A machine learning framework for flood risk assessment with imbalanced data

This study presents a machine-learning framework for identifying areas of elevated flood risk using imbalanced, high-dimensional geospatial datasets. Using the Mandra River Basin in Greece as a case study, an extreme gradient boosting (XGBoost) classifier was trained on 11 flood-related features including hydrological, meteorological, soil and topographic data, whereas 454 geolocated citizens' emergency calls were used as target information. Class imbalance was addressed through a weighted-only strategy to preserve the integrity of the original dataset, while adjusting class contributions during the training process. Bayesian optimization and stratified 10-fold cross-validation were combined to tune and evaluate the model performance. Threshold-sweep analysis enables the identification of multiple operational settings based on precision–recall trade-offs and cost-sensitive criteria. The results revealed physically interpretable feature importance patterns, with terrain elevation (TE) (45%) and accumulated rainfall (25%) emerging as the dominant predictors according to feature importance analysis, followed by curve number (16%) and the Manning coefficient (9.2%). The proposed framework provides a reliable, data-driven tool for flood risk assessment, and supports a flexible, two-tier decision strategy that differentiates between emergency escalation and broader situational awareness.