Google’s “Virtual Satellite” Is Here: AlphaEarth AI Model Maps the Planet in Unprecedented Detail
Engineers at Google DeepMind have unveiled AlphaEarth Foundations, a sophisticated artificial intelligence model capable of generating a digital representation of the Earth’s surface at an impressive resolution of 10×10 meters. Unlike conventional systems designed for visual display, AlphaEarth constructs what are known as embeddings—numerical vectors that encode the intrinsic properties of each point on the planet, enabling automated interpretation without the need for manual labeling.
AlphaEarth was trained on vast datasets derived from remote sensing, including satellite imagery and various observational streams. The system produces high-dimensional vectors that encapsulate the characteristics of terrestrial and coastal regions, taking into account climate patterns, seasonal variations, and topographic features. Rather than yielding images in the traditional sense, it generates structured representations tailored for consumption by other AI models and analytical tools.
Each embedding corresponds to a 100-square-meter cell and comprises 64 numerical parameters that describe the terrain. These features are synthesized from heterogeneous data sources, offering an abstract yet comprehensive annual summary of the region. This format facilitates tracking of environmental changes, classification of similar territories, and discovery of underlying patterns—all without human intervention.
One of AlphaEarth’s primary applications lies in Google Earth Engine, the geospatial analysis platform used for environmental monitoring, agricultural planning, forest inventory, and water resource management. The system has already demonstrated potential in assessing deforestation impacts, analyzing drought conditions, and calculating water consumption by data centers powering large-scale AI infrastructure. Thanks to its vector-based approach, users can identify similar land segments, generate thematic maps, and detect changes—without the burden of manual classification.
According to its developers, AlphaEarth surpasses its predecessors—SatMAE (2022) and SatCLIP (2025)—in both accuracy and efficiency. This leap in performance stems from key architectural innovations: harmonizing data from diverse sources, integrating seasonal awareness, and delivering fine-grained spatial resolution. Testing reveals the model requires 16 times less memory while achieving a 24% reduction in error rates.
In tandem with the AlphaEarth launch, Google introduced the Satellite Embedding Dataset—a structured, open-access repository of the model’s embeddings. This dataset is now available to researchers, developers, and academic teams who can integrate these vectors into their own analytical frameworks. It is particularly valuable in domains that demand unified structures from disparate data sources, such as climate modeling, agroanalytics, or risk forecasting.
Christopher Seeger, Professor of Geoinformatics and Landscape Architecture at Iowa State University, told reporters that merging scattered data into a coherent analytical space is an inherently resource-intensive challenge. AlphaEarth, he noted, streamlines this process, enabling coverage of expansive areas rather than isolated observation points.
What captivated the audience most was the spatial granularity: the division of Earth’s surface into discrete cells allows for analysis not only on a global scale but also at the level of individual regions. Seeger also underscored the importance of ground truth verification—testing the model’s accuracy against field measurements—to avoid the pitfalls that often arise from reliance solely on simulations.