Satellite Embedding V1

The Google Satellite Embedding dataset is a global, analysis-ready collection of learned geospatial embeddings. Each 10-meter pixel in this dataset is a 64-dimensional representation, or "embedding vector," that encodes temporal trajectories of surface conditions at and around that pixel as measured by various Earth observation instruments and datasets, over a single calendar year. Unlike conventional spectral inputs and indices, where bands correspond to physical measurements, embeddings are feature vectors that summarize relationships across multi-source, multi-modal observations in a less directly interpretable, but more powerful way.

The dataset covers terrestrial land surfaces and shallow waters, including intertidal and reef zones, inland waterways, and coastal waterways. Coverage at the poles is limited by satellite orbits and instrument coverage.

The collection is composed of images covering approximately 163,840 meters by 163,840 meters, and each image has 64 bands {A00, A01, …, A63}, one for each axis of the 64D embedding space. All bands should be used for downstream analysis as they collectively refer to a 64D coordinate in the embedding space and are not independently interpretable.

All images are generated in their local Universal Transverse Mercator projection as indicated by the UTM_ZONE property, and have system:time_start and system:time_end properties that reflect the calendar year summarized by the embeddings; for example, an embedding image for 2021 will have a system:start_time equal to ee.Date('2021-01-01 00:00:00') and a system:end_time equal to ee.Date('2022-01-01 00:00:00').

The embeddings are unit-length, meaning they have a magnitude of 1 and do not require any additional normalization, and are distributed across the unit sphere, making them well-suited for use with clustering algorithms and tree-based classifiers. The embedding space is also consistent across years, and embeddings from different years can be used for condition change detection by considering the dot product or angle between two embedding vectors. Furthermore, the embeddings are designed to be linearly composable, i.e., they can be aggregated to produce embeddings at coarser spatial resolutions or transformed with vector arithmetic, and still retain their semantic meaning and distance relationships.

The embeddings are produced by AlphaEarth Foundations, a geospatial embedding model that assimilates multiple datastreams including optical, radar, LiDAR, and other sources (Brown, Kazmierski, Pasquarella et al., in review).

Because representations are learned across many sensors and images, embedding representations effectively mitigate common issues such as clouds, scan lines, sensor artifacts, or missing data, providing seamless analysis-ready features that can be directly substituted for other Earth Observation image sources in classification, regression, and change detection analyses. While some large scale swath and data availability artifacts may be noticeable, these typically represent minor vector offsets and generally do not significantly affect downstream processing or results.