SSL4EO-S12-downstream

Welcome to the SSL4EO-S12-downstream dataset. This dataset is used in the Embed2Scale Challenge.

SSL4EO-S12-downstream is a Earth Observation (EO) dataset of downstream tasks. It is released as a standalone dataset together with the NeuCo-Bench neural compression benchmarking framework. Parts of the SSL4EO-S12-downstream dataset was used in the 2025 CVPR EarthVision data challenge and the dev and eval phases of the challenge can be recreated. For instructions to recreate the phases, please see the section "Recreate 2025 CVPR EarthVision data challenge".

The dataset is structure as follows:

• data contains image data from three EO modalities: Sentinel-1, Sentinel-2 L1C and Sentinel-2 L2A. For details on the contents of the data, please see section "Image data". There is one file for each modality per location. In each file, there are four time instances at this location.
• labels contains downstream task target labels in the form of csv files. Each csv file correpond to one task and contains a mapping between the image file name (column id) and the target label (column label). The downstream tasks may use a subset of the available image data file. Please see the section "Downstream tasks" for information on the tasks.

Image data

The dataset constitutes 13260 datacubes with a size of approximately 150 GB. The structure of the datacubes is visualized below.

The image data is organized under the data folder, which contains one subfolder per modality (s1, s2l1c and s2l2a). Within the modality subfolders the data is distributed over a set of parts, subfolders named part-XXXXXX, each containing 1000 samples. The part with the same name across the modalities contains identically named files, corresponding to the same location. Each datacube constitute one location, with S1 polarisations, S2 L1C and S2 L2A channels.

Each location is sampled at four times, one during the months March-May, one during June-August, one during September-November and finally one during months December-February, in this order.

The datacubes are stored in zipped zarr files; see here for instructions how to load the data. The data in the zarr files is structured as (number of locations, number of timestamps, number of channels, heigh, width) with the dimensions (1, 4, 27, 264, 264); the 27 channels coming from 2 S1 polarizations (VV and VH), 13 S2 L1C channels (B1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B10, B11, B12), and 12 S2 L2A channels (B1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, B12).

In addition to the full dataset, we provide an example version, containing the first part, i.e. the first 1000 samples of each modality.

The data is structured identically to the SSL4EOS12 v1.1 dataset:

@article{blumenstiel2025ssl4eos12,
  title={{SSL4EOS12 v1.1 - A Multimodal, Multiseasonal Dataset for Pretraining}},
  author={Blumenstiel, Benedikt and Braham, Nassim Ait Ali and Albrecht, Conrad M and Maurogiovanni, Stefano and Fraccaro, Paolo},
  journal={arXiv preprint arXiv:2503.00168},
  year={2025}
}

Downstream tasks

Currently, 11 downstream tasks are available

• biomass_mean__regr: Biomass mean across all pixels, with labels retrieved from GEDI.
• biomass_std__regr: Biomass standard deviation across pixels, same source as above.
• crops__regr: Combined fraction of Soybean and Corn in image, with labels retrieved from CDL.
• landcover_aggriculture_regr: Fraction of agriculture in image, with the labels retrieved from the Corine land cover dataset.
• landcover_forest__regr: Fraction of forest in image, same source as above.
• clouds_reg__regr: Average cloud cover percentage across the four seasons. Labels retrieved from CloudSen12. This task was not part of the Embed2Scale data challenge.
• heatisland_mean__regr: Mean summertime surface temperature in larger cities across the northern hemisphere. Labels estimated from LandSat-8 thermal bands.
• heatisland_std__regr: Summertime surface temperature standard deviation of the same locations as above.
• nodata__regr: The fraction of zero pizels in the Sentinel-2 image. Based on all 13260 currently available image data.
• random_reg__reg: Random task with majority zero entries. The samples are the same as from the clouds_reg__reg task and a minor set of the labels are also the same.
• random_cls__cls: Binary classification version of the random tasks where the positive class is given to samples with regression targets greater than 10.

Recreate 2025 CVPR EarthVision data challenge

To recreate the challenge phases, the downstream task label files used in the competition have two additional columns, in addition to the id and label columns. The two boolean columns cvpr_earthvision_phase_eval and cvpr_earthvision_phase_dev indicates which samples were included in which phase. Alternatively, use the commit https://huggingface.co/datasets/embed2scale/SSL4EO-S12-downstream/tree/5c5539acd6f42e2e3547ec816fcf577e3f6800fa to download the dataset in the state used in the competition.