Google DeepMind unveils AlphaEarth to build a digital twin of our planet

Every day, hundreds of satellites beam down petabytes of Earth imagery. The problem isn't getting the pictures—it's that optical shots, radar, and 3D laser scans don't speak the same language.

DeepMind and the Google Earth Engine team just introduced AlphaEarth Foundations, an AI model designed to act like a single virtual satellite that turns global geospatial chaos into a clean, searchable grid.

A 10-meter digital twin of the planet

AlphaEarth Foundations takes volumes of multimodal data from dozens of public sources—optical imagery, synthetic aperture radar, elevation mapping, and climate simulations—and fuses them together.

It breaks down all global land and coastal waters into sharp 10x10 meter squares. For every single patch, the model outputs a unified annual embedding, basically an information-dense digital fingerprint of what that piece of ground looks like and how it changes over time.

Google is rolling out these outputs directly into Google Earth Engine under the Satellite Embedding dataset, clocking in at over 1.4 trillion embedding footprints per year. It also sits inside Google Earth AI, a broader collection of geospatial models Google announced alongside it, which includes existing systems for flood forecasting, wildfire detection, and weather modeling.

16x smaller footprints, 24% lower error

The real engineering flex here comes down to compression and data scarcity.

Planetary-scale AI usually collapses under its own storage bill. DeepMind claims AlphaEarth creates summaries for each 10-meter square that take up 16 times less storage space than comparable AI systems they evaluated, drastically lowering the computing cost for researchers.

It also shines when ground truth data is missing. In benchmark tests across land-use classification and surface property estimation, AlphaEarth posted a 24% lower error rate on average than baseline systems, especially in label-scarce environments.

More than 50 organizations have spent the past year testing the dataset in the wild, including the UN's Food and Agriculture Organization, Stanford, and Harvard Forest. In Brazil, conservation group MapBiomas is using the embeddings to track agricultural shifts and deforestation across the Amazon, reporting unprecedented speed and mapping precision.

Why it matters

For environmental researchers, agricultural planners, and geospatial analysts, this turns months of painful sensor calibration and data cleaning into simple vector queries. If you already work inside Google Earth Engine, you don't have to stitch together Landsat and Sentinel passes anymore; the foundational features are already unified for you.

DeepMind played this smartly by prioritizing compact representation over sheer model weight. Shrinking footprint size by 16 times means real-world institutions without hyperscaler budgets can actually afford to run global inference.

We think this also redraws the competitive line in geospatial analytics. Google is essentially turning Earth Engine into the default operating system for planetary data, leaving proprietary satellite imagery providers to either integrate with Google's representations or fight for table scraps. The next test will be DeepMind's plan to wire these embeddings into reasoning agents like Gemini—moving from static mapping straight into automated planetary monitoring.