Urban expansion and waterlogging susceptibility in Dhaka city: an explainable ensemble machine learning model and urban morphology framework

Urban waterlogging is a growing challenge in rapidly urbanizing cities, yet existing assessments in Bangladesh often rely on expert-based, hydrological, or conventional geospatial approaches that may inadequately capture complex nonlinear interactions among environmental and urban factors. This study develops an integrated hydro-environmental and urban-morphological framework for assessing waterlogging susceptibility in Dhaka City. A composite waterlogging inventory was constructed by integrating historical recurrent hotspot records from the Bengal Institute map in 2024 with event-based field verification conducted across eight locations during the July 2026 record-rainfall event. Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM) were evaluated using hydro-climatic, topographic, environmental, infrastructural and socio-demographic predictors, with spatially blocked five-fold cross-validation used to assess model predictability. SHapley Additive exPlanations (SHAP) were employed to identify predictor importance and directional associations. Contemporary susceptibility patterns were linked to urban expansion, while urban morphological changes during 2010-2025 were analyzed to characterize long-term urban transformation. All models showed strong spatial discrimination under spatially blocked five-fold cross-validation, with fold-wise mean ROC-AUC across four folds of 0.87 ± 0.07 for XGBoost, 0.85 ± 0.08 for RF, and 0.84 ± 0.09 for LightGBM (pooled out-of-fold ROC-AUC: 0.91-0.93). SHAP revealed a consistent predictor hierarchy, with rainfall and elevation (DEM) as the predictors most strongly associated with higher modelled susceptibility based on cross-fold mean rankings and full-data models, followed by road and settlement densities. Model predictions represent a relative susceptibility index rather than calibrated probabilities due to the artificially balanced 50:50 training prevalence. High-susceptibility areas were concentrated in densely developed zones with high flow accumulation and intensive urban infrastructure. The findings demonstrate that integrating spatially robust machine learning, explainable modelling and urban morphological analysis can advance waterlogging assessment toward more interpretable and spatially targeted urban planning.