SSL4EO-S12 v1.1
March 24, 2026 ยท View on GitHub
SSL4EO-S12 v1.1
SSL4EO-S12 v1.1 is an updated multimodal version of the popular EO pre-training dataset SSL4EO-S12. Read more about the reasons behind our update and further improvements in our technical report on arXiv.
NEWS
- Mar 24, 2026: We provided metadata files to better search the samples.
- Feb 19, 2026: We fixed a temporal alignment issue in the Sentinel-1 GRD data which was not sorted by date (see issue). Thank you, Thomas, for reporting the issue!
- Feb 17, 2026: SSL4EO-S12 v1.1 is now available as a webdataset version for better usability at HuggingFace.
- Mar 11, 2025: SSL4EO-S12 v1.1 available as a Zarr chunk file version on HuggingFace.
- Mar 10, 2025: SSL4EO-S12 v1.1 utilized as pre-training dataset for 2025 CVPR EARTHVISION data challenge:
Data
The dataset includes 246,144 locations with four timestamps each from the modalities S2L1C, S2L2A, S1GRD, S2RGB, NDVI, LULC, and a single timestamp DEM. We refer to our technical report for details.
Sentinel-2 and Sentinel-1 time series examples with four seasonal images:
Modality examples in SSL4EO-S12 v1.1:
Download
You can download the dataset via the Hugging Face CLI (pip install huggingface_hub). Please note that the full dataset requires 2.3TB of storage.
hf download embed2scale/SSL4EO-S12-v1.1 --repo-type dataset --local-dir data/SSL4EOS12
If you like to download only a subset of the data, you can specify it with --include.
# Only download val data
hf download embed2scale/SSL4EO-S12-v1.1 --repo-type dataset --include "val/*" --local-dir data/SSL4EOS12
# Only download a single modality (e.g., S2L2A)
hf download embed2scale/SSL4EO-S12-v1.1 --repo-type dataset --include "*/S2L2A/*" --local-dir data/SSL4EOS12
For development, webdataset supports data streaming and does not need any local data.
Usage
Set up your env with
pip install -r requirements.txt
# or install the packages manually via
pip install huggingface_hub webdataset zarr==2.18.0 numcodecs==0.15.1 torch numpy albumentations fsspec braceexpand
We provide code for a PyTorch dataloader in ssl4eos12_dataset.py which you can download with
wget https://raw.githubusercontent.com/DLR-MF-DAS/SSL4EO-S12-v1.1/refs/heads/main/ssl4eos12_dataset.py
You can use the build_ssl4eos12_dataset function to initialize a dataset, which uses the WebDataset package to load samples from the shard files. You can stream the data from Hugging Face using the urls or download the full dataset and pass a local path (e.g, data/SSL4EOS12/).
from ssl4eos12_dataset import build_ssl4eos12_dataset
from torch.utils.data import DataLoader
# If you only pass one modality, the modality is loaded with the "image" key
dataset = build_ssl4eos12_dataset(
path="https://huggingface.co/datasets/embed2scale/SSL4EO-S12-v1.1/resolve/main/", # Streaming or local path
modalities=["S2L2A"],
split="val",
batch_size=8
)
# Batch keys: ["__key__", "__url__", "image"]
# If you pass multiple modalities, the modalities are returned using the modality names as keys
dataset = build_ssl4eos12_dataset(
path="https://huggingface.co/datasets/embed2scale/SSL4EO-S12-v1.1/resolve/main/", # Streaming or local path
modalities=["S2L2A", "S2L1C", "S2RGB", "S1GRD", "DEM", "NDVI", "LULC"],
split="val",
batch_size=8,
)
# Set batch size to None because batching is handled by WebDataset.
dataloader = DataLoader(dataset, batch_size=None, num_workers=4, persistent_workers=True, prefetch_factor=1)
# Iterate over the dataloader
for batch in dataloader:
print("Batch keys:", list(batch.keys()))
# Batch keys: ["__key__", "__url__", "S2L2A", "S2L1C", "S2RGB", "S1GRD", "DEM", "NDVI", "LULC"]
print("Data shape:", batch["S2L2A"].shape)
# Data shape: torch.Size([8, 4, 12, 264, 264]
# Dimensions [batch, time, channel, h, w]
break
The data in SSL4EO-S12 v1.1 is sorted by date, meaning that each timestep can be from any season.
If you like to load the data with fixed seasons, pass reindex_seasonal=True to build_ssl4eos12_dataset() and the loaded data is sorted by season while ignoring the year.
I.e., the first timestamp is from the first yearly quartal, followed by the second and so on.
Data transform
We provide some additional code for wrapping albumentations transform functions.
We recommend albumentations because parameters are shared between all image modalities (e.g., same random crop).
