ELLIOT-X-EXT

Derived and contextual layers for Major TOM ELLIOT-Pretrain, row-aligned with it. ELLIOT-Pretrain ships Sentinel-2, Landsat-8/9, Copernicus DEM and ESA WorldCover for 279,166 tiles of 10.56 × 10.56 km on the Major TOM grid. This extension adds, for the same tiles and in the same order:

• a cloud and shadow mask (OmniCloudMask; clear, thick cloud, thin cloud, shadow) for every Sentinel-2 and Landsat-8 acquisition, with its fractions;
• ERA5 weather at every acquisition (2 m temperature and dewpoint, relative humidity, precipitation, cloud cover, 10 m wind);
• one Sentinel-1 RTC image (VV, VH; linear gamma-0) per Sentinel-2 frame, acquired as close as possible to that frame's date, on ELLIOT's own 10 m grid;
• the tile's OpenStreetMap features and administrative units as GeoParquet (osm.parquet, admin.parquet), clipped to the tile;
• per tile: acquisition geometry, the ESA WorldCover histogram and Copernicus DEM statistics.

Sample i of a part here extends sample i of the same part of ELLIOT-Pretrain (elliot:data_index); majortom:code_10km is unique within a part.

Part	Tiles	Frames per tile	Zip parts	Size
monotemporal	250,000	1	50	1,066 GB
monthly	12,500	12, one per calendar month, not necessarily from the same year (a multi-year climatology)	30	628 GB
burst	16,666	6 (about five-day cadence)	20	419 GB

Each part is a TACO v3 container in zip parts of about 20 GB, with a .tacocat/ catalog that opens them as one dataset.

Reading it

git clone https://github.com/OscarPellicer/taco
pip install -e "taco/python[ml]"     # builds the native reader: needs CMake, Ninja and a C++20 compiler

from huggingface_hub import snapshot_download
from taco.ml import Dataset

root = snapshot_download("isp-uv-es/elliot-x-ext", repo_type="dataset", allow_patterns=["burst/*"])
ds = Dataset(f"{root}/burst")
sample = ds[0]

The majortom-elliot-tasks package reads ELLIOT-Pretrain and this extension together, builds a measured fact sheet per tile and generates vision-language tasks from it (captions, grounding, grids, phenology and change questions), with scorers and a PyTorch loader.

Corrected ELLIOT-Pretrain metadata

elliot-pretrain-metadata-overlay.zip holds corrected metadata for ELLIOT-Pretrain's three parts (COLLECTION.json and METADATA/*.parquet; no pixels). The release on Source Cooperative currently lacks the ml:contract and carries frame rows that do not match their tiles in the burst part. Reading the release together with this overlay gives the correct per-frame dates and geometry; majortom-elliot-tasks applies it automatically. It will be withdrawn once the release itself is corrected; the package README lists exactly what that correction involves.

Licence

CC-BY-SA-4.0, except osm.parquet and admin.parquet, which are extracts of OpenStreetMap under the Open Database License (ODbL-1.0) and stay under it.

Required credits:

• © OpenStreetMap contributors (ODbL-1.0)
• Contains modified Copernicus Sentinel data
• Contains modified Copernicus Climate Change Service information (ERA5)
• Sentinel-1 RTC: Catalyst, as published on Microsoft Planetary Computer (sentinel-1-rtc, CC-BY-4.0)
• ESA WorldCover (CC-BY-4.0) and Copernicus DEM, through Major TOM ELLIOT-Pretrain (CC-BY-SA-4.0)
• Cloud masks computed with OmniCloudMask

Citation

Please cite Major TOM, on which ELLIOT-Pretrain and this extension are built:

@inproceedings{francis2024majortom,
  title={Major TOM: Expandable Datasets for Earth Observation},
  author={Francis, Alistair and Czerkawski, Mikolaj},
  booktitle={IGARSS 2024},
  pages={2935--2940},
  year={2024},
  doi={10.1109/IGARSS53475.2024.10640760}
}

Acknowledgements

Built by the Image and Signal Processing Group (ISP), Universitat de València, within the ELLIOT project, funded by the European Commission under Horizon Europe (grant 101214398).