AEF Mosaic

A global mosaic of Google AlphaEarth Foundations Satellite Embeddings, covering 2017-2025 at ~0.00009° resolution. The dataset is projected in EPSG:4326 with a fixed pixel size in degrees. This means that pixels cover ~10m on the ground at the equator, but less area at higher latitudes. For area-based analysis, consider reprojecting to an equal-area projection, creating your own mosaic in a custom CRS, or using the underlying COGs which are already projected to UTM. This copy is not officially supported by Google.

This dataset implements the following Zarr conventions. These Zarr conventions are not mature and may change.

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."

Visualizations

RGB Composite

PCA visualization of the embedding space projected to RGB:

RGB PCA (Zoomed)

Detailed view showing embedding structure at higher resolution:

Cosine Similarity

Similarity search using a query embedding:

Change Detection

Temporal embedding differences highlighting landscape changes:

Quick Start

1import zarr
2import numpy as np
3                                                                                                                            
4root = zarr.open_group(                                                                                                       
5  "s3://us-west-2.opendata.source.coop/tge-labs/aef-mosaic/",                                                               
6  mode="r",                                                                                                                 
7  storage_options={"anon": True}                                                                                            
8)                                                                                                                             
9embeddings = root["embeddings"]
10
11# Access a spatial subset (year 0, all bands, y slice, x slice)
12data = embeddings[0, :, 516000:516100, 832000:832100]
13
14# De-quantize to float32
15data_float = ((data / 127.5) ** 2) * np.sign(data)
1import zarr
2import numpy as np
3                                                                                                                            
4root = zarr.open_group(                                                                                                       
5  "s3://us-west-2.opendata.source.coop/tge-labs/aef-mosaic/",                                                               
6  mode="r",                                                                                                                 
7  storage_options={"anon": True}                                                                                            
8)                                                                                                                             
9embeddings = root["embeddings"]
10
11# Access a spatial subset (year 0, all bands, y slice, x slice)
12data = embeddings[0, :, 516000:516100, 832000:832100]
13
14# De-quantize to float32
15data_float = ((data / 127.5) ** 2) * np.sign(data)

Xarray

1import xarray as xr
2
3
4def fetch_point(ds, lat, lon, time=0):
5    """Fetch AEF embedding for a single lat/lon point. Returns shape (64,)."""
6    return ds.isel(time=time).sel(x=lon, y=lat, method="nearest")
7
8
9def fetch_bbox(ds, bbox, time=0):
10    """Fetch AEF embeddings for a bounding box. Returns shape (64, y, x)."""
11    minx, miny, maxx, maxy = bbox
12    return ds.isel(time=time).sel(x=slice(minx, maxx), y=slice(maxy, miny))
13
14
15def main():
16    ds = xr.open_zarr(
17        "s3://us-west-2.opendata.source.coop/tge-labs/aef-mosaic/",
18        storage_options={"anon": True}
19    )
20
21    # Fetch embedding for a single point
22    lat, lon = 34.0522, -118.2437  # Los Angeles
23    point_data = fetch_point(ds, lat, lon).compute()
24    print(f"Point shape: {point_data.embeddings.shape}")
25
26    # Fetch embeddings for a bounding box
27    bbox = [-118.279668, 34.036242, -118.20862, 34.089643]
28    bbox_data = fetch_bbox(ds, bbox).compute()
29    print(f"Bbox shape: {bbox_data.embeddings.shape}")
30
31
32if __name__ == "__main__":
33    main()
1import xarray as xr
2
3
4def fetch_point(ds, lat, lon, time=0):
5    """Fetch AEF embedding for a single lat/lon point. Returns shape (64,)."""
6    return ds.isel(time=time).sel(x=lon, y=lat, method="nearest")
7
8
9def fetch_bbox(ds, bbox, time=0):
10    """Fetch AEF embeddings for a bounding box. Returns shape (64, y, x)."""
11    minx, miny, maxx, maxy = bbox
12    return ds.isel(time=time).sel(x=slice(minx, maxx), y=slice(maxy, miny))
13
14
15def main():
16    ds = xr.open_zarr(
17        "s3://us-west-2.opendata.source.coop/tge-labs/aef-mosaic/",
18        storage_options={"anon": True}
19    )
20
21    # Fetch embedding for a single point
22    lat, lon = 34.0522, -118.2437  # Los Angeles
23    point_data = fetch_point(ds, lat, lon).compute()
24    print(f"Point shape: {point_data.embeddings.shape}")
25
26    # Fetch embeddings for a bounding box
27    bbox = [-118.279668, 34.036242, -118.20862, 34.089643]
28    bbox_data = fetch_bbox(ds, bbox).compute()
29    print(f"Bbox shape: {bbox_data.embeddings.shape}")
30
31
32if __name__ == "__main__":
33    main()

