Uncertainty-Aware Flood Susceptibility Mapping Using Calibrated Ensemble Machine Learning in the Indus River Basin, Pakistan

Flood susceptibility mapping is an important component of spatial flood-risk assessment; however, conventional machine-learning approaches provide limited information about prediction reliability and spatial uncertainty. This study develops an uncertainty-aware flood susceptibility framework for the Indus River Basin (IRB), Pakistan, integrating satellite-derived flood observations, GIS-based environmental predictors, spatially blocked validation, probability calibration, ensemble learning, and uncertainty quantification. A flood inventory was derived from Landsat-9 observations of the July–August 2025 monsoon flood event, and environmental predictors were harmonized at a spatial resolution of 500 m. Four tree-based algorithms (LightGBM, XGBoost, Extra Trees, and HistGBM) were evaluated using five-fold spatially blocked cross-validation. Isotonic regression was applied to calibrate model probabilities, which were subsequently combined using a calibrated averaging ensemble (CAE). The CAE achieved mean area under the receiver operating characteristic curve (ROC-AUC) and area under the precision–recall curve (PR-AUC) values of 0.878 and 0.837, respectively, and demonstrated improved probabilistic reliability compared with the individual classifiers. High- and very-high-susceptibility zones were concentrated across the central and lower Indus floodplains and major tributary corridors. Entropy analysis identified areas of model agreement and greater predictive uncertainty, while SHAP analysis highlighted drainage density, proximity to rivers and water bodies, the Standard Precipitation Index (SPI), and the topographic wetness index as influential predictors. These findings demonstrate the value of integrating spatial validation, probability calibration, ensemble learning, and uncertainty quantification for basin-scale flood susceptibility mapping. Because the analysis is based on a single flood event, the resulting maps represent event-informed susceptibility patterns rather than deterministic predictions of future flood extent or depth.