However, it requires some code wrapping to bring the data into the expected shape.
import albumentations as A
from albumentations.pytorch import ToTensorV2
from ssl4eos12_dataset import (build_ssl4eos12_dataset, Transpose, MultimodalTransforms, MultimodalNormalize,
FlattenTemporalIntoChannels, UnflattenTemporalFromChannels, statistics)
# Define all image modalities
modalities = ["S2L2A", "S2L1C", "S2RGB", "S1GRD", "DEM", "NDVI", "LULC"]
# Define multimodal transform function that converts the data into the expected shape from albumentations
val_transform = MultimodalTransforms(
transforms=A.Compose([ # We use albumentations because of the shared transform between image modalities
Transpose([0, 2, 3, 1]), # Convert data to channel last (expected shape from albumentations)
MultimodalNormalize(mean=statistics["mean"], std=statistics["std"]),
# CenterCrop other transformations cannot handle temporal data. Needs to be applied after MultimodalNormalize
FlattenTemporalIntoChannels(),
A.CenterCrop(224, 224), # Use center crop in val split
# A.RandomCrop(224, 224), # Use random crop in train split
# A.D4(), # Optionally, use random flipping and rotation for the train split
ToTensorV2(), # Convert to tensor and back to channel first
UnflattenTemporalFromChannels(n_timesteps=4), # Add time dim back, apply after ToTensorV2()
],
is_check_shapes=False, # Not needed because of aligned data in TerraMesh
additional_targets={m: "image" for m in modalities}
),
non_image_modalities=["__key__", "__url__"], # Additional non-image keys
)
dataset = build_ssl4eos12_dataset(
path="https://huggingface.co/datasets/embed2scale/SSL4EO-S12-v1.1/resolve/main/",
modalities=modalities,
split="val",
transform=val_transform,
batch_size=8,
)
If you only use a single modality, you don't need to specify additional_targets but you need to change the normalization to:
` MultimodalNormalize(
mean={"image": statistics["mean"]["<modality>"]},
std={"image": statistics["std"]["<modality>"]}
),`
Returning metadata
You can pass return_metadata=True to build_ssl4eos12_dataset() to load center longitude and latitude, timestamps, and the S2 cloud mask as additional metadata.
The resulting batch keys include: ["__key__", "__url__", "S2L2A", "S1GRD", ..., "center_lon", "center_lat", "cloud_mask", "time_S2L2A", "time_S1GRD", ...].
If you are using the cloud_mask, update additional_targets in your transform:
val_transform = MultimodalTransforms(
transforms=A.Compose([...],
additional_targets={m: "image" for m in modalities + ["cloud_mask"]}
# additional_targets={"cloud_mask": "image"} # Setting for a single modality dataset
),
)
Note that center points are not updated when random crop is used. The cloud mask provides the classes land (0), water (1), snow (2), thin cloud (3), thick cloud (4), cloud shadow (5), and no data (6). DEM does not return a time value while LULC uses the S2 timestamp because of the augmentation using the S2 cloud and ice mask. Time values are returned as integer values but can be converted back to datetime with
batch["time_S2L2A"].numpy().astype("datetime64[ns]")
Metadata
We provide metadata Parquet files per split at Hugging Face with the following columns:
tar ssl4eos12_shard_000001.tar
zarr ssl4eos12_val_seasonal_data_0000001.zarr.zip
sample_id 0080717
split val
center_lon 125.637763
center_lat 42.659342
crs 32651
bounds [714870.0, 4725040.0, 717510.0, 4727680.0]
geometry POLYGON ((125.65334657340223 42.64709631345207...
S2_time_0 2020-11-16 02:30:01
S2_time_1 2021-03-04 02:36:09
S2_time_2 2021-05-18 02:35:51
S2_time_3 2021-08-08 02:25:49
S1_time_0 2020-11-27 21:46:46
S1_time_1 2021-02-26 21:38:39
S1_time_2 2021-05-26 21:46:46
S1_time_3 2021-08-30 21:46:51
cloud_cover_0 0.0
cloud_cover_1 0.0
cloud_cover_2 0.0
cloud_cover_3 0.0011
The sample_id is similar to the v1 dataset. The S2_time_{i} are shared between S2L1C, S2L2A, S2RGB, and NDVI. LULC is augmented with the clouds and ice/snow at that timestamp.
Zarr chunk file version
We released a previous version of SSL4EO-S12 v1.1 using Zarr chunk files with 64 samples each. The version ist still available at embed2scale/SSL4EO-S12-v1.1-Zarr. We moved on to a webdataset version for better usability.
zarr_dataset.py provides data loading code for previous version and the chunk file version is directly compatible with TerraTorch's GenericMultiModalDataModule which is showcased in the config terratorch_zarr_ssl4eos12.yaml.
License
This repository is released under the Apache 2.0 license. The dataset is released under the CC-BY-4.0 license.
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 Ait Ali Braham, Nassim 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 Ait Ali Braham, Nassim 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}
}