Attributed graph neural network-based flood vulnerability assessment of river basin systems

Abstract Flood vulnerability assessment is a critical component of flood risk management, particularly in regions with complex river networks and limited hydrological data. This study proposes a graph neural network-based framework to evaluate river system vulnerability, representing each river segment as a node with hydrological and geomorphological attributes. Two GNN models were employed to generate vulnerability scores by jointly considering node attributes and network structure. High-risk river segments were first identified based on these scores, and the results were then aggregated to delineate flood-sensitive sub-basins. A case study in Guangxi, China, using the Xijiang River system, shows that the two models converge on similar high-risk areas, with an overlap of 60% in identified high-risk segments, aligning well with observed flood patterns. This highlights the robustness and practical reliability of the proposed approach. The framework offers a practical, data-efficient tool for identifying vulnerable river segments and flood-prone sub-basins, supporting flood risk management and decision-making in complex river systems.