Flood Risk Assessment and Driver Analysis in Hengyang City Using Explainable Machine Learning

Flooding is among the most frequent and damaging natural hazards worldwide. Accurate flood risk assessment and rigorous analysis of its driving factors are of critical importance for flood prevention and mitigation. Traditional flood risk assessments tend to focus only on proximity to rivers or a single hydrological indicator, lacking in-depth and objective consideration of the influence of other drainage network structure characteristics. To address this gap, this study develops a comprehensive four-dimensional flood risk evaluation framework that integrates hazard factors, hazard-forming environment, exposed elements, and disaster prevention and mitigation capacity. The framework is validated through a consolidated risk mapping procedure, and the resulting indicator set is combined with explainable machine learning algorithms. A suite of drainage network metrics—including drainage density, channel curvature, tributary development coefficient, and water surface ratio—alongside topographic and soil variables, is incorporated to systematically reveal how water system structure interacts with other controls to shape spatial flood risk.The results indicate a distinct spatial pattern of flood risk across Hengyang City: ① Higher risk concentrates in the central, southeastern, and northeastern parts of the study area, while the western region exhibits relatively lower risk. At the administrative level, the municipal district of Hengyang, Hengshan County, and Changning City show notably elevated risk values. ② Explainable machine learning-based driver analysis reveals that flood risk in the Hengyang section of the Xiangjiang River Basin is predominantly influenced by elevation, drainage density, and soil moisture. Elevation exerts strong control via its effects on runoff generation and flow routing; densely dissected drainage networks accelerate runoff concentration and transmission, increasing inundation potential in downstream lowlands; high soil moisture reduces infiltration capacity and thereby elevates surface runoff and flooding likelihood. Importantly, the relationships between these predictors and flood risk exhibit threshold behavior rather than simple linear trends—risk escalates sharply once critical indicator values are exceeded. Identifying such thresholds enhances the practical utility of the results for flood risk zoning and management.Overall, by integrating a four-dimensional evaluation framework with explainable machine learning and explicit water system metrics, this study quantifies the coupled effects of drainage structure, topography, soil conditions, and human exposure, and pinpoints key thresholds and high-risk spatial patterns. These findings provide scientifically grounded guidance for flood hazard mapping, early warning prioritization, and targeted mitigation measures in the study area and in other basins with similar hydro-geomorphic and socio-environmental characteristics.