Multi-Source Remote Sensing and Dem-Derived Hydrogeomorphic Indicators for High-Resolution Urban Flood Susceptibility Mapping: A Spatial Cross-Validation Framework
DOI:
https://doi.org/10.63125/z4xt7k38Keywords:
Urban Flooding, Remote Sensing, Hydrogeomorphology, Spatial Cross-Validation, Susceptibility MappingAbstract
Urban flooding represents a spatially complex hazard influenced by interactions among terrain configuration, drainage connectivity, land-surface modification, and hydroclimatic conditions. This study developed and quantitatively evaluated a high-resolution urban flood susceptibility mapping framework integrating multi-source remote sensing with DEM-derived hydrogeomorphic indicators under geographically explicit spatial cross-validation. The analysis covered 1,286.4 km² at 10 m spatial resolution and incorporated 1,284 verified flood locations and 1,284 spatially screened non-flood observations, producing a balanced analytical sample of 2,568 locations with 99.3% geospatial data completeness. Seventeen predictors were retained following quality, correlation, and multicollinearity screening. Flood occurrence exhibited significant spatial clustering, with Moran’s I = 0.41 (p < 0.001). HAND was the most influential predictor, contributing 13.8% mean importance, followed by distance to drainage at 12.9%, flow accumulation at 11.7%, elevation at 10.8%, and TWI at 9.6%. Logistic regression, Support Vector Machine, Random Forest, gradient boosting, XGBoost, and ensemble models were comparatively evaluated. Under conventional random validation, the ensemble achieved 91.7% accuracy, 0.917 F1-score, 0.972 ROC-AUC, and 0.969 PR-AUC. Spatial cross-validation produced more conservative estimates of 85.6% accuracy, 0.856 F1-score, 0.914 ROC-AUC, and 0.907 PR-AUC. The ROC-AUC reduction was statistically significant (p = 0.001) with a large effect size of Cohen’s d = 1.26. Resolution analysis showed that ROC-AUC decreased from 0.914 at 10 m to 0.901 at 20 m and 0.886 at 30 m. Removing hydrogeomorphic predictors further reduced ROC-AUC to 0.852. The final ensemble produced a Brier score of 0.112, calibration slope of 0.93, residual Moran’s I of 0.07, exact susceptibility-class agreement of 84.7%, and adjacent-class agreement of 96.1%. The findings demonstrated that integrating high-resolution hydrogeomorphic and multi-source remote-sensing information with spatially explicit validation provided a robust approach for distinguishing predictive accuracy from genuine geographic generalization in urban flood susceptibility mapping.


