DEEP LEARNING–ENABLED LIDAR AND MULTISPECTRAL SIGNATURE FUSION FOR FLOOD HAZARD MAPPING AND LAND-SURFACE VULNERABILITY PREDICTION

Abstract

This study developed and validated a quantitative flood hazard and land-surface vulnerability modeling framework that integrated LiDAR-derived terrain information with multispectral surface-condition signatures using deep learning and benchmark statistical approaches. Flood vulnerability was operationalized as a continuous, probability-like spatial outcome, and model performance was evaluated under a decision-utility framework that emphasized discrimination, calibration, and ranking quality in a rare-event setting. The analysis was conducted using a retrospective observational design applied to a finalized dataset of approximately 1.25 million raster cells, of which 62,500 (5.0%) were labeled as flood-positive based on event-linked inundation references. The dataset was partitioned into training (70%), validation (15%), and holdout test (15%) subsets using spatially disciplined splits to reduce leakage from spatial autocorrelation. Descriptive results indicated that flood-positive observations were concentrated in low-elevation, low-slope, high-convergence terrain and were more frequently associated with elevated imperviousness and wetness signatures. Reliability analysis showed acceptable internal consistency for composite predictor constructs, with Cronbach’s alpha values ranging from 0.75 to 0.86 across terrain and multispectral feature families. Baseline regression modeling confirmed statistically significant associations for all major construct families, with the strongest effects observed for convergence and wetness proxies and negative associations for elevation and slope structure. Comparative model evaluation demonstrated that non-linear approaches outperformed the regression benchmark. The fused LiDAR–multispectral convolutional neural network achieved the highest discrimination on the holdout test set, with an AUC-ROC of 0.94 and an AUC-PR of 0.63, compared with 0.84 and 0.41, respectively, for the regression baseline. Calibration quality also improved, as reflected by a lower Brier score (0.084 versus 0.118) and reduced expected calibration error (0.025 versus 0.041). Ranking utility gains were substantial, with precision at the top 1% of ranked locations increasing from 0.34 for regression to 0.52 for the fused model, and top-decile capture improving from 0.52 to 0.74 while reducing false alerts from 620 to 470 per 10,000 evaluated cells. Robustness testing across regional, catchment, and terrain-class holdouts confirmed that performance gains persisted under spatial generalization. Overall, the findings demonstrated that integrating LiDAR microtopography and multispectral surface-condition information through disciplined data fusion and spatially credible validation produced statistically robust and decision-useful flood vulnerability maps.