Investigating the mechanisms of flood susceptibility with the use of multi-basin machine learning models in data-scarce environments in Cyprus

The island of Cyprus is dominated by small-scale watersheds that favor the occurrence of flash floods. Climate projections indicate the increase in frequency and intensity of these events. The development of rapid flood screening tools is essential for better urban planning. This study uses four different machine learning algorithms, namely support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF), and multilayer perceptron (MLP), to build models based on data collected from eight watersheds to enhance their within-region (Cyprus) generalization. Seven features were selected for tuning and testing the performance of these models. T-based confidence intervals were calculated to quantify uncertainty. All models achieved good agreement with the inventory database. RF model was selected to build multi-level susceptibility maps. Half of the Georskipou watershed is classified as highly susceptible to flooding, mostly urban and semi-urban regions, whereas 38 % of the test watershed is not expected to experience severe flood events. Simplified RF models were developed by selecting different combinations of the most important features, revealing that land-use, terrain slope, terrain elevation, and flow accumulation are sufficient to achieve good accuracy (95 %) with flood inventory data. The results highlight the ability of simple, computationally efficient data-driven models to provide rapid predictions, thus avoiding the compilation of fully detailed physically-based models. • A classification machine learning approach was adopted to assess flood susceptibility. • Data were collected from several sites to construct multi-basin machine learning models. • Hyperparameter optimization was conducted to ensure fair comparison among the models. • Land cover and topography are the most influential parameters in FS estimation. • Simplified models are proposed as fast screening tools in data-scarce environments.