# SSL4EO-S12 v1.1: Multimodal, Multiseasonal Sentinel-1/2 Pretraining Dataset

## Facts

- Provider: Blumenstiel, Ait Ali Braham, Albrecht, Maurogiovanni, Fraccaro (IBM Research, DLR, TUM; Embed2Scale)
- License: CC-BY-4.0 (https://github.com/DLR-MF-DAS/SSL4EO-S12-v1.1)
- Size: 3,378 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/embed2scale/SSL4EO-S12-v1.1
- Access: public, readable without an account
- Category: Earth Observation & Geospatial
- Source: https://arxiv.org/abs/2503.00168
- Page: https://scigantic.com/archive/0d77cf8f-a0c1-561e-98e4-808e255dad9e
- Keywords: Earth Observation & Geospatial, satellite imagery, remote sensing, Parquet, Hugging Face

## About

Updated release of SSL4EO-S12, the self-supervised learning dataset for Earth observation built from Sentinel-1 and Sentinel-2. It covers 246,144 locations worldwide, each sampled at four seasonal timestamps, with co-registered modalities S2L1C, S2L2A, S2RGB, S1GRD (VV and VH backscatter), land use and land cover (LULC, augmented with cloud and snow masks), a digital elevation model (DEM) and NDVI. Version 1.1 fixes alignment and metadata issues in the 2022 original, and it is one of the two sources of the TerraMesh pretraining corpus. Samples are Zarr v2 zip files with aligned metadata (cloud mask, centre coordinates, CRS, product ids), grouped into tar shards for loading with webdataset; the listing (measured 2026-11-14) shows train/ and val/ folders with one subfolder per modality plus train_metadata.parquet and val_metadata.parquet holding per-sample timestamps and cloud cover. The card gives about 2.3 TB for the full download and xarray or zarr opens single samples. A Zarr-chunk variant lives in embed2scale/SSL4EO-S12-v1.1-Zarr. The authors release the dataset under CC-BY-4.0 (code Apache-2.0); cite arXiv:2503.00168 and the original SSL4EO-S12 paper (Wang et al., IEEE GRSM 2023).

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("embed2scale/SSL4EO-S12-v1.1")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `train_metadata.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/0d77cf8f-a0c1-561e-98e4-808e255dad9e
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/0d77cf8f-a0c1-561e-98e4-808e255dad9e
- Over MCP: `get_schema_card` with id `0d77cf8f-a0c1-561e-98e4-808e255dad9e` at https://api.scigantic.com/mcp

## Related topics

- [Earth Observation & Geospatial datasets](https://scigantic.com/topics/earth-observation-and-geospatial)
- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Satellite imagery datasets](https://scigantic.com/topics/satellite-imagery)
- [Remote sensing datasets](https://scigantic.com/topics/remote-sensing)
