Probabilistic Spatial Completion of FEMA Special Flood Hazard Area Coverage in Louisiana Using Conditional Diffusion and Distributionally Trustworthy Explanation

Effective flood planning requires spatially complete hazard information, yet regulatory flood products can contain unresolved hazard classifications and provide limited uncertainty information. In addition, existing models for flood mapping often rely on random sampling that neglects spatial dependence among neighboring areas, struggle to model zero-inflated, bounded area shares (i.e., shares with many zero values and a bounded 0–1 range), and offer limited distribution-level explainability. To address these limitations, this study developed a calibrated hurdle conditional diffusion framework to estimate the Federal Emergency Management Agency (FEMA) Special Flood Hazard Area (SFHA) share and the probability that coverage exceeds 10% for Louisiana census block groups. Within this framework, a hurdle component separates zero from positive coverage, a conditional diffusion model estimates bounded positive-share distributions, spatial blocking evaluates geographic transfer, and probabilistic calibration supports exceedance probabilities and prediction intervals. Terrain, land-cover, wetland, hydrographic, climate, and soil predictors were evaluated across 3776 block groups with resolved FEMA information. The conditional regression component achieved a mean absolute error of 0.172 for the primary continuous SFHA-share prediction and strong parish-level agreement (Pearson r = 0.834). For the secondary screening task of identifying census block groups with ≥10% SFHA share, the calibrated hurdle diffusion model yielded an area under the receiver operating characteristic curve of 0.852. Statewide predictions were generated for all 4294 census block groups, including 518 unresolved units with a mean predicted SFHA share of 33.3%. Distributional Reliability Explanation Attribution (DREA) identified flooded soils, elevation, topographic wetness, wetlands, and water proximity as reliable predictors of distributional displacement, uncertainty, exceedance probability, and probabilistic skill. Overall, the framework supports leakage-safe, uncertainty-aware screening while complementing authoritative FEMA flood maps.