From Prediction to Prevention: An Explainable GeoAI Framework for Flood Susceptibility and Urban Exposure Assessment Using Machine and Deep Learning Models

Rapid urbanisation and intensifying rainfall have increased cities' vulnerability to flooding, posing major challenges to sustainable development. Although machine learning models have improved flood prediction accuracy, most remain limited by their black-box nature and lack of actionable insights. This study integrates geospatial techniques with explainable artificial intelligence (XAI) to create a transparent, prevention-oriented framework for flood susceptibility mapping. Five ensemble models (random forest, extremely randomised trees, extreme gradient boosting, light gradient boosting machine, and categorical boosting) and two deep learning models (convolutional neural network and deep neural network) were developed using topographical, hydrological, climatic, and anthropogenic predictors. Model performance was assessed using AUC, accuracy, precision, recall, F1-score, Brier score, and Matthews correlation coefficient (MCC). CatBoost performed best, achieving a cross-validation AUC of 0.975 ± 0.005 and an independent test AUC of 0.956, with strong calibration (Brier score = 0.075) and high accuracy (0.904). XGBoost (0.949), LightGBM (0.948), Random Forest (0.948), and Extra Trees (0.945) followed closely, whereas CNN and DNN achieved test AUCs of 0.928 and 0.927. To enhance model interpretability, SHapley Additive exPlanations and Accumulated Local Effects analyses were compared in their ability to characterise susceptibility-relevant response patterns. Elevation, Normalised Difference Vegetation Index, and proximity to rivers emerged as key drivers, with values of < 50 m, < 0.15, and < 525 m delineating high-susceptibility zones. Combining predictive precision with interpretability, the proposed explainable GeoAI framework bridges the gap between flood prediction and prevention, enabling data-driven planning for resilient, climate-adaptive cities.