Applying a Random Forest Modeling Approach for Delineating Flood Susceptibility in WV, USA

Flooding remains one of the most common and damaging natural hazards in West Virginia, where steep terrain, narrow valleys, and dense development along river corridors increase flood risk. Although FEMA Flood Insurance Rate Maps (FIRMs) provide the regulatory basis for floodplain management, these maps are binary, often outdated, and cannot reflect gradual changes in flood susceptibility across the landscape. Identifying flood-prone areas across the entire state requires an approach that combines multiple environmental indicators and links them to historical flood event data. Therefore, this study develops a machine-learning-based flood-susceptibility model to provide a continuous representation of relative flood susceptibility across West Virginia using environmental factors. Using high-water-mark observations and six environmental factors, including precipitation, topographic wetness index, ecological land units, land cover, impervious surface, and karst presence, a Random Forest model was trained to estimate flood susceptibility across West Virginia. Model performance was assessed with standard classification metrics and showed strong predictive ability, with an overall accuracy of 91.5%, a sensitivity of 94.4% for flooded locations, and a specificity of 88.7% for pseudo-absence locations. The held-out test dataset produced an AUC of 0.958 (95% CI: 0.945–0.971). For map visualization, the continuous Random Forest output was classified into five susceptibility classes using Jenks Natural Breaks. The output is interpreted as a relative susceptibility score, not an annual flood probability. These results suggest that environmental indicators can capture spatial patterns of flood susceptibility. The model is intended to complement, rather than replace, FEMA Flood Insurance Rate Maps by identifying areas with higher modeled susceptibility than current designations indicate. Since the model relies on widely available environmental data, it can be updated as input layers change and may be adapted to regions with comparable data, although local validation would be required. By providing a flexible, data-driven depiction of relative flood susceptibility, this approach can support flood resilience planning and risk communication at the county and state levels.