Satellite Imagery and Language Dataset for Disaster Damage Assessment

DisasterM3 is a remote sensing vision-language dataset for disaster damage assessment and response. It contains 26,988 bi-temporal satellite images and 123,000 instruction-response pairs. The dataset was created by researchers including Junjue Wang and Weihao Xuan, with a paper published in 2025.

Use Cases

• Train visual question answering models using the 123k instruction pairs to answer questions about disaster impact in satellite images.
• Fine-tune image segmentation models on bi-temporal satellite imagery to identify damaged buildings and infrastructure.
• Develop multi-task models for simultaneous image classification and text generation tasks on disaster scenes.
• Benchmark change detection algorithms by comparing pre- and post-disaster satellite image pairs.
• Create instruction-following AI agents for disaster response using the structured vision-language data.

Strengths

• Contains 26,988 bi-temporal satellite image pairs for change detection.
• Includes 123,000 instruction-response pairs for vision-language training.
• Covers multiple disaster types and tasks as indicated by platform tags.

Limitations

• Specific geographic coverage and disaster event details are not provided in the input.
• The dataset size of '1Bn-10Bn' from platform tags is ambiguous and lacks precise definition.
• Class distribution and potential label noise for damage assessment are unknown.

Provenance

Source

Researchers including Junjue Wang, Weihao Xuan, and Naoto Yokoya, associated with the paper 'DisasterM3: A Remote Sensing Vision-Language Dataset for Disaster Damage Assessment and Response'.

Collection Method

Collection of bi-temporal satellite imagery and generation of instruction-response pairs, as described in the associated arXiv paper.

Freshness

Dataset was last updated on the Hugging Face platform on 2026-01-13.

Geography

Platform tags indicate a 'us' region, but full spatial coverage is unspecified.