Google's "virtual satellite" allows for Earth to be seen like never before

A new AI model capable of mapping the Earth in unprecedented detail promises to give scientists a sharper, near-real-time view of our planet’s ongoing transformation.

Developed by Google DeepMind, the tool—called AlphaEarth Foundations—combines massive volumes of Earth observation data to monitor environmental and climate-related changes with remarkable accuracy, as reported by Newsweek.

According to DeepMind, the system can “accurately and efficiently” map all of the world’s land and coastal waters by integrating petabytes of information—equal to around 1,000 terabytes or about one million gigabytes.

“This breakthrough enables scientists to do something that was previously out of reach: create detailed, consistent, on-demand maps of our world,” DeepMind said in a statement. “Whether it’s assessing crop health, tracking forest loss, or monitoring urban development, researchers no longer have to wait for a single satellite flyover—they now have a new kind of geospatial foundation.”

AlphaEarth Foundations offers what DeepMind calls a “more complete and consistent picture” of Earth’s surface, enhancing decision-making around deforestation, food production, urban growth, and natural resource management.

Described as a “virtual satellite,” the AI system fuses multiple sources of data—including optical imagery, radar scans, 3D laser mapping, and climate models—into a unified, digital format known as an embedding. This approach allows for efficient machine analysis across vast geographic areas.

The model segments the planet’s land and coastal regions into 10-by-10-meter squares, enabling precise, time-based tracking of surface changes. Its key innovation is a highly compact data summary for each square—requiring 16 times less storage than other AI systems and significantly cutting the cost of global-scale monitoring.

In tests against traditional mapping techniques and other AI models, AlphaEarth Foundations consistently delivered the highest accuracy across various tasks and time periods, from land-use classification to surface property estimation.

By Nazrin Sadigova