[4N5-GS-10g] AI application

3:30 PM - 3:45 PM JST(6:30 AM - 6:45 AM UTC)

[4N5-GS-10g-01] Feasibility of Seismic Intensity Estimation from Earthquake Damage Photographs Using a Multimodal Large Language Model

〇Hisahiko Kubo¹, Masashi Shiomitsu² (1. National Research Institute for Earth Science and Disaster Resilience, 2. Yamagata University)

Keywords:

Earthquake Damage Photograph,Multimodal LLM,Seismic Intensity

This study investigated the feasibility of estimating Japan Meteorological Agency seismic intensity scales from earthquake-damage photographs using a multimodal large language model (LLM). When using a single photograph, the LLM could generally recognize the damage status and produce reasonable estimates corresponding to the recognized damage. Approximately half of the estimates fell within ±1 scale of the observed values. However, there was an overall tendency to underestimate, likely because a single photograph captures only a limited portion of the total damage. Incorporating multiple photographs taken at nearby locations reduced this underestimation, although accurately identifying seismic intensity 7 remained challenging.