Google DeepMind / Google Earth Engine / July 30, 2025

AlphaEarth Foundations

An Earth observation foundation model that turns massive, multimodal satellite and geospatial data into compact annual embeddings for global mapping and monitoring.

What It Is

AlphaEarth Foundations functions like a virtual satellite. It characterizes terrestrial land and coastal waters by combining diverse Earth observation streams into one model-ready representation.

Core Output

The model produces 64-dimensional embedding fields that summarize a location through time. Annual embeddings are available in Google Earth Engine as the Satellite Embedding dataset.

Primary Uses

It supports ecosystem classification, agricultural monitoring, deforestation analysis, urban expansion tracking, surface property estimation, and water resource assessment.

10x10 m Analysis squares for land and coastal waters

64 Embedding components per field

16x Less storage than other tested AI systems

24% Lower average error rate than tested alternatives

1.4T+ Annual embedding footprints in Earth Engine

50+ Organizations involved in real-world testing

Why It Matters

A consistent view of a changing planet

Satellites capture information-rich imagery and measurements every day, but the data is complex, multimodal, and refreshed at high frequency. AlphaEarth Foundations is designed to connect those disparate signals so analysts can work from a more complete and consistent picture of Earth.

DeepMind frames the model as useful for decisions around food security, ecosystem change, deforestation, urban growth, and water resources.

Visuals

Embedding views from the DeepMind article

AlphaEarth embedding visualization showing agricultural detail — Three embedding dimensions are mapped to RGB to reveal patterns in Earth observation data.

AlphaEarth embedding visualization of a difficult-to-image surface — The article highlights clear structure in areas that are hard to image consistently with conventional satellite passes.

AlphaEarth embedding visualization showing land-use variation — Embeddings can make land-use variation visible even when it is subtle in raw imagery.

How It Works

A compact representation built from many sources

AlphaEarth Foundations combines public data sources such as optical satellite imagery, radar, 3D laser mapping, and climate simulation outputs. It then produces a compact summary that can be used for downstream mapping and monitoring tasks.

DeepMind reports that this compact representation reduces storage requirements and makes planetary-scale analysis more practical.

Diagram animation illustrating AlphaEarth Foundations processing Earth observation frames through time — DeepMind diagram showing the model building a continuous view from unevenly sampled observations.

Representation

Global embedding fields

Each embedding has 64 components. DeepMind describes the result as a global embedding field that can be queried and used by mapping systems in Earth Engine.

Diagram of a global embedding field broken down into a single 64-component embedding — DeepMind diagram of a global embedding field and one 64-component embedding.

Dataset

Satellite Embedding dataset in Earth Engine

The annual Satellite Embedding dataset is available through Google Earth Engine. It contains more than 1.4 trillion embedding footprints per year and is intended to help researchers and organizations build custom maps more quickly.

Global Ecosystems Atlas MapBiomas Food and Agriculture Organization Harvard Forest Group on Earth Observations Oregon State University Spatial Informatics Group Stanford University

Model Details

Category Foundation Model

Organization Google DeepMind / Google Earth Engine

Spatial unit 10x10 m land and coastal water squares

Embedding size 64 components

Dataset access Satellite Embedding dataset in Google Earth Engine

Test partners 50+ organizations

Annual dataset scale More than 1.4 trillion embedding footprints per year

Reported result 24% lower error rate on average than tested alternatives

Resources

Google DeepMind article Satellite Embedding dataset Paper with Code Google Earth Engine blog