Researchers have developed a sophisticated artificial intelligence model that functions as a virtual satellite, creating a unified digital representation of the planet’s land and coastal waters. This new system integrates vast quantities of Earth observation data from multiple public sources, providing a more complete and consistent picture of planetary changes. By processing this complex, multimodal information in near-real time, the model aims to revolutionize global mapping and monitoring, aiding more informed decisions on critical issues like deforestation, food security, and urban expansion.
The core innovation, named AlphaEarth Foundations, overcomes the challenges of data overload and informational inconsistency that have hampered previous satellite monitoring efforts. The system analyzes the world in high-resolution 10-by-10 meter squares, tracking changes with remarkable precision over time. A key breakthrough is its ability to create a highly compact summary for each square, significantly reducing data storage and the cost of planetary-scale analysis. To accelerate research, a public dataset derived from the model is now available through Google Earth Engine, and has already been tested by more than 50 organizations, including the United Nations’ Food and Agriculture Organization and Stanford University, to enhance their environmental analysis and mapping work.
A New Foundation for Geospatial Data
The challenge of monitoring global land cover is immense. Every day, a fleet of satellites captures a torrent of images and measurements, but the sheer complexity, varying formats, and rapid refresh rate of this data make it difficult to synthesize effectively. Traditional methods often rely on single data sources, like optical satellite images, which can be obscured by clouds or fail to capture subsurface details. The AlphaEarth Foundations model represents a paradigm shift by fusing information from dozens of different sources, including optical imagery, radar that can penetrate clouds, 3D laser mapping, and climate simulations. This integration allows the AI to “see through” persistent cloud cover and analyze areas that are notoriously difficult to image, such as Antarctica.
This AI model does not just combine images; it creates a sophisticated, unified digital representation known as an “embedding.” This embedding is a compact, information-rich summary for each 10-meter square of the Earth’s surface that computer systems can easily process. This efficient data format makes planetary-scale analysis practical for a wider range of users. In testing, these summaries required 16 times less storage space than those from other AI systems, dramatically lowering the barrier to entry for researchers and organizations without massive computational resources. This allows scientists to create detailed, consistent maps on-demand, transforming what was once a monumental data-wrangling task into a streamlined analytical process.
Powering Real-World Applications
To demonstrate its practical utility, the developers have released a collection of the model’s annual embeddings as the Satellite Embedding dataset. This resource, one of the largest of its kind with over 1.4 trillion data footprints per year, is already driving significant advancements in environmental science. Over 50 partner organizations have been using the dataset to pioneer new applications, from mapping previously unclassified ecosystems to monitoring agricultural shifts with unprecedented speed and accuracy.
Conservation and Biodiversity
One of the most promising applications is in global conservation. The Global Ecosystems Atlas initiative is using the dataset to help countries map and classify ecosystems that have never been charted before, such as specific types of coastal shrublands or hyper-arid deserts. This detailed classification is critical for helping nations prioritize which areas to conserve, where to focus restoration efforts, and how to combat biodiversity loss most effectively. Nick Murray, a leader of the initiative, noted that the dataset is “revolutionizing our work” by providing the tools needed to pinpoint conservation priorities.
Agriculture and Environmental Change
In Brazil, the conservation technology organization MapBiomas is leveraging the dataset to gain a deeper understanding of agricultural and environmental changes, particularly in critical areas like the Amazon rainforest. By producing more accurate and precise maps at a faster rate, the system informs sustainable development strategies and conservation initiatives. According to MapBiomas founder Tasso Azevedo, the dataset offers “new options to make maps that are more accurate, precise and fast to produce – something we would have never been able to do before.” This capability is vital for monitoring the agricultural frontier and its impact on natural ecosystems.
The Evolution of Land Cover Mapping
The new AI-driven system builds upon decades of progress in remote sensing and land cover analysis. For years, scientists have relied on satellite programs like Landsat, a joint mission of the U.S. Geological Survey and NASA, which provides a multidecadal record of Earth’s surface at a 30-meter resolution. While invaluable for historical analysis, Landsat’s temporal frequency and spatial resolution are surpassed by modern systems. Datasets derived from Landsat have provided foundational knowledge, quantifying long-term changes in forests, croplands, and surface water between 2000 and 2020.
More recent satellite constellations, such as the European Space Agency’s Sentinel-2, offer higher spatial resolution and more frequent revisits, enabling near-real-time monitoring. These advancements have fueled the development of dynamic mapping platforms. For instance, the Dynamic World dataset, a collaboration between Google and the World Resources Institute, already produces near-real-time global land cover labels at a 10-meter resolution using Sentinel-2 imagery and deep learning. This platform provides nine distinct land cover classes, offering a detailed, probabilistic view of the landscape. The AlphaEarth Foundations model takes this a step further by not only classifying but also integrating and summarizing data from an even wider array of sources to create its powerful embeddings.
Technical Performance and Future Directions
The new system’s performance was rigorously tested against both traditional mapping methods and other AI systems. AlphaEarth Foundations was consistently more accurate across a wide range of tasks and time periods, including land use identification and surface property estimation. It proved particularly effective in scenarios where labeled training data was scarce, demonstrating a 24% lower error rate on average compared to other models tested, highlighting its superior learning efficiency.
The development team is continuing to explore new ways to leverage the model’s capabilities. While the current public dataset consists of annual embeddings, there is potential for even more frequent updates. Future work may combine the system’s time-based capabilities with the general reasoning skills of large language models like Gemini. This could unlock new analytical possibilities, allowing for more nuanced and context-aware interpretations of land cover changes. The ongoing development is part of a broader effort at Google to build a collection of geospatial models and datasets, called Google Earth AI, aimed at addressing the planet’s most critical environmental challenges.
