Google has released a new AI model that maps 10 meters squares and allows machines to read.
The company's Alphaearth Foundations model creates embeddings of computer programs because it is trained on vast amounts of Earth observation data from satellites and other sources, rather than visual images or other forms of output intended for human interpretation.
According to researchers at Google Deepmind, the Alphaearth Foundations “alphaearth Foundations accurately and efficiently characterize the planet's entire land and coastal waters by integrating large amounts of earth observation data that computer systems can easily process into a digital representation, or “embedded.”
Therefore, this model can provide scientists with a comprehensive view of the planet over time and support the study of environmental and agriculture concerns such as deforestation and water resources emitted by data centers running AI models.
An embedding is a vector (numerical representation of data points) created by a machine learning model that encodes data about some objects. Defines how different objects in the model relate to each other.
Mathematical vectors are modeled in dimension space. For example, a two-dimensional vector can be expressed as two numbers, x and y.
The Alphaearth embedding has 64 dimensions, each representing 10 meters of pixels encoding data from multiple sources about ground surface conditions around that plot over a year.
Alphaearth makes Earth observation information available as an embedded dataset so that it can be efficiently processed by deep learning applications such as Google Earth Engine. According to its creator, a concise data summary of the 10-meter square model allows it to operate at 16 times more storage space and at lower cost than other AI systems. The model is also said to be more accurate, with an error rate of 24% lower than the other models.
Researchers at Google Deepmind have shown that Alphaearth improves previous geospatial infrastructure models such as Satmae (Cong et al., 2022) and Satclip (Klemmer et al., 2025), by averaging embeddings across multiple sources, including time in the modeling framework, providing high-precision spatial resolution.
With the announcement of Alphaearth, Google has released satellite embedded datasets generated by the model to use them in applications such as the Earth Engine.
Google Research Scientist Valerie Pasquarella and product manager Emily Schechter propose that using datasets it can perform similarity searches (identify areas similar to other regions based on specified criteria) and can be used to detect geographical changes, discover hidden patterns, and create maps without manual labels.
Christopher Seeger, professor and extension specialist in landscape architecture and geospatial technology at Iowa State University, Register In telephone interviews, he expects Alphaearth Foundations to be extremely beneficial.
“It can be difficult to deal with only individual datasets, but bringing in multiple different datasets and trying to perform the analysis can result in intensive use of the machine,” he said. “And at best, you can usually only do small areas of what you're studying.”
He said he leaned towards AI to see features across multiple datasets, and looked like an interesting use of technology.
“Of course, I'm happy to see that they do some ground truth with this, in order to find out how reliable the model is, and I look forward to seeing what this is possible on a more local scale, not just on a global scale.”
Seeger said it makes great sense to apply AI to this type of data.
“What's interesting is that they can descend into a 10 x 10 metre square, which is incredible,” he said. “It's great for decision makers.” ®
