Pixel-Level Earthquake Damage Segmentation from Social Media Images

Containing 3,266 ground-level images from nine major earthquakes between 2008 and 2023, each manually annotated at the pixel level. It provides a resource for semantic segmentation of infrastructure damage into five classes: Undamaged Building, Damaged Building, Destroyed Building, Undamaged Road, and Damaged Road.

Use Cases

• Train semantic segmentation models like Mask Transformers to classify pixels into the five damage classes (Undamaged Building, Damaged Building, etc.) from social media images.
• Develop domain-adaptive segmentation models for disaster response using imagery from nine different earthquake events across various geographic regions.
• Benchmark computer vision architectures on fine-grained damage assessment tasks using the provided train, validation, and test subsets.
• Analyze the distribution and co-occurrence of damage severity classes (e.g., Damaged Building vs. Destroyed Building) within ground-level scenes.

Strengths

• Large-scale collection of 3,266 annotated images focused on a specific, high-impact application.
• Pixel-level annotations for five distinct semantic classes enable fine-grained computer vision tasks.
• Images sourced from nine major earthquake events between 2008 and 2023, providing temporal and geographic diversity.
• Annotation process achieved an inter-annotator agreement above seventy percent, indicating reliable labels.

Limitations

• Annotations are derived from social media images, which may introduce perspective, lighting, and composition biases not found in aerial imagery.
• The five-class taxonomy may not capture all nuances of infrastructure damage present in real-world scenarios.
• Reliance on non-expert annotators, despite a rigorous protocol, could introduce label noise for complex damage states.

Provenance

Source

Images collected from social media platforms following nine major earthquakes.

Collection Method

Manual pixel-level annotation following a three-phase cross-disciplinary protocol with iterative reviews.

Time Range

2008 to 2023

Freshness

Data covers events up to 2023, with the dataset record last updated in November 2025.

Geography

Covers geographic regions affected by the nine sourced earthquake events.