terrafloww/aef-v1-annual-rasteret

AEF v1 Annual Rasteret Collection

This dataset is a read-ready Rasteret collection derived from the public AlphaEarth Foundations (AEF) annual index and enriched with per-band Cloud Optimized GeoTIFF header metadata.

It is intended for Rasteret users who want to skip the build/enrich step and start directly from a published narrow index plus read-ready Rasteret collection shards.

What this dataset contains

Each row corresponds to one AEF annual tile and includes:

• index.parquet: a narrow GeoParquet 1.1 index with fid, crs, path, year, utm_zone, bbox_utm, bbox, geom, and location
• data/part-xxxxx.parquet: read-ready Rasteret collection shards with canonical runtime columns id, datetime, geometry, bbox, assets, proj:epsg, year, utm_zone, plus A00_metadata through A63_metadata

This dataset does not store embedding pixel arrays inside Parquet. It stores metadata that lets Rasteret fetch pixels directly from the original COGs.

Why this exists

The source AEF STAC-geoparquet is already public, but Rasteret normally needs one more step: COG header enrichment. This collection publishes that work once so users can:

• filter remotely before reading any pixels
• join the collection to their own labels, manifests, splits, or experiment metadata
• reuse the same read-ready index across notebooks, jobs, and teams

License and attribution

This derived metadata index is licensed under CC-BY-SA-4.0. It contains no pixel data — only Parquet metadata and COG headers of the imagery.

This dataset was built from the public AlphaEarth Foundations Satellite Embedding Dataset published on Source Cooperative by Taylor Geospatial Engine Labs:

• Source Cooperative product page: https://source.coop/tge-labs/aef
• Source Cooperative AEF index used as metadata input: https://source.coop/tge-labs/aef/v1/annual/aef_index.parquet

The underlying AlphaEarth Foundations Satellite Embedding data is produced by Google and Google DeepMind and is licensed under CC-BY-4.0. This Terrafloww dataset is derived from the Source Cooperative AEF index and the Source Cooperative-hosted AEF COGs.

When using this derived index or the underlying embeddings, include the required upstream attribution text:

"The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind."

Reference links:

• Google Earth Engine dataset page: https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL
• Google AEF GCS README: https://developers.google.com/earth-engine/guides/aef_on_gcs_readme
• Taylor Geospatial Engine article on the Source Cooperative release: https://tgengine.org/building-frictionless-geospatial-ai-making-alphaearth-foundations-embeddings-accessible/
• This index license (CC-BY-SA-4.0): https://creativecommons.org/licenses/by-sa/4.0/
• Underlying data license (CC-BY-4.0): https://creativecommons.org/licenses/by/4.0/

Citation

Use this citation for the derived Terrafloww dataset:

@dataset{terrafloww_aef_v1_annual_rasteret_2026,
  author    = {Terrafloww Labs, Inc.},
  title     = {AEF v1 Annual Rasteret Collection},
  year      = {2026},
  publisher = {Terrafloww Labs, Inc.},
  url       = {https://huggingface.co/datasets/terrafloww/aef-v1-annual-rasteret},
  note      = {Derived Rasteret metadata index built from the Source Cooperative AEF index and Source Cooperative-hosted AlphaEarth Foundations COGs published by Taylor Geospatial Engine Labs}
}

Also acknowledge the upstream data source in accompanying text:

• Source Cooperative AEF product published by Taylor Geospatial Engine Labs: https://source.coop/tge-labs/aef
• AlphaEarth Foundations data producer: Google and Google DeepMind
• Required attribution: "The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind."

Recommended access patterns

Rasteret (simplest)

import rasteret

collection = rasteret.load("hf://datasets/terrafloww/aef-v1-annual-rasteret/data")

# Filter to your area and time range
sub = collection.subset(
    bbox=(-95.5, 38.4, -95.0, 38.8),
    date_range=("2024-01-01", "2024-12-31"),
)

# Read pixels — choose your output format
ds = sub.get_xarray(geometries=(-95.4, 38.5, -95.1, 38.7), bands=["A00", "A01"], geometry_crs=4326)

Data layout

• index.parquet: narrow GeoParquet 1.1 metadata index for remote filtering
• data/part-xxxxx.parquet: read-ready Rasteret collection shards
• rows are clustered by year, utm_zone, and WGS84 spatial order
• Parquet layout is tuned for repeated filtered reads rather than maximum full-scan throughput

Intended use

This dataset is for metadata filtering, selection, and read-time handoff into Rasteret. It is especially useful when you want to combine AEF tiles with your own Arrow-native metadata workflows before reading embeddings.

Limitations

• It depends on the upstream AEF COG URLs remaining available.
• It is optimized for Rasteret and Arrow-native tooling, not for storing embedding payloads in Parquet.
• It is a derived metadata index, not a replacement for the upstream source dataset.