Science

AlphaEarth, Google's Ai that maps the Earth with unprecedented accuracy

Available for free through Google Earth Engine for researchers and organisations, it is the new frontier of global maps according to DeepMind

5 August 2025

3' min read

Google DeepMind has just unveiled AlphaEarth Foundations, an artificial intelligence model that functions as a 'virtual satellite', capable of mapping the entire planet with an accuracy of up to 10 by 10 metres. The system processes petabytes of Earth observation data from multimodal sources, including optical satellites such as Sentinel 2 and Landsat, synthetic aperture radar (SAR), 3D lidar surveys, climate models, elevation data and even georeferenced text. All this information is merged into a coherent digital representation of the planet, made available through the Satellite Embedding dataset on Google Earth Engine.

The annual dataset is ready to use, distributed under an open source licence (Apache 2.0) and accompanied by tutorials and a grant programme of up to $5,000 for scientific applications. Unlike many previous solutions, it does not require complex pre-processing steps (such as cloud masking or atmospheric correction), and thus provides more immediate and cost-effective access to advanced Earth observation capabilities. In this way, technologies hitherto reserved for large research centres or companies with significant computational resources also become accessible to local authorities, start-ups, NGOs and universities.

The numbers of the new application

AlphaEarth is not a satellite, but it acts as one. The system interprets dozens of heterogeneous sources - optical, radar, elevation, meteorological and textual data - and merges them into a single numerical representation, called 'embedding', which condenses the information collected over an entire year into a 64-dimensional vector space. It is like translating the appearance, function and composition of every square metre of the Earth into a machine-readable formula. This not only reduces the space needed to store the data, but also improves its quality. Tests show a 24% reduction in average error compared to previous approaches, with superior performance even in complex conditions such as the Amazon, the Sahel or Antarctica.

The result is a drastic reduction (sixteen times less) in storage space. Translated into economic terms, this means a clear reduction in costs for geospatial analysis on a planetary scale, a crucial factor in making accessible technologies that until recently were reserved for large companies and research centres. The system analyses the planet's land surfaces and coastal waters by dividing them into 10×10 metre squares, and allows changes over time to be tracked with an extraordinary level of detail. Each pixel, in other words, covers an area smaller than a football field, enough to monitor individual city blocks or small agricultural plots.

The dataset generated by AlphaEarth is among the largest of its kind ever produced: over 1.4 trillion geospatial embeddings per year. The system does not just provide images: it transforms raw data into operational intelligence, also enabling predictive modelling of environmental changes.

Beyond the clouds: when AI sees what satellites cannot

AlphaEarth excels in complex scenarios and edge cases, such as producing accurate maps in constantly cloudy areas. In Ecuador, for example, the model can see through persistent cloud cover to detail agricultural plots at various stages of development. In Antarctica, an area notoriously difficult to image due to irregular satellite imaging, it can map a complex surface in clear detail. And in Canada it has made visible variations in agricultural land use invisible to the naked eye. The key lies in the 'Space Time Precision' (STP) architecture, which processes spatial, temporal and resolution data simultaneously. For the first time, an Earth observation system supports 'continuous time': it can create accurate maps for any specific period, even interpolating between observations or extrapolating over periods without direct satellite coverage.

Compared to traditional satellite maps, the main advantage is that DeepMind's model eliminates much of the 'noise' associated with daily variability. It does not work on instantaneous images but on annual average data, allowing for a more stable and representative view. Unlike many machine learning models, the system does not attempt to reconstruct pixel by pixel the visual appearance of the territory, but its statistical, functional and environmental nature.

Partners and use in the field

Over the past 12 months, more than 50 organisations have already tested the dataset on real-world applications, including the FAO, Stanford University and the Brazilian MapBiomas project, the platform for monitoring deforestation in Brazil, launched in collaboration with Google. The use cases are many: from combating deforestation to monitoring coastal erosion, from agricultural planning to choosing ideal sites for photovoltaic plants. In academia, some researchers are using it to classify complex ecosystems, analyse land consumption or study the effects of climate change on a local scale.

Next Future: when AI thinks about space

Google is exploring the integration of AlphaEarth with general reasoning LLM agents such as Gemini, opening up even more ambitious scenarios: a system that does not just see changes, but understands them, anticipates them and suggests actions to mitigate problems.