Earthquake Infrastructure Damage Images from Social Media

Comprising 13,513 high-quality images from five significant earthquakes, labeled for infrastructure damage assessment. It uses a 4-class classification system aligned with established damage scales like HAZUS and EMS-98, with labeling guidelines that improved annotator agreement from 39.7% to 10.4% disagreement.

Use Cases

• Train a 4-class damage classifier using the HAZUS/EMS-98 aligned labels on 13,513 social media images.
• Benchmark computer vision models for disaster reconnaissance using the multi-class classification system.
• Analyze the impact of structured labeling guidelines on annotator agreement metrics within disaster image datasets.

Strengths

• 13,513 high-quality images provide a substantial sample for model training.
• Labeling guidelines reduced annotator disagreement from 39.7% to 10.4%.
• Images are sourced from five significant earthquakes, offering event diversity.
• Classification system is aligned with established damage scales (HAZUS, EMS-98).

Limitations

• Images are compiled from other social media databases, which may introduce source bias or duplication.
• Limited to five earthquake events, which may not represent all damage scenarios.
• Focus is on infrastructure damage, excluding other disaster impacts like human casualties.

Provenance

Source

Compiled from several social media image databases.

Collection Method

Images were labeled with a novel 4-class system using comprehensive guidelines to enhance quality and annotator agreement.

Freshness

Last updated February 2025.

Geography

Covers five significant earthquake events (specific locations not listed).