a classical statistical model, weights of evidence (WoE), with three machine learning models (K-star (K*); logistic model tree (LMT); random forest (RF)) optimized through the iterative classifier optimizer (ICO): ICO-K*-WOE, ICO-LMT-WOE and ICO-RF-WOE. The analysis was carried out in the Comana Basin (Romania) and was based on ten relevant conditioning factors (morphometric, hydrological and land use) integrated within a geographic information system (GIS) framework. Model training and hyperparameter tuning were performed using 10-fold cross-validation, while performance was independently assessed on a 30% hold-out validation sample through ROC curves (area under curve (AUC)), confusion-matrix-derived statistics (sensitivity, specificity, accuracy) and the Kappa index. All the models demonstrate excellent performance (AUC > 0.93), with ICO-RF-WOE performing the best. High-flood-risk areas are concentrated mainly on steep slopes. Combining bivariate statistics with iteratively optimized classifiers improves the accuracy and interpretability of the assessments.