The SAR data are obtained from the UMBRA Open Data Program (https://umbra.space/open-data/). The dataset has been created combining the WorldCover by ESA (https://esa-worldcover.org/en) with the UMBRA images. BIOME information has been extracted from the RESOLVE biome dataset (https://ecoregions.appspot.com/). Reverse geo-coding with OSM nominatim (https://nominatim.openstreetmap.org). (more details asap) Authors: Federico Ricciuti, Federico Serva, Alessandro Sebastianelli License: Same as… See the full description on the dataset page: https://huggingface.co/datasets/fedric95/umbra.

Overview:

OpenEarthMap-SAR is a benchmark synthetic aperture radar dataset, for global high-resolution land cover mapping. It consists of 5033 images covering 35 regions from Japan, France and USA, with partically manually annotated and fully pesudo 8-class land cover labels at a 0.15–0.5m ground sampling distance.

IEEE GRSS Data Fusion Contest 2025

OpenEarthmap-SAR also serves as the official dataset of IEEE GRSS DFC 2025 Track I.

Please download dfc25_track1_test.zip and unzip it. It contains test images.

Benchmark code related to the DFC 2025 can be found at this Github repo.

The official leaderboard is located on the Codalab-DFC2025-Track I page.

Paper & Reference

Details of OpenEarthMap-SAR can be refer to our paper.

If OpenEarthMap-SAR is useful to research, please kindly consider cite our paper

@misc{xia2025openearthmapsar, title={OpenEarthMap-SAR: A Benchmark Synthetic Aperture Radar Dataset for Global High-Resolution Land Cover Mapping}, author={Junshi Xia and Hongruixuan Chen and Clifford Broni-Bediako and Yimin Wei and Jian Song and Naoto Yokoya}, year={2025}, eprint={2501.10891}, archivePrefix={arXiv}, primaryClass={eess.IV}, url={https://arxiv.org/abs/2501.10891}, }

License of Track 1Optical images are provided by the National Institute of Geographic and Forest Information (IGN), France, under the CC BY 2.0 license, with contributions from the Geospatial Information Authority of Japan (GSI) and the National Agriculture Imagery Program (NAIP), USA.

SAR images are supplied by the Umbra Open Data Program under the CC BY 4.0 license.

Label datasets are shared under the same license as the original optical images, with specific terms varying by source dataset.

Data

The datasets will be distributed step-by-step at the Zenodo platform. In Track 1, we provide a multimodal dataset of optical and SAR images, consisting of approximately 4,000 aerial RGB and SAR image pairs with pseudo 8-class land cover labels. The 8 classes include bareland, rangeland, developed space, road, tree, water, agriculture land, and building. The pseudo labels are generated from pre-trained OpenEarthMap models. The images cover 38 regions in Japan, USA, and France, with a ground sampling distance (GSD) between 0.15-0.5 meters. Both the images and labels are provided in TIFF format. For training, paired optical and SAR images with pseudo labels are available. For validation and testing, only SAR images are provided for metric evaluation.

• Aerial RGB images: Aerial images are from the National Agriculture Imagery Program (NAIP), the French National Institute of Geographic and Forest Information (IGN), and the Geospatial Information Authority of Japan (GSI). Each aerial image is available as an 8-bit RGB TIFF tile, with a standard tile size of 1,024 × 1,024 pixels to match the SAR resolution.
• Umbra SAR images: The SAR images are all provided by Umbra. These SAR images are provided as 8-bit single-channel TIFF tiles, with a pixel spacing ranging from 0.15 to 0.5 meters per pixel and a title size of 1,024 × 1,024 pixels.
• Land cover labels: Pseudo land cover labels for training are generated using pre-trained OpenEarthMap models. For testing, selected areas within urban regions are manually labeled by experts to provide high-quality evaluation data. https://open-earth-map.org/assets/img/DFC2025_Track1_codalab.png" alt="https://open-earth-map.org/assets/img/DFC2025_Track1_codalab.png"> Visual examples of the training data used in Track 1. SAR image © 2024 Umbra Lab, Inc., used under CC BY 4.0 license. Optical images of the first and third columns © 2024 National Institute of Geographic and Forest Information (IGN), France, used under CC BY 2.0 license; Optical image of the second column courtesy of Geospatial Information Authority of Japan (GSI); Optical images of the fourth and fifth columns courtesy of the National Agriculture Imagery Program (NAIP), USA. ### Land Cover Classes We provide annotations with eight classes: bareland, rangeland, developed space, road, tree, water, agriculture land, and building. Their color and proportion of pixels are summarized below. https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F18690771%2F15f69349a8d4a33c0c6312d6875ab08f%2Fdesc_color.png?generation=1737791806641629&alt=media" alt="color desc"> ### License Optical images are provided by the National Institute of Geographic and Forest Information (IGN), France, under the CC BY 2.0 license, with contributions from the Geospatial Information Authority of Japan (GSI) and the National Agriculture Imagery Program (NAIP), USA.

