AlphaEarth Foundations helps map our planet in unprecedented detail
Google DeepMind ● Covered by 2 sources
Google DeepMind built an AI that turns satellite data into a compact 'virtual satellite' map of Earth. It's already helping groups track deforestation, farming, and unmapped ecosystems way faster and cheaper than before.
Based on reporting by Google DeepMind — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Satellites generate an absurd amount of data every single day, but stitching all those images, radar scans, and climate readings into something usable has always been a mess. Google DeepMind's answer is AlphaEarth Foundations, a model that chews through petabytes of Earth observation data and spits out compact digital summaries, called embeddings, for every 10x10 meter patch of land and coastal water on the planet.
The trick isn't just aggregation. It's compression. Each embedding takes up 16 times less storage than comparable outputs from other AI systems DeepMind tested, which matters enormously when you're trying to analyze the entire globe rather than a single county. That efficiency, combined with a 24% lower error rate than rival models in DeepMind's benchmarks, is what lets researchers build detailed, consistent maps without waiting on one specific satellite pass over one specific spot.
DeepMind isn't keeping this locked away. The annual embeddings are now public as the Satellite Embedding dataset inside Google Earth Engine, and more than 50 organizations have spent the past year kicking the tires. The Global Ecosystems Atlas is using it to categorize ecosystems that have never been formally mapped, things like hyper-arid deserts and coastal shrublands, work that James Cook University's Nick Murray says is reshaping how countries decide where to focus conservation money. In Brazil, MapBiomas is applying the dataset to track agricultural and environmental shifts across the Amazon, with founder Tasso Azevedo calling it a way to produce maps that are faster and more precise than anything his team could manage before.
What's notable here is the scale: over 1.4 trillion embedding footprints per year, feeding into work at Harvard Forest, Stanford, Oregon State, and the UN's Food and Agriculture Organization. DeepMind is already talking about pairing this with reasoning models like Gemini down the line, treating it as one more piece of Google Earth AI rather than a finished product. For now, though, the real story is that a notoriously fragmented type of data just got a lot easier to actually use.
My take — AI-written commentary, not fact-checked reporting
This is the kind of AI application that deserves more attention than another chatbot demo, because it's solving a genuinely boring but critical problem: making Earth data cheap enough to use at planetary scale. I'll believe the climate and conservation upside once independent groups outside Google's partner list start publishing results, but locking the raw embeddings into Earth Engine rather than open-sourcing weights is a very Google move, and worth watching.
Read more about this at: Google DeepMind