Spatial Uncertainty in Fluvial Flood Susceptibility Mapping: A Focused Analysis within a Polygon-Based Monte Carlo Framework

Fluvial flood susceptibility (FFS) maps are widely used in flood-risk management, yet their reliability may be affected by the point-based representation of spatially heterogeneous flood inventories. This study developed a polygon-based Monte Carlo framework to evaluate prediction uncertainty under repeated within-polygon sampling and applies it, in this focused analysis, to the Random Forest (RF) algorithm. The framework was applied to the Kysuca catchment, Slovakia, using ۲۰۸ inventory polygons (۱۱۳ flood and ۹۵ non-flood polygons) derived from the September ۲۰۲۴ flood event. Twelve flood-conditioning factors at a ۱۰-m resolution were screened for multicollinearity using Pearson correlation analysis and the Variance Inflation Factor (VIF, threshold = ۵), which led to the exclusion of the Topographic Wetness Index (TWI) from the final predictor set. RF was then evaluated across ۵۰۰ Monte Carlo realizations, with prediction uncertainty quantified using the standard deviation (SD), coefficient of variation (CV), and a proposed Uncertainty-Adjusted Susceptibility Index (UASI). RF achieved a mean AUC of ۰.۹۸۳ ± ۰.۰۰۹, Accuracy of ۰.۹۳۳ ± ۰.۰۲۶, and F۱-score of ۰.۹۳۸ ± ۰.۰۲۵ with SD values ranging from ۰.۰۰۱ to ۰.۲۱۹ and no area of the catchment falling into the high or very-high SD classes. The lowest uncertainty was concentrated along river channels and adjacent floodplains, while elevated and steep upland areas showed greater prediction variability. These findings show that RF combined with polygon-based Monte Carlo sampling offers a stable, transparent basis for uncertainty-aware flood susceptibility mapping.

کلیدواژه ها:

Fluvial flood susceptibility ، Random Forest ، Spatial uncertainty ، Polygon-based sampling