Machine learning-based assessment of building evacuation vulnerability at the pre-disaster stage

Natural disasters frequently threaten human life worldwide, and evacuation is a key strategy to reduce casualties. However, evacuation-based vulnerability assessments remain limited and often lack practical guidance. This study proposes a temporal–spatial evacuation vulnerability assessment framework using machine learning–estimated travel time. The framework was applied to three Japanese cities—Adachi, Katori, and Takahashi—to evaluate its generalizability. Among the seven machine learning algorithms tested, XGBoost with optimal feature selection performed best, with over 98 % of predicted travel times falling within a 5-minute error margin and mean absolute errors (MAE) kept below 1 minute across all cities. The results demonstrate that the proposed approach effectively identifies spatial distribution and different levels of evacuation vulnerability. This framework offers practical insights for prioritizing high-vulnerability areas, improving evacuation preparedness, and supporting sustainable disaster risk reduction strategies in urban environments.