Can Large Language Models Quantify Floods from Crowd-Sourced Data?

Abstract Submission: Flash floods in urban areas occur due to extensive impermeable surfaces and intense rainfall. Accurate observations of flood depths are essential for quantifying and validating this flood risk. Crowd-sourced data, including citizen and government reports, social media, and news outlets, contain valuable information about flood inundation and could contribute to such a database. They remain underutilized due to their unstructured nature. Large Language Models (LLMs) have recently gained prominence in processing and extracting information from unstructured data. In this study, we explore the potential of LLMs to quantify and curate flash flood-related inundation depths from crowd-sourced textual data. The National Weather Service damage reports, which contain descriptive event narratives that indirectly reference inundation depths in New York City (NYC), were used as the benchmark. These narratives pose challenges for conventional methods, which struggle to extract reliable inundation information efficiently. In this context, we attempted to address the following research questions: 1) Can LLMs effectively translate flash flood reports into quantitative flood depths, and how skillful are these estimations? 2) What are the best LLM adaptation techniques that can lead to accurate information extraction comparable to ground truths? Preliminary results from Chat-GPT for NYC indicate that LLMs can appropriately quantify flood inundation depths when the textual information includes explicit inundation measurement or relevant keywords. However, the model occasionally produced inaccurate results, such as misinterpreting unrelated numbers as inundation depths in narratives requiring deeper inference. The findings of this study offer insights for improving strategies in applying LLMs for natural hazards engineering.

Learning Objectives/Expected Outcome (Optional) :