SEN2VENµS, a dataset for the training of Sentinel-2 super-resolution algorithms

<strong>1 Description</strong> <strong>SEN2VENµS</strong> is an open dataset for the super-resolution of Sentinel-2 images by leveraging simultaneous acquisitions with the VENµS satellite. The dataset is composed of 10m and 20m cloud-free surface reflectance patches from Sentinel-2, with their reference spatially-registered surface reflectance patches at 5 meters resolution acquired on the same day by the VENµS satellite. This dataset covers 29 locations with a total of 132 955 patches of 256x256 pixels at 5 meters resolution, and can be used for the training of super-resolution algorithms to bring spatial resolution of 8 of the Sentinel-2 bands down to 5 meters. <strong>2 Files organization</strong> The dataset is composed of separate sub-datasets, one for each site, as described in table 1. Table 1: Number of patches and pairs for each site, along with VENµS viewing zenith angle Site Number of patches Number of pairs VENµS Zenith Angle FR-LQ1 4888 18 1.795402 NARYN 3814 25 5.010906 FGMANAUS 129 4 7.232127 MAD-AMBO 1443 19 14.788115 ARM 15859 39 15.160683 BAMBENW2 9018 34 17.766533 ES-IC3XG 8823 35 18.807686 ANJI 2314 16 19.310494 ATTO 2258 9 22.048651 ESGISB-3 6057 19 23.683871 ESGISB-1 2892 13 24.561609 FR-BIL 7105 30 24.802892 K34-AMAZ 1385 21 24.982675 ESGISB-2 3067 13 26.209776 ALSACE 2654 17 26.877071 LERIDA-1 2281 6 28.524780 ESTUAMAR 912 13 28.871947 SUDOUE-5 2176 20 29.170244 KUDALIAR 7269 20 29.180855 SUDOUE-6 2435 14 29.192055 SUDOUE-4 935 7 29.516127 SUDOUE-3 5363 14 29.998115 SO1 12018 36 30.255978 SUDOUE-2 9700 27 31.295256 ES-LTERA 1701 19 31.971764 FR-LAM 7299 22 32.054056 SO2 738 22 32.218481 BENGA 5858 29 32.587334 JAM2018 2564 18 33.718953 For each site, the sub-dataset folder contains a set of files for each date, following this naming convention as the pair id: <code>{site_name}_{mgrs_tile}_{acquisition_date}</code>. For each pair, 5 files are available, as shown in table 2. Patches are encoded as ready-to-use tensors as serialized by the well known Pytorch library¹. As such they can be loaded by a simple call to the <code>torch.load()</code> function. Note that bands are separated into two groups (10m and 20m Sentinel2 bands), which leads to four separate tensor files (2 groups of bands \(\times\) source and target resolution). Tensor shape is <code>[n,c,w,h]</code> where \(n\) is the number of patches, \(c=4\) is the number of bands, \(w\) is the patch width and \(h\) is the patch height. In order to save storage space, they are encoded as 16 bits signed integers and should be converted back to floating point surface reflectance by dividing each and every value by 10 000 upon reading. Table 2: Naming convention for files associated to each pair. <code>{id}</code> is <code>{site_name}_{mgrs_tile}_{acquisition_date}</code>. File Content <code>{id}_05m_b2b3b4b8.pt</code> 5m patches (\(256\times256\) pix.) for S2 B2, B3, B4 and B8 (from VENµS) <code>{id}_10m_b2b3b4b8.pt</code> 10m patches (\(128\times128\) pix.) for S2 B2, B3, B4 and B8 (from Sentinel-2) <code>{id}_05m_b5b6b7b8a.pt</code> 5m patches (\(256\times256\) pix.) for S2 B5, B6, B7 and B8A (from VENµS) <code>{id}_20m_b5b6b7b8a.pt</code> 20m patches (\(64\times64\) pix.) for S2 B5, B6, B7 and B8A (from Sentinel-2) <code>{id}_patches.gpkg</code> GIS file with footprint of each patch Each file comes with a master <code>index.csv</code> CSV (Comma Separated Values) file, with one row for each pair