README.md

---
license: cc-by-4.0
size_categories:
- 10M<n<100M
task_categories:
- image-classification
- image-feature-extraction
pretty_name: Copernicus-Pretrain
tags:
- earth-observation
- remote-sensing
- foundation-model
- pretrain
- self-supervised-learning
- sentinel
library_name: datasets
---

# Dataset Card for Copernicus-Pretrain

[Paper](https://arxiv.org/abs/2503.11849) | [Repository](https://github.com/zhu-xlab/Copernicus-FM)

<!-- Provide a quick summary of the dataset. -->

Copernicus-Pretrain is a large-scale EO pretraining dataset with 18.7M aligned images covering all major Sentinel missions (S1,2,3,5P).

*Officially named **Copernicus-Pretrain**, also referred to as SSL4EO-S ("S" means Sentinel), as an extension of [SSL4EO-S12](https://github.com/zhu-xlab/SSL4EO-S12) to the whole Sentinel series.*

## Dataset Details

<!-- Provide a longer summary of what this dataset is. -->

Copernicus-Pretrain contains 18.7M aligned imagery from all major Sentinel missions in operation (Sentinel-1 SAR, Sentinel-2 multispectral reflectance, Sentinel-3 multispectral radiance, and Sentinel-5P atmospheric variables), as well as an elevation product Copernicus DEM GLO-30.
The images are organized into ~310K regional grids (0.25°x0.25°, consistent with ERA5), densely covering the whole land surface and near-land ocean with eight distinct Sentinel modalities.

|     |  Modality            | GSD   | Image size            | # Grid cells | # Patches | # Timestamps | # Total images |
|-----------|----------------------|-------|-----------------------|--------------|-----------|--------------|----------------|
| Sentinel-1 GRD  | SAR   | 10 m  | 264×264×2  | 247,723     | 1,067,267 | ~4          | 4,227,387     |
| Sentinel-2 TOA  | MS    | 10 m  | 264×264×13 | 247,723     | 1,067,267 | ~4          | 4,218,065     |
| Sentinel-3 OLCI | MS    | 300 m | 96×96×21   | 281,375     | 281,375   | ~8          | 2,189,561     |
| Sentinel-5P CO  | atmos.| 1 km  | 28×28      | 306,097     | 306,097   | 1–12        | 2,104,735     |
| Sentinel-5P NO2 | atmos.| 1 km  | 28×28      | 291,449     | 291,449   | 1–12        | 1,752,558     |
| Sentinel-5P SO2 | atmos.| 1 km  | 28×28      | 262,259     | 262,259   | 1–12        | 1,366,452     |
| Sentinel-5P O3  | atmos.| 1 km  | 28×28      | 306,218     | 306,218   | 1–12        | 2,556,631     |
| Copernicus DEM  | elevation | 30 m | 960×960   | 297,665     | 297,665   | 1           | 297,665       |
| **Copernicus-Pretrain** |  |  |  | **312,567** | **3,879,597** |  | **18,713,054** |

## Sample Usage

You can load the dataset using the Hugging Face `datasets` library. This dataset is very large and may require specific handling such as streaming or selecting specific configurations if available.

```python
from datasets import load_dataset

# Load the dataset. For large datasets, consider streaming or specific data_files if available.
# This dataset offers raw GeoTiff and streaming WebDataset formats.
dataset = load_dataset("wangyi111/Copernicus-Pretrain")

# Print the dataset structure (e.g., available splits)
print(dataset)

# Example of accessing a sample from a split (uncomment and adjust if applicable)
# For example, if 'train' split exists:
# print(dataset["train"][0])
```

## License

CC-BY-4.0.

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

```
@misc{wang2025unifiedcopernicusfoundationmodel,
      title={Towards a Unified Copernicus Foundation Model for Earth Vision}, 
      author={Yi Wang and Zhitong Xiong and Chenying Liu and Adam J. Stewart and Thomas Dujardin and Nikolaos Ioannis Bountos and Angelos Zavras and Franziska Gerken and Ioannis Papoutsis and Laura Leal-Taixé and Xiao Xiang Zhu},
      year={2025},
      eprint={2503.11849},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.11849}, 
}
```