Google DeepMind recently launched an artificial intelligence system called AlphaEarth Foundations, which aims to convert vast amounts of satellite data into a unified digital representation to improve the accuracy of environmental analysis and support decision-making on issues such as food security, deforestation, and water resources. AlphaEarth Foundations can be considered a "virtual satellite," depicting all land areas and coastal waters of the Earth at a resolution of 10x10 meters.
This model integrates multiple data sources, including optical satellite imagery, radar, 3D laser mapping, and climate simulations. By compressing these input data into 64-dimensional embeddings, DeepMind achieved an efficient representation of the data. During training, AlphaEarth Foundations used over 3 billion observations from more than 5 million locations around the world, with data sources covering satellite missions such as Sentinel-2 and Landsat, as well as text information such as Wikipedia articles and species observations.
The system's goal is to address two core challenges: data overload and inconsistent information. AlphaEarth Foundations can penetrate persistent cloud cover, map the complex surfaces of Antarctica, and reveal subtle changes in Canadian wheat cultivation that are imperceptible to the human eye. In comparative tests with traditional methods and other AI mapping systems, AlphaEarth Foundations had an average error rate that was 24% lower. The model performed exceptionally well on 15 evaluation datasets, including land use classification, biophysical variable estimation, and change detection.
AlphaEarth Foundations is also capable of working effectively in data-scarce situations. Its continuous time analysis feature allows the system to make precise predictions for periods that are not fully aligned. The model's "spatiotemporal precision" (STP) architecture treats satellite images from different periods of the same location as frames in a video. This approach enables the system to learn relationships between space, time, and measurements, generating embedding representations that capture local environmental and temporal trajectories.
Currently, more than 50 organizations have tested this system in real-world applications. The Global Ecosystems Atlas uses this data to classify previously unmapped ecosystems, including coastal shrubs and hyper-arid deserts. MapBiomas in Brazil uses this data to deeply analyze agricultural and environmental changes, especially in key ecosystems such as the Amazon rainforest.
In addition, Google will release an annual Satellite Embedding Dataset on Google Earth Engine. According to Google Earth Engine data, this dataset generates over 1.4 trillion embedding footprints annually, providing various application scenarios for identifying similar global environmental conditions, change detection, automatic clustering, and smarter classification.
To accelerate scientific research, Google also offers research funding of up to $5,000 to support case studies on satellite embedding-based applications. The development team at DeepMind believes that AlphaEarth Foundations is an important step in understanding our changing planet and its dynamics and looks forward to combining it with general reasoning large language models (LLMs) to create more powerful applications.
Key Points:
🌍 AlphaEarth Foundations is a virtual satellite AI system developed by Google DeepMind that can depict all land areas and coastal waters of the Earth at a resolution of 10x10 meters.
📊 The system integrates multiple data sources and efficiently represents them through 64-dimensional embeddings, improving the accuracy of environmental monitoring, with an average error reduction of 24%.
💡 Google will release a satellite embedding dataset to support global research and provide research funding to promote scientific applications.