links-ads/fmars-dataset

FMARS: Foundation Model Annotations for Remote Sensing Images

FMARS is a large-scale dataset of Very High Resolution (VHR) remote sensing images with annotations generated using Vision Foundation Models. The dataset focuses on disaster management applications and provides pre-event imagery and annotations for major crisis events worldwide from 2021 to 2023.

Dataset Features

• VHR Imagery: The dataset uses pre-event VHR satellite imagery from the Maxar Open Data Program, covering a total surface area of over 200,000 km^2.
• Automatic Annotations: Annotations are generated using a novel pipeline that combines the Segment Anything Model (SAM) and GroundingDINO to extract segmentation masks for buildings, roads, and high vegetation.
• Disaster Management Focus: The dataset is designed for use in disaster management applications such as damage assessment and risk analysis.

Annotation Pipeline

The annotation workflow uses a combination of open data sources and Vision Foundation Models:

1. Building footprints and road graphs are obtained from Microsoft's Building Footprints and Road Detection datasets and converted into prompts.
2. High vegetation bounding boxes are generated using GroundingDINO with text queries.
3. The bounding box prompts are fed into SAM to extract fine-grained segmentation masks for each category.
4. The resulting masks are stored to allow for both instance and semantic segmentation tasks.

Dataset Structure

FMARS dataset provides annotations in parquet format. The corresponding VHR images can be obtained from the Maxar Open Data Program website. We keep tha naming convention of the original files, to facilitate the match between images and labels.

For the full list of events and their details, please refer to the original paper, linked above.

Applications and Benchmarks

FMARS represents a first attempt at large-scale mapping, but it can be used to train semantic segmentation models for disaster management tasks, with precautions.

License and Citation

FMARS annotations are licensed under MIT. If you use this dataset or want to talk about it in your research, please cite the following paper:

@inproceedings{fmars2024,
  title={FMARS: Annotating Remote Sensing Images for Disaster Management using Foundation Models},
  author={Arnaudo, Edoardo and Lungo Vaschetti, Jacopo and Innocenti, Lorenzo and Barco, Luca and Lisi, Davide and Fissore, Vanina and Rossi, Claudio},
  booktitle={IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium},
  year={2024},
  organization={IEEE}
}