Identification of tropical cyclone–related flash floods from hazard narratives using a large language model–based approach
摘要
This study employs a Large Language Model (LLM)–based approach to identify tropical cyclones (TCs)-related flash floods and understand their interrelationships using hazard narratives from the National Centers for Environmental Information Storm Events Database. A processing pipeline comprising two LLMs and a series of validation procedures was developed to extract the contributing factors of flash floods and classify them classified as either TC-related factors or external weather and environmental conditions. The analysis identified 6,470 flash floods across the contiguous United States related to 135 North Atlantic and East Pacific TCs between 2007 and 2024. The spatial proximity of flash floods to TC tracks varied substantially, suggesting multiple TC-related flood-generating mechanisms. The contributing factors to flash floods extracted from hazard narratives indicated that floods primarily driven by TC processes occurred closer to the TCs, whereas events influenced by external factors extended well beyond the typical TC rainfall range.