Floods are one of the most dramatic natural events threatening human life and habitation, which make efficient and timely responding to floods highly valuable. Commonly flood monitoring is based on satellite based synthetic aperture radar detection, but effective early identification of floods is needed to direct satellite resources to the affected areas. Here, the early warning and identification of floods through real-time social media data is studied. Large language models (LLMs) are utilized as the core reasoning mechanism for the early identification of flooded areas for the early allocation of active monitoring resources. We introduce a framework that uses a chatbot application and LLMs in the early identification of floods and illustrate it with two experiments using real-world storm data. We also utilize two different LLMs to compare their performance in the task. The results highlight the potential and the limitations of leveraging LLMs for early flood warning analysis.