A Transformer-based Graph Network for Flash Flood Disaster Classification

Flash floods are one of the most devastating natural disasters, posing significant risks to both human life and infrastructure. The classification of their underlying drivers—such as high precipitation, dam breaches, landslides, and melting snow—remains a critical yet challenging task, especially in regions like China, where diverse geographical and climatic factors exacerbate disaster complexity. In this study, we propose a Transformer-based Graph Network (TGN) designed to tackle these challenges by leveraging a dataset of over 53,000 flash flood events, each characterized by non-uniform geographical attributes and varying levels of data completeness. Unlike traditional graph neural networks (GNNs) that depend on predefined graph structures, TGN dynamically learns and refines edge weights during training, enabling it to uncover asymmetric dependencies. This adaptability is particularly valuable when explicit relationships between nodes are unavailable or incomplete.

Integrating multi-head self-attention mechanisms from Transformer architectures, TGN captures complex interdependencies across watershed features while maintaining interpretability through sparsity and diversity constraints. A distinguishing feature of this framework is its ability to identify meaningful graph structures without prior knowledge, offering insights into critical connections and interactions within disaster-prone regions. For instance, our experiments demonstrate how TGN emphasizes high-risk upstream-downstream relationships, providing actionable knowledge for localized flood management. The model significantly outperforms traditional GNNs and machine learning methods in accuracy and robustness, achieving superior classification performance across all four disaster categories. Furthermore, the TGN framework is supported by rigorous evaluation metrics, including Precision, Recall, F1-score, and Overall Accuracy, ensuring its reliability in real-world applications.

By combining innovative graph-based modeling with interpretable mechanisms, this study bridges the gap between theoretical advancements and practical disaster management. The proposed approach not only enhances prediction capabilities but also provides an analytical lens for understanding the intricate relationships among flash flood drivers, paving the way for more effective mitigation strategies and informed decision-making. This work underscores the transformative potential of adaptive graph neural networks in addressing complex environmental challenges and advancing the state of flood risk assessment.