From social media to flood 3D scenes: leveraging large language models and knowledge graphs for flood spatiotemporal evolution representation

Flood disaster evolution exhibits strong spatiotemporal coupling and complex causality. Intuitive representations of flood disaster evolution can help the public better understand disaster processes and their underlying evolution mechanisms, thereby enhancing risk awareness. However, most existing studies focus on numerical modeling of flood disaster evolution, which has limited capability in effectively capturing and representing the evolving processes of flood disasters. Social media posts generated during floods contain rich information and provide valuable insights into the evolution of flood events. In this context, this study leverages large language models (LLMs) and knowledge graphs (KGs) to extract flood spatiotemporal knowledge from social media, thereby reconstructing the disaster evolution pathway for representation as 3D scenes. Specifically, ontology-constrained LLMs for flood knowledge extraction, evolutionary pathway reconstruction and 3D representation are introduced in detail. We collected a total of 34,731 tweets related to the 2021 European floods for experimental analysis and developed a prototype system for 3D flood representation. The experimental results demonstrate that the proposed method achieves reliable performance in disaster knowledge extraction and evolutionary pathway reconstruction. Moreover, the augmented 3D representation improves the interpretability and clarity of flood disaster evolution, thereby facilitating a better understanding of disaster processes and supporting disaster prevention and mitigation.