SAR images are supplied by the Umbra Open Data Program under the CC BY 4.0 license.

Label datasets are shared under the same license as the original optical images, with specific terms varying by source dataset.

Paper & Reference

If OpenEarthMap-SAR is useful to research, please kindly consider cite the paper text @misc{xia2025openearthmapsar, title={OpenEarthMap-SAR: A Benchmark Synthetic Aperture Radar Dataset for Global High-Resolution Land Cover Mapping}, author={Junshi Xia and Hongruixuan Chen and Clifford Broni-Bediako and Yimin Wei and Jian Song and Naoto Yokoya}, year={2025}, eprint={2501.10891}, archivePrefix={arXiv}, primaryClass={eess.IV}, url={https://arxiv.org/abs/2501.10891}, }

Overview

BRIGHT is the first open-access, globally distributed, event-diverse multimodal dataset specifically curated to support AI-based disaster response. It covers five types of natural disasters and two types of man-made disasters across 14 regions worldwide, with a particular focus on developing countries. About 4,200 paired optical and SAR images containing over 380,000 building instances in BRIGHT, with a spatial resolution between 0.3 and 1 meters, provides detailed representations of individual buildings, making it ideal for precise damage assessment.

IEEE GRSS Data Fusion Contest 2025 (Closed, All Data Available)

BRIGHT also serves as the official dataset of IEEE GRSS DFC 2025 Track II. Now, DFC 25 is over. We recommend using the full version of the dataset along with the corresponding split names provided in our Github repo. Yet, we also retain the original files used in DFC 2025 for download.

1. Please download dfc25_track2_trainval.zip and unzip it. It contains training images & labels and validation images.
2. Please download dfc25_track2_test.zip and unzip it. It contains test images for the final test phase.
3. Please download dfc25_track2_val_labels.zip for validation labels, redownload dfc25_track2_test_new.zip for test images with geo-coordinates and dfc25_track2_test_labels.zip for testing labels.
4. The official leaderboard is located on the Codalab-DFC2025-Track II page.

Benchmark for multimodal disaster scenes

1. For building damage assessment, please download pre-event.zip, post-event.zip, and target.zip. Note that for the optical pre-event data in Ukraine, Myanmar, and Mexico, please follow our instructions/tutorials to download.
2. For the benchmark code and evaluation protocal for supervised building damage assessment, cross-event transfer, and unsupervised multimodal change detection, please see our Github repo. You can download our provided models' checkpoint in Zenodo repo.
3. BRIGHT supports the evaluation of Unsupervised Multimodal Image Matching (UMIM) algorithms for their performance in large-scale disaster scenarios. Please download data with the prefix "umim", such as umim_noto_earthquake.zip, and use our code to test the exsiting algorithms' performance.

Paper & Reference

Details of BRIGHT can be refer to our paper.

If BRIGHT is useful to research, please kindly consider cite our paper

@article{chen2025bright,
   title={BRIGHT: A globally distributed multimodal building damage assessment dataset with very-high-resolution for all-weather disaster response}, 
   author={Hongruixuan Chen and Jian Song and Olivier Dietrich and Clifford Broni-Bediako and Weihao Xuan and Junjue Wang and Xinlei Shao and Yimin Wei and Junshi Xia and Cuiling Lan and Konrad Schindler and Naoto Yokoya},
   journal={arXiv preprint arXiv:2501.06019},
   year={2025},
   url={https://arxiv.org/abs/2501.06019}, 
}

License

Label data of BRIGHT are provided under the same license as the optical images, which varies with different events.

With the exception of two events, Hawaii-wildfire-2023 and La Palma-volcano eruption-2021, all optical images are from Maxar Open Data Program, following CC-BY-NC-4.0 license. The optical images related to Hawaii-wildifire-2023 are from High-Resolution Orthoimagery project of NOAA Office for Coastal Management. The optical images related to La Palma-volcano eruption-2021 are from IGN (Spain) following CC-BY 4.0 license.

The SAR images of BRIGHT is provided by Capella Open Data Gallery and Umbra Space Open Data Program, following CC-BY-4.0 license.