Multimodal models for assessing earthquake-induced building damage using metadata and satellite imagery
Journal
Journal of Building Engineering
Journal Volume
111
Start Page
113467
ISSN
2352-7102
Date Issued
2025-10
Author(s)
Kuo, Wen-Ni
Abstract
The accurate assessment of building damage following earthquakes is crucial for effective disaster response and recovery efforts. In place of on-site manual damage investigation, many automated methods and artificial intelligence models were developed for large-scale disaster damage assessment to enhance the efficiency and accuracy of disaster response efforts. Common single-modal models, which depend only on either metadata or imagery, often fail to comprehensively capture the complexity of post-disaster scenarios, leading to suboptimal predictive accuracy. This study develops a multimodal model that integrates satellite imagery with detailed building metadata to enhance the assessment process. By leveraging the strengths of both data types, the model provides a more robust and nuanced understanding of earthquake-induced damage. Testing on datasets from the 2010 Haiti and 2015 Nepal earthquakes demonstrates that the multimodal approach outperforms traditional methods, offering improved accuracy and reliability in damage prediction, with an F1-score increase from 0.368 (metadata-only) and 0.348 (image-only) to 0.555 (multimodal). Overall, this research underscores the potential of multimodal frameworks to enhance post-disaster evaluations, providing valuable insights for more effective disaster management.
Subjects
Building metadata
Machine learning
Multimodal models
Post-earthquake building damage assessment
Satellite imagery
Publisher
Elsevier BV
Type
journal article