earth-compress/georvq-ae-v2-sample

AlphaEarth Foundations embeddings, georvq ae-v2 compressed (sample)

Six AlphaEarth Foundations annual embedding tiles (64 bands, 10 m, 8192 x 8192 px), as distributed on source.coop, compressed with the georvq ae-v2 codec. Each source tile is one Zarr store that can be read by window, locally or over HTTP, at any of 23 rates. Reading returns int8 embeddings with the source's shape, CRS, geotransform and -128 nodata.

Layout

annual/{year}/{utm_zone}/{stem}.georvq.zarr
aef_georvq_index.parquet

The tree mirrors the source bucket (tge-labs/aef/v1/annual/{year}/{utm_zone}/{stem}.tiff), and each store is a plain Zarr v3 group. The index is GeoParquet with one row per store. Its first columns are the same in every georvq release's index: the store's path relative to this directory, product_id, source_url (the source tile on source.coop), release, crs, the size and bounds in the tile's UTM zone, the footprint in longitude and latitude as geometry and bbox, the codec ids, the store's sizes, and rungs, the bytes a whole-tile read fetches at each rung. The columns of the source index (aef_index.parquet from source.coop) follow, with its path as source_path; location is copied from it and points at the source tile, not at the store here.

year	zone	stem	lat	valid	store
2018	36N	x145gbb3orzpkjlqv-0000008192-0000008192	32.2	1.00	51.6 MB
2021	35S	xawei966i326mixy3-0000008192-0000000000	-15.2	1.00	63.7 MB
2025	32N	xh72hhha3ih6mroc7-0000008192-0000008192	52.8	1.00	63.5 MB
2021	44N	x21yfwngg902u0lsh-0000008192-0000000000	17.4	0.89	57.3 MB
2021	40N	xtuw00yqbjn5szkgd-0000008192-0000008192	74.9	0.55	33.4 MB
2024	25S	xny2ld0nqj92sstbu-0000000000-0000000000	-10.0	0.51	28.3 MB

The six source tiles take 0.8 to 3.3 GB each. The records are compressed with zstd, and a window with no valid pixel is stored as zeros, which take almost no space, so a store's size follows its valid area: the two tiles near 50% valid are 28 and 33 MB and the 44N tile is 57.3 MB. The fully valid stores differ, from 51.6 to 63.7 MB, because zstd compresses the records of some tiles more than others.

Opening a store

# pip install "georvq @ git+https://github.com/earth-compress/georvq"
import georvq
url = "https://huggingface.co/datasets/earth-compress/georvq-ae-v2-sample/resolve/main/annual/2024/25S/xny2ld0nqj92sstbu-0000000000-0000000000.georvq.zarr"
with georvq.open(url) as ds:
    a = ds.read(window=((0, 1024), (0, 1024)), quality="8-8-2")   # int8 (64, 1024, 1024)

georvq.open needs nothing but the URL. On the first pixel read it downloads the decoder named in the store's header from the model repository earth-compress/georvq-ae-v2. The stores are plain Zarr v3, so zarr-python and xarray also open them, and pyarrow or DuckDB can query the index without opening any store. The layout is described in docs/GEORVQ_ZARR.md and the release in docs/AE_RELEASE.md.

Rates and quality

The 23 rungs run from 0.17 to 7.6 bits per pixel. A rung is a depth per tier (coarse-mid-fine), and a read fetches only the bytes its rung needs, which for the whole 2018 36N tile are the sizes below. The quality column is 1 - cos of the int8 reconstruction against the int8 source, averaged over the valid pixels of 16 held-out source tiles.

rung	bits/px	whole-tile read	1 - cos
2-0-0	0.169	1.30 MB	0.0266
8-0-0	0.186	1.30 MB	0.0195
8-4-0	0.389	2.1 MB	0.0087
8-8-0	0.592	3.3 MB	0.0067
8-8-2	2.342	11.2 MB	0.0036
8-8-4	4.092	23.8 MB	0.0026
8-8-8	7.592	51.4 MB	0.0019

Each store also records the measured error of every 64 px window at every rung, as the mean and 99th percentile over the window's valid pixels of the per-pixel 1 - cos between the decoded field and the source (the header's quality.name). A read can therefore ask for a bound such as quality="p99<=0.02" and receive, per window, the cheapest rung meeting it. The values are stored log-quantised, with 32 codes per decade, rounded up, so a window read to a bound that one of its rungs meets does not exceed the bound in the decode the values were measured with.

Licence and attribution

The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind. It is licensed CC-BY 4.0. This dataset is a lossy compressed derivative of it and carries the same licence and attribution, which every store's header also records under licence. The codec weights are licensed Apache 2.0.

Development

Developed by Asterisk Laboratories Co-operative Ltd. (Asterisk Labs) as part of the Earth Compress project, a partnership between Asterisk Labs and the National Oceanography Centre.

Funding

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

Citation

@misc{georvq_ae_v2_sample,
  title  = {{AlphaEarth Foundations embeddings, georvq ae-v2 compressed (sample)}},
  author = {{Asterisk Laboratories Co-operative Ltd.}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/earth-compress/georvq-ae-v2-sample},
  note   = {Earth Compress project}
}

The source data is described in Brown et al., "AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data", arXiv:2507.22291 (2025).