aef-loader

Virtualizarr access for AEF embeddings as an analysis ready data cube, alongside rapid querying of the GCS and Source Coop index. 2x quicker than rioxarray for single tile downloads.

What is AEF?

Alpha Earth Foundations embeddings is a dataset produced by Google Deepmind, providing a yearly 64-channel embeddings derived from numerous satellite image sources with numerous downstream applications. The embeddings are stored as multi-band Cloud-Optimised GeoTIFFs (COGs), alongside a parquet index file.

AEF is stored by two hosts:

• Google Cloud Storage - requester pays (requires gcp_project)
• Source Cooperative - AWS hosted and free to access

More in the docs.

What does aef-loader do?

aef-loader provides two key functionalities:

1. Rapid download, and querying of indexes for source_coop + gcs with obstore and geopandas
2. Lazily load the COGs as VirtualiZarr as a datatree by UTM zone, COG headers are cached, so repeated reads are cheap(er)

As additional utilities:

• dequantize, requantize, or raw-cast (int8 → float32) the embeddings
• split the "embeddings" dataset into 64 datasets
• use odc-geobox for dask aware reprojections for creating multi-zone composites

Overview

Alpha Earth Foundations embeddings is a dataset produced by Google Deepmind, providing a yearly 64-channel embeddings derived from numerous satellite image sources with numerous downstream applications. The embeddings are stored as multi-band Cloud-Optimised GeoTIFFs (COGs).

aef-loader supports two dataset hosts, both having tradeoffs:

1. Google Cloud Storage - maintained by the Earth Engine team, more up to date but requiring authentication and "requester pays", meaning users must pay egress and other charges.
2. Source Cooperative - Hosted on AWS S3 and free to access, includes the full time range (2017-2025) Recommended

Installation

pip install aef-loader

or:

uv add aef-loader

Quick Start

import asyncio
from aef_loader import AEFIndex, VirtualTiffReader, DataSource
from aef_loader.utils import reproject_datatree
from odc.geo.geobox import GeoBox

async def main():
    # Initialize index (Source Cooperative - no auth needed)
    index = AEFIndex(source=DataSource.SOURCE_COOP)
    await index.download()
    index.load() # returns a gdf for alternative use

    # Query for tiles
    tiles = await index.query(
        bbox=(-122.5, 37.5, -122.0, 38.0),
        years=(2020, 2023),
    )

    # Load tiles organized by UTM zone
    async with VirtualTiffReader() as reader:
        tree = await reader.open_tiles_by_zone(tiles)

    # Each zone is a separate Dataset with its native CRS
    for zone in tree.children:
        ds = tree[zone].ds
        print(f"{zone}: {ds.odc.crs}, {dict(ds.sizes)}")

    # Optionally reproject all zones to a common CRS
    target = GeoBox.from_bbox(
        bbox=(-122.5, 37.5, -122.0, 38.0),
        crs="EPSG:4326",
        resolution=0.0001,
    )
    combined = reproject_datatree(tree, target)

asyncio.run(main())

Attribution and Dataset License

This dataset is licensed under CC-BY 4.0 and requires the following attribution text: "The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind."

Special notes

Thanks to Max Jones, Virtual-tiff and Virtualizarr.