Andrew Kagan
Research Mentor: Xin Wei
Mentor Department: Michigan Institute for Data and AI in Society (MIDAS), Other
Author(s): Not Available
Session: Session 7 (4:00 PM – 4:50 PM)
Presentation Type: Poster 76
Abstract
While landslides are pervasive across the U.S., their specific social, economic, and ecological impacts remain underappreciated due to their localized and episodic nature, co-occurrence with other hazards, and variability across broad regions. Efforts to systematically quantify landslide impacts have often been limited to localized case studies, carried out by numerous agencies with differing standards, and constrained to limited time frames due to labor-intensive data collection. This lack of standardized, comprehensive reporting limits the ability of researchers, policymakers, and emergency managers to fully understand past impacts and to better anticipate future risk. To address these gaps, this project develops a generative AI-based workflow for scalable landslide impact documentation and synthesis: (i) we verify and refine the USGS National Landslide Damages and Losses (NLDL) database through multi-LLM cross-checking with human review to identify and correct inconsistencies; (ii) we evaluate and summarize failure modes in long-document extraction, including output truncation, missing landslide records, and hallucinations, particularly when each event requires extraction of many structured fields; and (iii) we develop a preliminary agentic framework that extracts impact information from heterogeneous sources (PDFs and web pages), spanning traditional materials such as government reports and journal articles and non-traditional sources such as news coverage and NGO reports, and compare performance with manual extraction and direct LLM prompting. Overall, this study presents a transferable framework for leveraging generative AI to improve the reliability and efficiency of disaster impact data extraction from unstructured sources, with a current focus on landslides and potential extension to other natural hazards.
