embed2scale/SSL4EO-S12-v1.1-Zarr

SSL4EO-S12-v1.1 Zarr Chunk Files

This is the Zarr Chunk file version of SSL4EO-S12-v1.1. You find a webdataset version at embed2scale/SSL4EO-S12-v1.1 which allows for data streaming and is easier to handle.

Important: This initial Zarr Chunk file version includes Sentinel-2 L1C and L2A data with a +1000 offset to align with newest S-2 processing version from ESA. This offset is removed in the webdataset version to improve usability.

The dataset includes 246,144 locations with four timestamps each from the modalities S2L1C, S2L2A, S1GRD, S2RGB, and NDVI. We refer to our technical report for details.

The samples are stored in 3,846 Zarr Zip files (zarr version 2) that enable efficient storage and data loading. The metadata is stored directly with the samples as additional data variables. Each Zarr files contains 64 samples (unique locations) with four timestamps each. The timestamps are chunked separately, which enables efficient loading of single timestamps.

You can read a Zarr file with:

import xarray as xr
ds = xr.open_zarr('filename.zarr.zip')  # load xarray dataset
data = ds.bands.values  # load numpy array with dims [B, T, C, H, W]

Zarr was recently updated to version 3 which might lead to errors. You can fix easily with pip install zarr==2.18.0 numcodecs==0.15.3.

Example of a S2L2A xarray dataset:

<xarray.Dataset> Size: 446MB
Dimensions:     (band: 12, sample: 64, time: 4, y: 264, x: 264)
Coordinates:
  * band        (band) <U3 144B 'B01' 'B02' 'B03' 'B04' ... 'B09' 'B11' 'B12'
  * sample      (sample) <U7 2kB '0080717' '0060573' ... '0179869' '0012333'
  * time        (time) int64 32B 0 1 2 3
  * x           (x) int64 2kB 0 1 2 3 4 5 6 7 ... 257 258 259 260 261 262 263
  * y           (y) int64 2kB 0 1 2 3 4 5 6 7 ... 257 258 259 260 261 262 263
Data variables:
    bands       (sample, time, band, y, x) int16 428MB 1463 1457 ... 1777 1673
    center_lat  (sample) float64 512B 42.66 -30.64 50.47 ... 27.29 -23.06 29.99
    center_lon  (sample) float64 512B 125.6 121.4 128.3 ... -104.8 43.64 48.18
    cloud_mask  (sample, time, y, x) uint8 18MB 0 0 0 0 0 0 0 ... 0 0 0 0 0 0 0
    crs         (sample) int64 512B 32651 32751 32652 ... 32613 32738 32639
    file_id     (sample, time) <U38 39kB '20201116T023001_20201116T022955_T51...
    sample_id   (sample, time) <U9 9kB '0080717_0' '0080717_1' ... '0012333_3'
    time_       (sample, time) datetime64[ns] 2kB 2020-11-16T02:30:01 ... 202...
    x_          (sample, x) float64 135kB 7.149e+05 7.149e+05 ... 2.289e+05
    y_          (sample, y) float64 135kB 4.728e+06 4.728e+06 ... 3.319e+06

Example of a S1GRD xarray dataset:

