Google DeepMind's AlphaEarth Foundations Maps the Planet Using AI and Satellite Data

Google DeepMind has unveiled AlphaEarth Foundations, an AI system that acts as a virtual satellite, synthesizing vast amounts of Earth observation data into a unified digital format that scientists can readily access and analyze.

The model tackles two critical challenges in planetary monitoring: managing the overwhelming volume and complexity of satellite imagery and other Earth data, and reconciling inconsistencies across different sources. By combining optical imagery, radar, laser mapping, and climate simulations, AlphaEarth Foundations analyzes the world's land and coastal waters in 10×10 meter squares, tracking changes over time with high precision.

How It Works#

A key innovation is the system's compression efficiency. It creates summaries of satellite data that require 16 times less storage than competing AI systems, drastically reducing the computational cost of planetary-scale analysis. The model achieved 24% lower error rates on average compared to other tested systems, particularly excelling when label data was limited.

Real-World Applications#

The team is releasing the Satellite Embedding dataset through Google Earth Engine, containing over 1.4 trillion embedding footprints annually. Partners including the UN's Food and Agriculture Organization, Harvard Forest, MapBiomas, and Stanford University are already using the dataset to advance conservation and agricultural monitoring.

In Brazil, MapBiomas is leveraging the data to better understand environmental changes across the Amazon rainforest. The Global Ecosystems Atlas initiative is using it to classify unmapped ecosystems, helping countries prioritize conservation areas and combat biodiversity loss.

As Tasso Azevedo, founder of MapBiomas, noted: the dataset can transform how teams work with maps that are more accurate, precise and fast to produce.

Looking Ahead#

Google DeepMind believes the technology could become even more powerful when combined with large language models like Gemini, enabling broader reasoning about planetary dynamics. The development represents part of Google Earth AI, a broader initiative offering geospatial models and datasets for addressing critical global challenges.