Extracting spatiotemporal disaster knowledge from massive, heterogeneous social media data is crucial for urban flood management but it remains technically challenging. This study proposes a unified framework that integrates Chinese Multimodal Large Language Models with agents to automate cross-modal extraction, using water depth as a representative case. Systematic evaluations against deep learning baselines demonstrate that knowledge-guided prompting enhances geographic extraction precision by 10%–25%. Specifically, DeepSeek R1 and Doubao-1.5-thinking-vision-pro excel in textual and visual tasks, respectively, and they are jointly adopted to maximize overall performance. The framework is applied to urban flood events in China (July–August 2021) to demonstrate practical utility. The framework constructs a structured database containing 96,826 spatiotemporal water depth records by processing 1.53 million texts and 240,000 images. The results successfully capture the evolution of flood events, verifying the robustness of the proposed approach. This study establishes a validated, data-driven approach using agent-based Multimodal Large Language Models, providing essential knowledge infrastructure for post-event analysis and urban flood risk assessment.
Yang et al. (Mon,) studied this question.