<xarray.Dataset> Size: 72MB
Dimensions:     (band: 2, sample: 64, time: 4, y: 264, x: 264)
Coordinates:
  * band        (band) <U2 16B 'vv' 'vh'
  * sample      (sample) <U7 2kB '0080717' '0060573' ... '0179869' '0012333'
  * time        (time) int64 32B 0 1 2 3
  * x           (x) int64 2kB 0 1 2 3 4 5 6 7 ... 257 258 259 260 261 262 263
  * y           (y) int64 2kB 0 1 2 3 4 5 6 7 ... 257 258 259 260 261 262 263
Data variables:
    bands       (sample, time, band, y, x) float16 71MB -15.13 -17.44 ... -29.0
    center_lat  (sample) float64 512B 42.66 -30.64 50.47 ... 27.29 -23.06 29.99
    center_lon  (sample) float64 512B 125.6 121.4 128.3 ... -104.8 43.64 48.18
    crs         (sample) int64 512B 32651 32751 32652 ... 32613 32738 32639
    file_id     (sample, time) <U67 69kB 'S1B_IW_GRDH_1SDV_20201127T214646_20...
    sample_id   (sample, time) <U9 9kB '0080717_0' '0080717_1' ... '0012333_3'
    time_       (sample, time) datetime64[ns] 2kB 2020-11-27T21:46:46 ... 202...
    x_          (sample, x) float64 135kB 7.149e+05 7.149e+05 ... 2.289e+05
    y_          (sample, y) float64 135kB 4.728e+06 4.728e+06 ... 3.319e+06

Download

You can also download the dataset from HuggingFace to your local data/ folder via various ways.

Python API

import huggingface_hub as hf # pip install huggingface_hub, if not installed

hf.snapshot_download(
  repo_id="embed2scale/SSL4EO-S12-v1.1", 
  allow_patterns=["*.txt", "*.zarr.zip"], 
  repo_type="dataset", 
  local_dir="data/SSL4EO-S12-v1.1", 
  local_dir_use_symlinks="auto"
)

Git LFS

1. Install git (see this page) and git lfs (see this page).
2. Run the following commands in a terminal from within your local folder:

mkdir data && cd data
git lfs install
git clone https://huggingface.co/datasets/embed2scale/SSL4EO-S12-v1.1

Julich DataHub

You can download the data from Julich DataHub with the following script:

# Download all data
wget --recursive --no-parent --reject "index.html*" --exclude-directories=ssl4eo-s12/raw --execute robots=off -nH -P data https://datapub.fz-juelich.de/ssl4eo-s12/

The script will download the data to your data/ folder with the following format:

data/
└── ssl4eo-s12
    ├── splits
    │   ├── ssl4eos12_train.txt
    │   └── ssl4eos12_val.txt
    ├── train
    │   ├── S1GRD
    │   │   ├── ssl4eos12_train_seasonal_data_000001.zarr.zip
    │   │   ├── ssl4eos12_train_seasonal_data_000002.zarr.zip
    │   │   ├── ...
    │   │   └── ssl4eos12_train_seasonal_data_003812.zarr.zip
    │   ├── S2L1C
    │   ├── S2L2A
    │   └── S2RGB
    └── val
        ├── S1GRD
        │   ├── ssl4eos12_val_seasonal_data_000001.zarr.zip
        │   ├── ssl4eos12_val_seasonal_data_000002.zarr.zip
        │   ├── ...
        │   └── ssl4eos12_val_seasonal_data_000034.zarr.zip
        ├── S2L1C
        ├── S2L2A
        └── S2RGB

You can specify a subdirectory for downloading a subset of the data:

# Download validation data
wget --recursive --no-parent --reject "index.html*" --execute robots=off -nH -P data https://datapub.fz-juelich.de/ssl4eo-s12/val/
# Download S2L2A validation data
wget --recursive --no-parent --reject "index.html*" --execute robots=off -nH -P data https://datapub.fz-juelich.de/ssl4eo-s12/val/S2L2A/

If you like to access the raw tif files, you can download tar archives for each modality (~1TB per file) from https://datapub.fz-juelich.de/ssl4eo-s12/raw/.