sampled in the given site, and columns as described in table 3, separated with tabs. Table 3: Columns of the <code>index.csv</code> file indexing pairs for each site. For file naming conventions, refer to table 2. Column Description <code>venus_product_id</code> ID of the sampled VENµS L2A product <code>sentinel2_product_id</code> ID of the sampled Sentinel-2 L2A product <code>tensor_05m_b2b3b4b8</code> Name of the 5m tensor file for S2 B2, B3, B4 and B8 (from VENµS) <code>tensor_10m_b2b3b4b8</code> Name of the 10m tensor file for S2 B2, B3, B4 and B8 (from Sentinel-2) <code>tensor_05m_b5b6b7b8a</code> Name of the 5m tensor file for S2 B5, B6, B7 and B8A (from VENµS) <code>tensor_20m_b5b6b7b8a</code> Name of the 20m tensor file for S2 B5, B6, B7 and B8A (from Sentinel-2) <code>s2_tile</code> Sentinel-2 MGRS tile <code>vns_site</code> Name of VENµS site <code>date</code> Acquisition date as <code>YYYY-MM-DD</code> <code>venus_zenith_angle</code> VENµS zenith viewing angle in degrees <code>patches_gpkg</code> Name of the GIS file with footprint for each patch <code>nb_patches</code> Number of patches for this pair Each site folder is compressed to a different 7z file. <strong>3 Licencing</strong> <strong>3.1 Sentinel-2 patches</strong> <strong>3.1.1 Copyright</strong> Value-added data processed by CNES for the Theia data centre www.theia-land.fr using Copernicus products. The processing uses algorithms developed by Theia's Scientific Expertise Centres. Note: Copernicus Sentinel-2 Level 1C data is subject to this license: https://theia.cnes.fr/atdistrib/documents/TC_Sentinel_Data_31072014.pdf <strong>3.1.2 Licence</strong> Files <code>{id}_05m_b2b3b4b8.pt</code> and <code>{id}_05m_b5b6b7b8a.pt</code> are distributed under the the original licence of the Sentinel-2 Theia L2A products, which is the Etalab Open Licence Version 2.0 ². <strong>3.2 VENµS patches</strong> <strong>3.2.1 Copyright</strong> Value-added data processed by CNES for the Theia data centre www.theia-land.fr using VENµS satellite imagery from CNES and Israeli Space Agency. The processing uses algorithms developed by Theia's Scientific Expertise Centres. 3.2.2 <strong>Licence</strong> Files <code>{id}_05m_b2b3b4b8.pt</code> and <code>{id}_05m_b5b6b7b8a.pt</code> are distributed under the original licence of the VENµS products, which is Creative Commons BY-NC 4.0 ³. <strong>3.3 Remaining files</strong> All remaining files are distributed under the Creative Commons BY 4.0 ⁴ licence. <strong>4 Note to users</strong> Note that even if the VenµS2 dataset is sorted by sites and by pairs, we strongly encourage users to apply the full set of machine learning best practices when using it : random keeping separate pairs (or even sites) for testing purpose, and randomization of patches accross sites and pairs in the training and validation sets. <strong>5 Citing</strong> Please cite the following data paper (preprint, submitted to <em>MDPI Data</em>) and zenodo link when publishing work derived from this dataset: Michel, J.; Vinasco-Salinas, J.; Inglada, J.; Hagolle, O. SEN2VENµS, a Dataset for the Training of Sentinel-2 Super-Resolution Algorithms. <em>Data</em> <strong>2022</strong>, <em>7</em>, 96. https://doi.org/10.3390/data7070096 https://zenodo.org/deposit/6514159 <strong>Footnotes:</strong> ¹ https://pytorch.org/ ² https://theia.cnes.fr/atdistrib/documents/Licence-Theia-CNES-Sentinel-ETALAB-v2.0-en.pdf ³ https://creativecommons.org/licenses/by-nc/4.0/ ⁴ https://creativecommons.org/licenses/by/4.0/