Presentation Information
[U15-P13]Disaster Risk Assessment for Cultural Heritage Monuments Using Integrated Damage and Restoration Records
*Hideyuki Uesugi1,2, Naoya Mitani1,2, Hajime Baba2 (1.Cultural Heritage Disaster Risk Management Center , 2.Nara National Research Institute for Cultural Properties)
Keywords:
Cultural Heritage,Disaster Risk Assessment,RAG,Historic Sites,Places of Scenic Beauty,Natural Monuments
Recent extreme rainfall and earthquakes have increased damage to protected cultural heritage monuments in Japan. Because real-time monitoring is rarely available for all sites, risk assessment and post-disaster reporting often rely on manual literature searches and expert judgement. We develop an interactive, evidence-grounded disaster risk assessment and reporting system that integrates: (i) a structured database of monument attributes, disaster events, and restoration costs; (ii) machine-learning models (e.g., decision trees and random forests) for scenario-based risk estimation by hazard intensity (continuous cumulative rainfall and seismic intensity); (iii) local explainability that ranks influential factors; (iv) retrieval from restoration reports and guidelines with page-level citations; and (v) a mitigation catalog that stores countermeasures with applicability rules and literature references. To prevent hallucinations, numeric results are strictly taken from model inference outputs, and the language model is constrained to summarize only retrieved evidence spans; countermeasures are selected from the catalog without inventing new ones. We also validate that every paragraph in the generated report has traceable citations (document title, year, and page). As a pilot corpus, we ingested 11 volumes (8 volumes of Sendai Castle Great East Japan Earthquake restoration reports, 2 volumes of Onojo site restoration reports for the July 2003 heavy rainfall, and 1 volume of the July 2020 heavy rainfall cultural property restoration record), producing 4,094 RAG chunks (heading/paragraph granularity with print-page numbers) and a mitigation catalog of 183 measures with 229 citations. The prototype can generate on-demand, evidence-backed PDF/HTML reports that summarize hazard-by-intensity risk outlooks, key factors, and prioritized mitigation options for four hazard types (heavy rain, flood, landslide, and earthquake). Future work will expand the monument master data and restoration-cost dataset and incorporate OCR for scanned documents to scale the system nationwide.