# Download raw tif files in tar archives
wget --recursive --no-parent --reject "index.html*" --execute robots=off -nH -P data https://datapub.fz-juelich.de/ssl4eo-s12/raw/

Usage

We provide code for a PyTorch dataset in zarr_dataset.py. You can initialize a data loader with the following code:

from torch.utils.data import DataLoader
from torchvision import transforms
from zarr_dataset import ZarrDataset, collate_fn, S2L1C_MEAN, S2L1C_STD, S2L2A_MEAN, S2L2A_STD, S1GRD_MEAN, S1GRD_STD

# We concatenate the modalities for the transform function. Depending on the parameter concat=True/False, 
# the data is returned as a concatenated tensor or split into the single modalities after the transform.
train_transform = transforms.Compose([
    transforms.RandomCrop(224),  # The data has size 264x264. We recommend RandomCrop for train and CenterCrop for val.
    transforms.Normalize(mean=S2L1C_MEAN + S2L2A_MEAN + S1GRD_MEAN, std=S2L1C_STD + S2L2A_STD + S1GRD_STD)
    # Data is loaded as torch Tensor, so no ToTensor() needed.
])

train_dataset = ZarrDataset(
    data_dir='data/ssl4eo-s12/train',
    split_file='data/ssl4eo-s12/splits/ssl4eos12_train.txt',  # optional, speeds up the initialization.
    modalities=['S2L1C', 'S2L2A', 'S1GRD'], # optional, list of modality folders.
    transform=train_transform,  # optional, torchvision transforms. Returns tensors if not provided.
    concat=False,  # Concatenate all modalities along the band dimension.
    single_timestamp=False,  # Load single timestamps rather than time series.
    num_batch_samples=64,  # optional, subsample samples in each zarr file.
)

train_loader  = DataLoader(
    dataset=train_dataset,
    batch_size=1,  # Note that each batch file contains already 64 samples!
    shuffle=True,
    collate_fn=collate_fn,  # Data needs to be concatenated along sample dimension instead of being stacked
)

Alternatively, you can use the GenericMultiModalDataModule from TerraTorch if you like to use TorchGeo or TerraTorch for your pre-training. We provide an example config here: terratorch_zarr_ssl4eos12.yaml.

Standardization values:

Attention: These S2 mean statistics include a +1000 offset to align with the newest ESA processing version. The offset was removed in the later webdataset version of SSL4EO-S12 v1.1 for better usability.

{
  "S2L1C": {
    "mean": [2607.345, 2393.068, 2320.225, 2373.963, 2562.536, 3110.071, 3392.832, 3321.154, 3583.77, 1838.712, 1021.753, 3205.112, 2545.798],
    "std": [786.523, 849.702, 875.318, 1143.578, 1126.248, 1161.98, 1273.505, 1246.79, 1342.755, 576.795, 45.626, 1340.347, 1145.036]
  },
  "S2L2A": {
    "mean": [1793.243, 1924.863, 2184.553, 2340.936, 2671.402, 3240.082, 3468.412, 3563.244, 3627.704, 3711.071, 3416.714, 2849.625],
    "std": [1160.144, 1201.092, 1219.943, 1397.225, 1400.035, 1373.136, 1429.17, 1485.025, 1447.836, 1652.703, 1471.002, 1365.307]
  },
  "S2RGB": {
    "mean": [100.708, 87.489, 61.932],
    "std": [68.550, 47.647, 40.592]
  },
  "S1GRD": {
    "mean": [-12.577, -20.265],
    "std": [5.179, 5.872]
  }
}

Citation

If you use this dataset in your work, please cite:

@article{blumenstiel2025ssl4eos12,
  title={{SSL4EO-S12} v1.1: A Multimodal, Multiseasonal Dataset for Pretraining, Updated},
  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}
}

This dataset is an updated version of:

@article{wang2022ssl4eo,
  title={{SSL4EO-S12}: A large-scale multimodal, multitemporal dataset for self-supervised learning in Earth observation [Software and Data Sets]},
  author={Wang, Yi and Braham, Nassim Ait Ali and Xiong, Zhitong and Liu, Chenying and Albrecht, Conrad M and Zhu, Xiao Xiang},
  journal={IEEE Geoscience and Remote Sensing Magazine},
  volume={11},
  number={3},
  pages={98--106},
  year={2023},
  publisher={IEEE}
}