Can large language models quantify urban floods from multi-source textual information

Multi-source textual flood information remains underused because its unstructured format limits quantitative analysis. This study proposes DeepFlood, a large language model (LLM)-based framework for estimating flash flood inundation depths from event narratives through reference-based inference. DeepFlood was evaluated using multiple LLM backbones, including ChatGPT and Gemini, to assess its robustness. For flash flood narratives from the Greater New York Metropolitan Area, the best-performing configuration achieved a flood-depth category accuracy of 95.5% and a reference-based reasoning accuracy of 92.4% compared to domain expert estimates. Expert-based estimation showed a 17.9% mismatch rate in reference-ID selection, whereas DeepFlood performance across training data conditions had coefficients of variation of 0.5–0.7% for flood-depth category accuracy and 0.6–1.5% for reference-based reasoning accuracy. DeepFlood demonstrates that LLMs can transform underused textual flood narratives into quantitative inundation information, supporting post-event assessment and urban flood model validation.