The gs://alphaearth_foundations_custom GCS bucket contains sample datasets of
Custom Satellite Embeddings made up of COG (Cloud Optimized GeoTIFF) files
generated by AlphaEarth Foundations.
License
This collection of datasets is licensed under CC-BY 4.0 and requires the following attribution text: "Custom Satellite Embeddings from AlphaEarth Foundations are produced by Google and Google DeepMind." Use of these public datasets is subject to the Google Cloud Acceptable Use Policy.
Available datasets
The following is a list of datasets in this bucket. It contains a dataset's name, description, and general area encompassed by the dataset. Note that time intervals here use an inclusive start time and an exclusive end time.
name: corcoran_city_2026_agricultural_change
area: Corcoran and surrounding agricultural region, California, USA
description: Winter bare soil (Dec 15, 2025 - Feb 15, 2026) and spring peak greenness (Mar 15 - May 15, 2026) embeddings capturing agricultural crop dynamics over two seasonal windows.name: futaleufu_argentina_2026_fire_scar
area: Los Alerces National Park and surrounding region, Futaleufú Department, Chubut Province, Argentina
description: Pre-fire (Dec 1-31, 2025) and post-fire (Mar 1-31, 2026) embeddings capturing wildfire burn scar extent and forest canopy disturbance from the 2026 Los Alerces Wildfire.name: king_county_2025_monthly
area: King County, Washington, USA
description: Calendar-month embeddings for the King County administrative area for the first six months of 2025, covering diverse urban city infrastructure and forested land.name: meloland_california_2026_crop_cycles
area: Meloland and surrounding agricultural fields, Imperial Valley, California, USA
description: Six sequential monthly (15th to 15th) embedding time steps (mid-November 2025 through mid-May 2026) tracking multi-crop rotation cycles.name: pacific_palisades_2025_fire
area: Pacific Palisades and surrounding coastal hills, Los Angeles, California, USA
description: Pre-fire (Dec 7, 2024 - Jan 7, 2025) and post-fire (Jan 7 - Feb 7, 2025) 31-day embeddings capturing Palisades Fire wildfire impact and land cover change.name: theewaterskloof_dam_2026_water_change
area: Theewaterskloof Dam, Theewaterskloof Local Municipality, South Africa
description: Dry-season (Jan 1-31, 2026) and wet-season recharge (Apr 15 - May 15, 2026) embeddings capturing reservoir water surface area changes.
Directory structure
All sample datasets are organized under the
gs://alphaearth_foundations_custom/custom_satellite_embedding/v1/ prefix:
gs://alphaearth_foundations_custom/custom_satellite_embedding/v1/
└── {dataset_name}/
├── aef_index.parquet
├── manifest.txt
└── {start_date}_{end_date}/
└── {image_id}/
├── data.tiff
└── image_manifest.json
Each dataset has its own directory containing a dataset-level
aef_index.parquet GeoParquet index file, a dataset-level manifest.txt file,
and one or more time interval subdirectories.
These time interval subdirectories are named using the pattern
{start_date}_{end_date}, where the start date is inclusive and the end date is
exclusive. Both dates are formatted as yyyyMMdd and are in the UTC timezone.
For example, a subdirectory for the month of June 2025 would be named
20250601_20250701, which corresponds to a time interval from
2025-06-01T00:00:00Z (inclusive) to 2025-07-01T00:00:00Z (exclusive).
Each time interval directory contains a subdirectory for each image embedding
within that time interval, named with an image identifier ({image_id}) that
uniquely identifies the image within the dataset. Each {image_id} subdirectory
contains a data.tiff COG file of the pixel data and a sidecar
image_manifest.json file that is used to register the data in Earth Engine.
File structure
The internal structure of each COG file—including its 8192x8192 pixel
dimensions, 64 channels (A00 through A63), normalized overview pyramids,
signed 8-bit integer quantization, and -128 NoData mask value—is identical
to the annual Satellite Embedding COGs. Unlike the annual COGs, valid pixels in
each sample tile typically cover only a subset of the 8192x8192 grid
corresponding to the region of interest, with all remaining pixels masked as
-128 (NoData). For complete details on the COG layout, see the
File structure section
of the AlphaEarth Foundations GCS data guide.
The path of each COG file also carries some information. For example, consider
the file at
gs://alphaearth_foundations_custom/custom_satellite_embedding/v1/pacific_palisades_2025_fire/20250107_20250207/x1evicvaet1mupt13-0000008192-0000000000/data.tiff.
As evident in the path, this file is part of the embeddings for the 2025
Palisades Fire in Los Angeles County. This particular file contains embeddings
that summarize a time period from January 7th, 2025 (inclusive) to February 7th,
2025 (exclusive). The x1evicvaet1mupt13-0000008192-0000000000 subdirectory
name uniquely identifies the data.tiff COG in the
pacific_palisades_2025_fire dataset. This identifier is an opaque UUID.
Alternatively, this file metadata and more can be extracted from the COG's
image_manifest.json file found at
gs://alphaearth_foundations_custom/custom_satellite_embedding/v1/pacific_palisades_2025_fire/20250107_20250207/x1evicvaet1mupt13-0000008192-0000000000/image_manifest.json.
