SD4EO: Sentinel-2 Northern France Patch Dataset (RGB+NIR)

This dataset has been created as part of the deliverables for ESA’s SD4EO project.

It consists of orthogonal patches from real satellite images of Sentinel-2, covering the visible and near-infrared bands, taken over large areas in northern France. The following tiles are included:

Tile ID

Sampling Date

Number of patches

T31TCJ

2021/04/15

3306

T31TCN

2021/04/25

3120

T31TDN

2021/04/25

3481

T31UDP

2021/04/25

3540

T31UEP

2021/04/25

3599

T30UWU

2023/11/07

2928

T31TCL

2024/05/09

3422

T31TDL

2024/05/09

3306

T31TDM

2021/03/31

2891

T31TEM

2021/04/27

3480

T31TEN

2021/04/27

3038

T31TFM

2021/04/27

2576

T31UDQ

2023/10/07

3599

T31UEQ

2023/10/07

3540

T30TXT

2023/10/10

3420

T30TYS

2023/10/10

3534

T30UXU

2023/10/10

3599

T30UYU

2023/10/10

3654

T31TFN

2024/05/11

3363

T31UFP

2024/05/11

3658

T31UFQ

2024/05/11

3660

T30TVK

2024/05/11

2970

TOTAL

73,684

The tiles were selected to minimize cloud cover, which is why the sampling dates are spread across a wide range. Additionally, care was taken to capture images close to winter and late spring, ensuring that lighting conditions were either at dawn or dusk, which helps reduce saturation effects on highly reflective surfaces such as flat industrial rooftops.

Each patch covers an area of approximately 1700x1700 meters (though the exact size may vary slightly with latitude). Patches located at the edges of tiles were excluded to avoid potential cutting or distortion issues. Likewise, patches near large bodies of water, including coastlines, were also discarded.

It's important to note that, unlike the synthetic image dataset [link], in this case, the patches are always disjoint.

To facilitate the use of these images for training generative AI models (such as GANs and Diffusion models), the patch size has been standardized to 512x512 pixels, providing an effective resolution of 3.3 meters per pixel. The color channels were scaled according to the accumulated histograms in order to embrace most of the energy (>90%): Thus, the values from 0 to 256 in the red channel correspond to the range of 460.0 to 6700.0 in the irradiance captured by Sentinel-2 for the patches in this dataset. The values in the green channel correspond to a range of 740.0 to 5800.0, and the blue channel values span from 400.0 to 4840.0 as recorded by Sentinel-2 sensors.

To simplify their usage, the images are stored in lossless PNG format.

Each patch is also accompanied by pixel-level labels based on the color-coded annotations from OpenStreetMap data in the dataset doi:10.5281/zenodo.13958096 Both this dataset and the OpenStreetMap-labeled dataset have been used to train a conditional diffusion model, which generated a synthetic dataset of 46.5 GB that can be accessed at this [link].

The file naming convention is straightforward:

                                          Patch____S2.png

Columns and rows are always referenced within the same tile.

The images from Sentinel satellites, part of the European Union's Copernicus program, are available under an open access policy. Specifically, they are distributed under the Creative Commons Attribution-ShareAlike 3.0 IGO (CC BY-SA 3.0 IGO) license. This means that you are free to share, use, and adapt the images, even for commercial purposes, as long as appropriate credit is given, and any derivative work is shared under the same license: Open Access at ESA. This open access policy allows wide usage for research, education, and even commercial applications, with the goal of supporting environmental monitoring and other societal needs. So, we have selected the closest avaliable licence in Zenodo, as this dataset is a derivative work.

The SD4EO Project is funded by ESA’s FutureEO programme under contract no. 4000142334/23/I-DT and is supervised by ESA Φ-lab.