From headlines to databases: leveraging LLMs for structured disaster event extraction

Abstract Disasters, especially those driven by climate change, are putting growing pressure on cities, threatening both their economic stability and social fabric. To build more resilient communities, we need to develop mitigation and preparedness plans. The effectiveness of these plans highly depends on understanding and accessing accurate and comprehensive disaster data. While databases like the Emergency Events Database (EM-DAT) provide useful information, they often exclude smaller, local disasters that do not meet their reporting thresholds. To address this gap, this paper presents a flexible workflow that leverages a large language model (LLM)-driven web scraping tool to collect data on underreported disaster events systematically. The workflow was tested through a case study focused on flood-related incidents in Granada (Spain), demonstrating its ability to capture relevant information. The results indicate strong potential for supporting more granular disaster damage assessments, with the system achieving an F-score of 76% in identifying requested information. This approach gives communities and policymakers a more complete picture of disaster risks, helping them make better decisions about preparedness and mitigation efforts.