Physically consistent and spatially conformalized machine learning for trustworthy flood susceptibility mapping of the 2024 Feni flash flood

Data-driven flood susceptibility mapping (FSM) has converged on a common template: a gradient-boosting classifier, a random train/test split, a global feature-attribution chart, and a five-class map. Using the catastrophic August-2024 Feni flash flood (Bangladesh) with 11 Google-Earth-Engine conditioning factors ( $$n=1{,}510$$ inventory samples; 982, 866 mapped pixels at 30m), we examine three limitations it can obscure. First, random splitting inflates the area under the ROC curve (AUC) by about 0.06–0.09 relative to spatially blocked cross-validation. Second, an unconstrained learner can attain high accuracy while encoding hydrologically implausible response functions, monotone in the physically expected direction on only $$54.5\%$$ of its response surface. Third, its outputs are neither calibrated nor accompanied by valid uncertainty. The proposed framework injects hydrological priors as hard monotonicity constraints, enforcing physically consistent behaviour at no measurable cost in skill (spatial AUC 0.938; Matthews correlation 0.700). Because the inventory is balanced by design, we report a relative, event-conditioned susceptibility score rather than an absolute occurrence probability. Predictions are calibrated with cross-fitted isotonic regression (expected calibration error 0.150–0.016) with spatially stratified (Mondrian) conformal sets exposing a conditional-coverage degradation under extrapolation (worst-region coverage 0.48 at the 0.90 level). A Physical Consistency Score (PCS) audits whether learned relationships obey their hydrological priors. The imposed directions are additionally confirmed against the observed inundation, and sensitivity analyses over the constraint set, spatial-block count and calibrator leave the conclusions unchanged. Validated against the spatially independent (same-event) observed 2024 inundation extent, the map yields a strictly monotone class-frequency response (observed flooding rising from $$1.0$$ to $$45.4\%$$ ; AUC 0.75). The contribution reframes FSM around physical consistency, calibrated probability and honest spatial uncertainty.