Zarr-python

1import asyncio
2from affine import Affine
3from obstore.store import S3Store
4from zarr.storage import ObjectStore
5from zarr.api.asynchronous import open_group
6
7
8store = S3Store(
9    "us-west-2.opendata.source.coop",
10    prefix="tge-labs/aef-mosaic/",
11    region="us-west-2",
12    skip_signature=True,
13)
14zarr_store = ObjectStore(store, read_only=True)
15
16
17def latlon_to_rowcol(lat, lon, transform):
18    """Convert lat/lon to row/col using the inverse transform."""
19    inv = ~transform
20    col, row = inv * (lon, lat)
21    return int(row), int(col)
22
23
24def bbox_to_slices(bbox, transform):
25    """Convert a bounding box to row/col slices."""
26    minx, miny, maxx, maxy = bbox
27    inv = ~transform
28    col1, row1 = inv * (minx, maxy)
29    col2, row2 = inv * (maxx, miny)
30    row_start, row_stop = int(min(row1, row2)), int(max(row1, row2)) + 1
31    col_start, col_stop = int(min(col1, col2)), int(max(col1, col2)) + 1
32    return slice(row_start, row_stop), slice(col_start, col_stop)
33
34
35async def fetch_point(embeddings, transform, lat, lon, year=0):
36    """Fetch AEF embedding for a single lat/lon point. Returns shape (64,)."""
37    row, col = latlon_to_rowcol(lat, lon, transform)
38    return await embeddings.getitem((year, slice(None), row, col))
39
40
41async def fetch_bbox(embeddings, transform, bbox, year=0):
42    """Fetch AEF embeddings for a bounding box. Returns shape (64, rows, cols)."""
43    row_slice, col_slice = bbox_to_slices(bbox, transform)
44    return await embeddings.getitem((year, slice(None), row_slice, col_slice))
45
46
47async def main():
48    root = await open_group(store=zarr_store, mode="r")
49    embeddings = await root.get("embeddings")
50    transform = Affine(*root.metadata.attributes["spatial:transform"])
51
52    # Fetch embedding for a single point
53    lat, lon = 34.0522, -118.2437  # Los Angeles
54    point_data = await fetch_point(embeddings, transform, lat, lon)
55    print(f"Point shape: {point_data.shape}")
56
57    # Fetch embeddings for a bounding box
58    bbox = [-118.279668, 34.036242, -118.20862, 34.089643]
59    bbox_data = await fetch_bbox(embeddings, transform, bbox)
60    print(f"Bbox shape: {bbox_data.shape}")
61
62
63if __name__ == "__main__":
64    asyncio.run(main())
1import asyncio
2from affine import Affine
3from obstore.store import S3Store
4from zarr.storage import ObjectStore
5from zarr.api.asynchronous import open_group
6
7
8store = S3Store(
9    "us-west-2.opendata.source.coop",
10    prefix="tge-labs/aef-mosaic/",
11    region="us-west-2",
12    skip_signature=True,
13)
14zarr_store = ObjectStore(store, read_only=True)
15
16
17def latlon_to_rowcol(lat, lon, transform):
18    """Convert lat/lon to row/col using the inverse transform."""
19    inv = ~transform
20    col, row = inv * (lon, lat)
21    return int(row), int(col)
22
23
24def bbox_to_slices(bbox, transform):
25    """Convert a bounding box to row/col slices."""
26    minx, miny, maxx, maxy = bbox
27    inv = ~transform
28    col1, row1 = inv * (minx, maxy)
29    col2, row2 = inv * (maxx, miny)
30    row_start, row_stop = int(min(row1, row2)), int(max(row1, row2)) + 1
31    col_start, col_stop = int(min(col1, col2)), int(max(col1, col2)) + 1
32    return slice(row_start, row_stop), slice(col_start, col_stop)
33
34
35async def fetch_point(embeddings, transform, lat, lon, year=0):
36    """Fetch AEF embedding for a single lat/lon point. Returns shape (64,)."""
37    row, col = latlon_to_rowcol(lat, lon, transform)
38    return await embeddings.getitem((year, slice(None), row, col))
39
40
41async def fetch_bbox(embeddings, transform, bbox, year=0):
42    """Fetch AEF embeddings for a bounding box. Returns shape (64, rows, cols)."""
43    row_slice, col_slice = bbox_to_slices(bbox, transform)
44    return await embeddings.getitem((year, slice(None), row_slice, col_slice))
45
46
47async def main():
48    root = await open_group(store=zarr_store, mode="r")
49    embeddings = await root.get("embeddings")
50    transform = Affine(*root.metadata.attributes["spatial:transform"])
51
52    # Fetch embedding for a single point
53    lat, lon = 34.0522, -118.2437  # Los Angeles
54    point_data = await fetch_point(embeddings, transform, lat, lon)
55    print(f"Point shape: {point_data.shape}")
56
57    # Fetch embeddings for a bounding box
58    bbox = [-118.279668, 34.036242, -118.20862, 34.089643]
59    bbox_data = await fetch_bbox(embeddings, transform, bbox)
60    print(f"Bbox shape: {bbox_data.shape}")
61
62
63if __name__ == "__main__":
64    asyncio.run(main())

Dataset Details

Source Code

The processing pipeline used to create this mosaic is available at: https://github.com/geospatial-jeff/aef-mosaic

Contact

Please send any comments, questions, or points of feedback to the #opening-aef CNG slack channel.