AlphaEarth Foundations in 2026: Inside Google's 10-Metre Map of the Planet

What AlphaEarth Foundations actually is

AlphaEarth is not a satellite. It's a foundation model trained on Sentinel-1 and Sentinel-2, Landsat, GEDI LiDAR, MODIS, ERA5 climate reanalysis and dozens of other public Earth observation streams. The model compresses all of that into a single continuous-time embedding per 10m pixel, per year. That one number (well — sixty-four numbers) encodes everything the satellites saw about that patch of ground: crop health, tree cover, built surfaces, moisture, snow, shipping activity.

Because the embeddings are analysis-ready, a basic Python user can classify land cover, detect deforestation or predict yield without training their own vision model, labelling imagery or wrangling tile pyramids. DeepMind's published benchmarks show the model beating prior approaches on ten out of eleven standard remote sensing tasks, with accuracy gains of 23.9% on average.

The 2026 Satellite Embedding release, explained

The original AlphaEarth announcement landed in late July 2025. The 2026 rollout is where it starts mattering operationally:

• March 2026 — state-of-the-planet 2025 embeddings went live in the Earth Engine Data Catalog. The annual collection now spans 2017 through 2025, so year-over-year change detection is a straight subtraction in embedding space.
• BigQuery vector search integration launched alongside it. You can now do similarity search across 1.4 trillion embedded points per year — ‘show me everywhere on Earth that looks like this 500-hectare deforestation front in Pará in 2023.’
• Google Cloud Storage availability means teams no longer have to run inside Earth Engine. The embeddings are downloadable as standard cloud-optimised GeoTIFFs, which opens up Snowflake, Databricks and any Python pipeline.
• Earth Engine tier quotas take effect 27 April 2026. Non-commercial projects must select a quota tier; commercial users continue to pay metered rates. Budget accordingly.

Who’s already using AlphaEarth in the wild

The ‘50 pilot organisations’ line from DeepMind’s 2025 launch has filled in with real users. Among the named adopters as of early 2026:

• MapBiomas — Brazil’s authoritative land-use monitoring coalition — uses AlphaEarth embeddings to track Amazon deforestation and Cerrado conversion at 10m resolution.
• The UN Food and Agriculture Organization (FAO) uses the dataset for crop classification and food-security monitoring across data-poor regions.
• Harvard Forest, Stanford University, Oregon State University and the Spatial Informatics Group use it for ecosystem mapping and biodiversity research.
• The Group on Earth Observations (GEO) has incorporated it into its open data commons.

This is the list of users sustainability teams and climate-disclosure auditors should be calling before commissioning their own bespoke remote-sensing stack. The gap between ‘build your own’ and ‘subscribe to AlphaEarth plus a BigQuery seat’ is now enormous.

How to pull AlphaEarth into Google Earth Engine

For technical readers, here’s the 60-second version:

1. In the Earth Engine code editor, load GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL.
2. Filter by year and geometry to get a 64-band image for your area of interest.
3. For change detection, take two annual embeddings and compute cosine distance per pixel — high distance = high change.
4. For classification, train a small classifier (even a basic Random Forest) on the 64-band embeddings plus a few hundred labelled points. You will routinely beat bespoke CNNs trained from scratch.
5. Export to BigQuery for similarity search, or to GCS for downstream pipelines.

The critical thing to internalise: you’re not classifying pixels, you’re classifying embeddings. That’s why it’s fast, that’s why it’s accurate, and that’s why the storage is 16× smaller.

What this means for marketers and climate-forward brands

You might reasonably ask why a marketing platform is writing about satellite embeddings. Two reasons.

First, sustainability claims are no longer defensible without geospatial evidence. The EU’s Corporate Sustainability Reporting Directive is in full force, the SEC’s climate-disclosure rule is back in court, and COP31 opens in Antalya in November 2026 with the operationalisation of loss-and-damage finance and the New Collective Quantified Goal on the table. Any brand claiming ‘deforestation-free supply chain’ or ‘regenerative sourcing’ will face auditors who can now verify those claims at 10m resolution, retroactively, for free. Marketing teams using Anjin need to know which of their campaign claims are about to be checked.

Second, AlphaEarth is a preview of what’s coming for every other dataset your brand touches. Five years ago, ‘predict flood risk in Jakarta’ required a PhD in remote sensing, a GPU farm and six months. Now it’s a Python notebook and a BigQuery query. The same collapse in friction is happening to consumer research, creative production, media planning and performance measurement. The firms that spot it early get the compounding advantage; the rest keep paying agencies to do it slowly.

Anjin: The Marketing Operating System for a data-rich, attention-poor world

Anjin is the Marketing Operating System — the climate-monitoring equivalent of what AlphaEarth delivers for the planet: a unified embedding of everything your organisation knows about its market, delivered to the people who need it, ready to reason over.

Anjin is a Marketing Operating System. It’s where your briefs, audiences, campaign creative, performance data, customer research, brand guidelines and AI agents live in a single workspace. Instead of sixteen tools and four agencies, one substrate. Instead of waiting six weeks for a strategy deck, agents that draft, critique and refine in hours. Instead of dashboards, instrumented outcomes.

The same logic that made AlphaEarth inevitable — compress the messy inputs into a queryable representation, let agents reason over it, let humans steer — is about to run through every marketing function that still depends on human coordination overhead. We’re just building the workspace for it.

Sources: Google DeepMind, Google Earth (Medium), BigQuery + AlphaEarth, AlphaEarth on GCS, Earth Engine Catalog, IEEE Spectrum, VentureBeat, UNFCCC COP31, MapBiomas, UN FAO