The hardest part of machine learning on satellite imagery was never the algorithm — it was the sheer logistics of the data. Google DeepMind's AlphaEarth Foundations, announced July 2025, attacks that directly: it analyzes the Earth in 10×10-meter squares and compresses petabytes of optical, radar, lidar, and climate observations into a 64-dimensional embedding vector for every pixel [1].
Why embeddings matter for ordinary GIS work
An embedding is a learned numerical summary — and the point is that you don't have to train anything. Classification, similarity search, change detection: they become operations on pre-computed vectors instead of deep-learning projects. DeepMind reports a 24% lower error rate than the models it tested against, with embeddings needing 16 times less storage than comparable systems [1]. The annual Satellite Embedding dataset contains over 1.4 trillion embedding footprints per year, and more than 50 organizations — from the UN FAO to Harvard Forest, MapBiomas, and Stanford — tested it pre-release [1].
AlphaEarth is one piece of Google's broader "Earth AI" umbrella — weather, flood, wildfire, and population models that "already support millions of users" through alerts in Search and Maps [2].
The November twist: the data left the walled garden
The original dataset lived in Google Earth Engine. In November 2025, the Taylor Geospatial Engine and Radiant Earth replicated the full 465 TB archive as Cloud-Optimized GeoTIFFs on Source Cooperative (source.coop/tge-labs/aef) [3] — meaning you can now read AlphaEarth embeddings with plain open-source tooling, no Earth Engine account required. It is hard to overstate how unusual that is: a frontier-lab Earth observation product, in an open format, on open infrastructure.
How to dip a toe in
Pick a small area you know well. Pull the embedding tiles for two different years, cluster or difference them, and compare against what you know changed on the ground. The workflow fits in a notebook, runs on a laptop, and teaches you more about what foundation-model geography can and cannot see than any launch post — including this one.