This file is a JSON representation of an Earth Engine Image Manifest. The JSON
structure of an Image Manifest is defined in the
Image Manifest guide. For a full
description of the contents in a given image_manifest.json file, see the
Image manifest section.
De-quantization
Raw pixel values in each channel are signed 8-bit integers in the range [-127,
127] (with -128 reserved as the "no data" value) and must be de-quantized to
produce analysis-ready floating-point values in the range [-1, 1]. For the
de-quantization mapping and instructions on creating downsampled pyramids, see
the De-quantization
and
Create downsampled pyramids
sections of the AlphaEarth Foundations GCS data guide.
Manifests and indexes
To facilitate discovering, iterating over, and registering embedding tiles, each
dataset provides two types of manifests—a dataset-level list of COG files
(manifest.txt) and a per-image metadata manifest (image_manifest.json)—along
with a dataset-level GeoParquet spatial index (aef_index.parquet).
Dataset manifest
Every dataset includes a manifest.txt file located at the root of its
directory:
gs://alphaearth_foundations_custom/custom_satellite_embedding/v1/{dataset_name}/manifest.txt
This is a plain-text file containing the full Google Cloud Storage URI for each
data.tiff file in the dataset, separated by newlines.
Image manifest
Alongside every data.tiff file is a sidecar image_manifest.json file:
gs://alphaearth_foundations_custom/custom_satellite_embedding/v1/{dataset_name}/{start_date}_{end_date}/{image_id}/image_manifest.json
This JSON file follows the Earth Engine Image Manifest specification and describes the tile's source files, band structure, temporal bounds, and spatial coverage.
While standard manifest fields (such as tilesets and bands) follow standard
Earth Engine conventions, there are several key details to note:
name: Left empty ("") in GCS. When using this manifest to register an image in an Earth Engine asset collection, the client must supply the target asset ID (e.g.,projects/{project}/assets/{collection}/{image_id}). See Load a dataset into Earth Engine for details.start_time&end_time: Specify the start (inclusive) and end (exclusive) timestamps for the composite time period in UTC formatted as ISO 8601 (e.g.,2026-03-15T00:00:00+00:00).footprint: A closed polygon in pixel coordinates outlining a region containing valid pixels. Note that missing pixels can still exist within the polygon - such is the case for complex embedding geometries like that ofking_county_2025_monthly.properties: Key-value metadata specific to the embedding, including:IMAGE_ID: The unique identifier of the image within the dataset.DATASET: The dataset name.DATASET_VERSION: The dataset version.START_DATE&END_DATE: The start (inclusive) and end (exclusive) dates of the time interval covered by the image, formatted asYYYY-MM-DD.MODEL_VERSION: The version string uniquely identifying the model version used to produce the image.PROCESSING_SOFTWARE_VERSION: The version string uniquely identifying the model data processing software used to produce the image.UTM_ZONE: The UTM zone of the coordinate reference system used to produce the image.
GeoParquet index
Every dataset includes a GeoParquet spatial index file located at the root of
its directory:
gs://alphaearth_foundations_custom/custom_satellite_embedding/v1/{dataset_name}/aef_index.parquet
This index contains one entry for each data.tiff file in the dataset and
provides spatial and temporal metadata to help determine which files cover a
given area and time period:
geometry: The geometry of the file as a WGS84 (EPSG:4326) polygon. For details on how this polygon is computed from the pixel array's UTM bounding box and clipped to UTM zone bounds, see the Geometry processing section of the AlphaEarth Foundations GCS data guide. Note that this polygon reflects the bounds of the full 8192x8192 pixel array and does not account for the validity of individual pixels.crs: The CRS of the UTM zone the image belongs to as an EPSG code, such asEPSG:32610.path: The Google Cloud Storage URI (gs://...) of thedata.tiffCOG file.dataset: The dataset name.start_date&end_date: The start (inclusive) and end (exclusive) dates of the time interval covered by the image, formatted asYYYY-MM-DD.image_id: The unique identifier of the image within the dataset.utm_zone: The UTM zone of the image, such as10N.utm_west,utm_south,utm_east,utm_north: The UTM bounds of the raw pixel array. This does not reflect any geometry processing and includes all pixels whether or not they are valid.wgs84_west,wgs84_south,wgs84_east,wgs84_north: The minimum and maximum longitude and latitude of the WGS84geometry.
Load a dataset into Earth Engine
NOTE: As a convenience, all sample datasets for Custom Satellite Embeddings will be added to Google Earth Engine as public ImageCollections using the method described in the linked notebook guide and will already be de-quantized. The linked guide can be used by users registering newly delivered Custom Satellite Embedding datasets as Google Earth Engine ImageCollections.
COGs can be loaded into Earth Engine by either registering them as COG-backed assets or loading them directly, see the Custom Satellite Embeddings Data Access notebook on GitHub for a complete guide to both methods. For a general overview and comparison of both these methods, see Cloud Optimized GeoTIFFs (COGs) in Earth Engine.