DeepMind's AlphaEarth AI Maps Planet in Unprecedented Detail

Satellites continuously capture vast amounts of data about our planet, offering scientists and experts a near real-time perspective. However, the sheer volume, diverse formats, and rapid refresh rates of this Earth observation data pose a significant challenge: effectively integrating disparate datasets to gain a comprehensive understanding.

To address this, DeepMind has introduced AlphaEarth Foundations, an artificial intelligence (AI) model designed to act as a "virtual satellite." This model efficiently processes petabytes of Earth observation data, converting it into a unified digital representation, or "embedding," that computer systems can easily interpret. This capability provides scientists with a more complete and consistent view of the planet's evolution, aiding informed decision-making on critical global issues such as food security, deforestation, urban expansion, and water resource management.

To accelerate research and practical applications, a collection of AlphaEarth Foundations’ annual embeddings has been released as the Satellite Embedding dataset within Google Earth Engine. Over the past year, more than 50 organizations have collaborated to test this dataset in real-world scenarios. Partners have reported significant benefits, utilizing the data to classify previously unmapped ecosystems, analyze agricultural and environmental changes, and enhance the accuracy and speed of their mapping efforts.

The model's ability to discern fine details is evident in its visualizations. For instance, in Ecuador, AlphaEarth Foundations can penetrate persistent cloud cover to accurately map agricultural plots at various stages of development. It can also clearly detail complex surfaces in Antarctica, an area notoriously challenging for satellite imaging due to irregular coverage. Furthermore, it reveals subtle variations in Canadian agricultural land use that are imperceptible to the human eye.

AlphaEarth Foundations tackles two primary challenges in Earth observation: data overload and inconsistent information. Firstly, it integrates massive volumes of data from dozens of public sources, including optical satellite images, radar, 3D laser mapping, and climate simulations. This diverse information is then woven together to analyze the world’s land and coastal waters in precise 10x10 meter squares, enabling remarkably accurate tracking of changes over time.

Secondly, the model makes this data practical and cost-effective to use. Its core innovation lies in generating highly compact summaries for each square. These summaries require 16 times less storage space compared to those produced by other tested AI systems, dramatically reducing the computational cost associated with planetary-scale analysis. This breakthrough empowers scientists to create detailed, consistent, on-demand maps, a capability previously unachievable. Whether monitoring crop health, tracking deforestation, or observing new construction, researchers are no longer reliant on single satellite passes but instead have a robust new foundation for geospatial data.

Rigorous testing has validated AlphaEarth Foundations' performance. When compared against both traditional methods and other AI mapping systems, it consistently demonstrated superior accuracy across a wide range of tasks and time periods, including identifying land use and estimating surface properties. Crucially, the model maintained its high performance even in scenarios where labeled data was scarce. On average, AlphaEarth Foundations exhibited a 24% lower error rate than the models it was tested against, highlighting its efficient learning capabilities.

The Satellite Embedding dataset, powered by AlphaEarth Foundations, is one of the largest of its kind, containing over 1.4 trillion embedding footprints per year. This extensive collection is already being utilized by organizations worldwide, including the United Nations’ Food and Agriculture Organization, Harvard Forest, the Group on Earth Observations, MapBiomas, Oregon State University, the Spatial Informatics Group, and Stanford University. These collaborations are generating powerful custom maps that yield real-world insights.

For example, the Global Ecosystems Atlas, an initiative focused on creating the first comprehensive resource to map and monitor the world’s ecosystems, is leveraging this dataset to help countries classify uncharted ecosystems into categories such as coastal shrublands and hyper-arid deserts. This resource is expected to play a vital role in enabling countries to better prioritize conservation areas, optimize restoration efforts, and combat biodiversity loss. Nick Murray, Director of the James Cook University Global Ecology Lab and Global Science Lead of Global Ecosystems Atlas, stated, “The Satellite Embedding dataset is revolutionizing our work by helping countries map uncharted ecosystems – this is crucial for pinpointing where to focus their conservation efforts.”

In Brazil, MapBiomas is testing the dataset to gain a deeper understanding of agricultural and environmental changes across the country. Maps generated from this data inform conservation strategies and sustainable development initiatives in critical ecosystems like the Amazon rainforest. Tasso Azevedo, founder of MapBiomas, commented, "The Satellite Embedding dataset can transform the way our team works – we now have new options to make maps that are more accurate, precise and fast to produce – something we would have never been able to do before."

AlphaEarth Foundations marks a significant advancement in understanding the state and dynamics of our changing planet. The team is currently generating annual embeddings and believes their utility could be further enhanced by combining them with general reasoning LLM agents like Gemini in the future. Further exploration of the model’s time-based capabilities is ongoing as part of Google Earth AI, a collection of geospatial models and datasets aimed at addressing the planet’s most critical needs.