Using large language models for early flood analysis : a case study

Abstract

Floods, intricate in causation and profound in local impact, necessitate early and intelligent responses. This thesis introduces an early flood warning system leveraging the capabilities of Large Language Models (LLMs). The system extracts insights from textual information, encompassing factors like flood warnings, evacuation orders and casualties. Using a case study approach, the paper explores flash floods in the UK and Scotland and floods in Japan, all of which occurred in 2023, by drawing data from social media platform X (formerly Twitter). The integration of Retrieval Augmented Generation (RAG) augments LLM knowledge with supplementary text data from the sources.

This thesis presents a theoretical foundation for LLMs, and the various components involved in RAG. Additionally, it offers a concise overview of LLM application development utilizing the Langchain framework. This is followed by a review of existing techniques for flood analysis using social media data.

Separate evaluations are conducted for the two case studies. The evaluations for the case study involving UK and Scotland floods, demonstrate high-quality results, successfully identifying close to 90% flood-prone locations under certain experimental settings. Conversely, the results for the Japanese flood are less satisfactory, highlighting an opportunity to explore alternative techniques for improvement. Additionally, a summary of utilizing LLMs as evaluators is presented.