OpenEarthMap-SAR

The OpenEarthMap Synthetic Aperture Radar (OpenEarthMap-SAR) is a high-resolution SAR benchmark dataset for land cover mapping under all-weather conditions. The motivation of this benchmark dataset is to facilitate advancements in SAR-based geospatial analysis for global high-resolution land cover mapping.

The OpenEarthMap-SAR dataset was served as the official dataset for the 2025 IEEE GRSS Data Fusion Contest Track 1 organized by the IEEE GRSS Image Analysis and Data Fusion Technical Committee, the University of Tokyo, RIKEN, and ETH Zurich. The contest aims to foster the development of innovative solutions for all-weather land cover and building damage mapping using multimodal SAR and optical EO data at submeter resolution. Check out the winners of the contest at here.

Dataset

The OpenEarthMap-SAR dataset consists of 1.5 million segments of 5033 aerial and satellite images covering 35 regions from Japan, France and the USA; and with partially manually annotated labels and fully pseudo labels of 8 land cover classes. Each image has a size of 1024x1024 pixels at a ground sampling distance of 0.15m--0.5m. The dataset has been made publicly available at Zenodo, where you can download it. Below are examples of the dataset.

Baseline Models

The baseline models for three different tasks (semantic segmentation, UDA, and image translation) on the OpenEarthMap-SAR benchmark dataset are provided as follows.

Semantic Segmentation

Yon can download all the pre-trained basleine models for the semantic segmentation task from here.

Method	Modality	Labelling Scenario	mIoU	Pretrained
U-Net with EfficientNet-B4	Optical	Pseudo labels	56.56	Download
		Pseudo labels + 5 real labels per region	58.19	Download
		Only 5 real labels per region	65.10	Download
	SAR	Pseudo labels	35.13	Download
		Pseudo labels + 5 real labels per region	36.84	Download
		Only 5 real labels per region	33.86	Download
	SAR + Optical	Pseudo labels	56.32	Download
		Pseudo labels + 5 real labels per region	49.79	Download
		Only 5 real labels per region	60.16	Download

The code for the U-Net with EfficientNet-B4 is borrowed from the Segmentation Models Pytorch. Thanks to the authors for making their code publicly available.

 

UDA Semantic Segmentation

Country-wise UDA task with France as the source domain and Japan and USA as target domains. Yon can download all the pre-trained basleine models for the UDA semantic segmentation task from here.

Method	Modality	UDA Setup	Labelling Scenario	mIoU	Pretrained
DAFormer	Optical	France-to-USA	Pseudo labels	39.38	Download
			Pseudo labels + 5 real labels per region	39.60	Download
		France-to-Japan	Pseudo labels	44.35	Download
			Pseudo labels + 5 real labels per region	43.74	Download
	SAR	France-to-USA	Pseudo labels	19.35	Download
			Pseudo labels + 5 real labels per region	21.26	Download
		France-to-Japan	Pseudo labels	13.49	Download
			Pseudo labels + 5 real labels per region	12.50	Download
	SAR + Optical	France-to-USA	Pseudo labels	40.45	Download
			Pseudo labels + 5 real labels per region	37.21	Download
		France-to-Japan	Pseudo labels	40.03	Download
			Pseudo labels + 5 real labels per region	39.97	Download

The code for the UDA semantic segmentation task is borrowed from the DAFormer. Thanks to the authors for making their code publicly available.

 

Image Translation

Yon can download all the pre-trained basleine models for the image translation task from here.

Method	Task	Sub-Task	Modality	PSNR	SSIM	Pretrained
CUT	RGB Generation	Pseudo Semantic2RGB	Pseudo Semantic Map	4.62	0.2799	Download
		Real Semantic2RGB	Real Semantic Map	14.57	0.2141	Download
		SAR2RGB	SAR	12.47	0.1075	Download
	SAR Generation	Pseudo Semantic2SAR	Pseudo Semantic Map	9.48	0.0213	Download
		Real Semantic2SAR	Real Semantic Map	9.27	0.0239	Download
		RGB2SAR	RGB	9.18	0.0292	Download

The code for the image translation task is borrowed from the Contrastive Unpaired Translation (CUT). Thanks to the authors for making their code publicly available.

 

Citation

For any scientific publication using this data, the following paper should be cited:

@article{xia2025openearthmap,
  title   = {OpenEarthMap-SAR: A Benchmark Synthetic Aperture Radar Dataset for Global High-Resolution Land Cover Mapping},
  author  = {Xia, Junshi and Chen, Hongruixuan and Broni-Bediako, Clifford and Wei, Yimin and Song, Jian and Yokoya, Naoto},
  journal = {arXiv preprint arXiv:2501.10891},
  year    = {2025}
}

Acknowledgements

We are most grateful to the authors of the following works for making their source codes publicly available:

• Semantic Segmentation PyTorch
• DAFormer
• CDAC
• BicycleGAN
• CUT
• CycleGAN-Pix2Pix
• Decent
• NICE-GAN
• SPADE
• UNSB