Earth Compress benchmark corpora

The 585 samples of the Earth Compress benchmark, in 27 corpora of satellite imagery, reanalysis fields, elevation, land cover and embeddings. Each sample is a window of a published product on its native grid, with bands or time steps on the first axis; ERA5 samples are whole global fields, and Clay samples are blocks of MajorTOM cells. A sample holds the published integers, with their scale and offset, when these reproduce the published values exactly, and the published floats otherwise. The compressed files take 2.6 GB.

Layout

{corpus}/manifest.csv
{corpus}/spec.md
{corpus}/npz/{sample}.npz.zst
{corpus}/png/{sample}.png

manifest.csv has one row per sample with its source asset, CRS, GDAL geotransform, start and end dates, bands, shape, dtype, unit, scale, offset, NoData and the SHA-256 of the array, which is enough to write the sample as a GeoTIFF. spec.md describes the corpus and how its samples were chosen. The PNG is a preview.

corpus	contents	shape	dtype	samples
s2-l1c-10m-4	Sentinel-2 L1C, bands B2, B3, B4, B8 at 10 m	4×512×512	uint16	50
s2-l1c-20m-6	Sentinel-2 L1C, red-edge, narrow NIR and SWIR bands at 20 m	6×512×512	uint16	50
s2-l1c-60m-3	Sentinel-2 L1C, aerosol, water vapour and cirrus bands at 60 m	3×512×512	uint16	50
lm-mss-4	Landsat 1–5 MSS, four bands at 60 m	4×512×512	uint8	15
l5-tm-6	Landsat 5 TM, six reflective bands at 30 m	6×512×512	uint8	15
l5-tm-tir	Landsat 5 TM, thermal band B6	512×512	uint8	15
l7-etm-6	Landsat 7 ETM+, six reflective bands at 30 m	6×512×512	uint8	15
l7-etm-pan	Landsat 7 ETM+, panchromatic band at 15 m	1024×1024	uint8	15
l7-etm-tir-2	Landsat 7 ETM+, thermal band at low and high gain	2×512×512	uint8	15
l8-oli-8	Landsat 8 OLI, eight reflective bands at 30 m	8×512×512	uint16	15
l8-oli-pan	Landsat 8 OLI, panchromatic band at 15 m	1024×1024	uint16	15
l8-tirs-2	Landsat 8 TIRS, thermal bands B10 and B11	2×512×512	uint16	15
goes-abi-16	GOES-16 ABI, sixteen bands as counts with per-band scale and offset	16×512×512	uint16	15
s1-grd-iw-2	Sentinel-1 GRD IW, VV and VH backscatter, linear	2×512×512	float32	50
emit-l2a-285	EMIT L2A reflectance, 285 bands at 60 m	285×128×128	float32	10
cop-dem-glo30	Copernicus DEM GLO-30 elevation	512×512	float32	50
esa-wc-v200	ESA WorldCover 2021 land-cover class codes at 10 m	512×512	uint8	50
imerg-pcp-12	GPM IMERG precipitation rate, 12 half-hourly fields	12×512×512	uint16	15
era5-msl-1h	ERA5 mean sea-level pressure, 24 hourly global fields	24×721×1440	uint16	10
era5-t2m-1h	ERA5 2 m temperature, 24 hourly global fields	24×721×1440	uint16	10
era5-tcc-1h	ERA5 total cloud cover, 24 hourly global fields	24×721×1440	uint16	10
era5-tp-1h	ERA5 total precipitation, 24 hourly global fields	24×721×1440	uint16	10
era5-u10-1h	ERA5 10 m eastward wind, 24 hourly global fields	24×721×1440	uint16	10
alphaearth-v1-64	AlphaEarth Foundations embeddings, 64 dimensions at 10 m	64×512×512	int8	15
tessera-v2-codes	TESSERA v2 embedding codes, 128 dimensions	128×512×512	int8	15
tessera-v2-scales	TESSERA v2 per-pixel scale of the codes	512×512	float32	15
clay-embeddings	Clay v1.5 Sentinel-2 embeddings, 1024 dimensions on 32×32 MajorTOM cells of 2560 m	1024×32×32	float32	15

Reading a sample

import io
import numpy as np
import zstandard

raw = open("s2-l1c-10m-4/npz/s2-l1c-10m-4-00.npz.zst", "rb").read()
sample = np.load(io.BytesIO(zstandard.decompress(raw)))
data = sample["data"]  # "scale", "offset" and "nodata" when the corpus has them

The benchmark reads the corpora from its corpora/ folder, which has the same layout as this repository.

hf download earth-compress/benchmark --repo-type dataset --local-dir corpora
uv run earthcompress corpora/s2-l1c-10m-4

Each corpus can also be rebuilt from its manifest with the build.py of its folder in the benchmark repository, which checks every sample against its SHA-256.

Licence and attribution

Each corpus keeps the licence of its source.

source	corpora	licence and attribution
Sentinel-1, Sentinel-2	s1-grd-iw-2, s2-l1c-*	Copernicus data policy. Contains modified Copernicus Sentinel data 2023.
Landsat	lm-mss-4, l5-*, l7-*, l8-*	Public domain. Landsat data courtesy of the U.S. Geological Survey.
GOES-16 ABI	goes-abi-16	Public domain. GOES-16 imagery courtesy of NOAA/NESDIS.
GPM IMERG	imerg-pcp-12	No restrictions. Global Precipitation Measurement (GPM) IMERG data provided by the NASA/GSFC Precipitation Processing System.
EMIT	emit-l2a-285	No restrictions. EMIT L2A Estimated Surface Reflectance V001, NASA.
ERA5	era5-*	Copernicus licence. Contains modified Copernicus Climate Change Service information 2023. Neither the European Commission nor ECMWF is responsible for any use of this data.
Copernicus DEM	cop-dem-glo30	Produced using Copernicus WorldDEM-30 © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA; all rights reserved.
ESA WorldCover	esa-wc-v200	CC BY 4.0. © ESA WorldCover project 2021 / Contains modified Copernicus Sentinel data (2021) processed by the ESA WorldCover consortium.
AlphaEarth Foundations	alphaearth-v1-64	CC BY 4.0. The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind.
TESSERA	tessera-v2-*	CC0 1.0.
Clay	clay-embeddings	CC BY 4.0. Clay v1.5 embeddings produced by LGND with the Clay Foundation Model.

Development

Developed in the Earth Compress project, led by Asterisk Labs and the National Oceanography Centre.

Funding

This work was funded by the Advanced Research and Invention Agency (ARIA).

Citation

@misc{earth_compress_benchmark,
  title  = {{Earth Compress benchmark corpora}},
  author = {{Asterisk Laboratories Co-operative Ltd.} and {National Oceanography Centre}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/earth-compress/benchmark},
  note   = {Earth Compress project}
}