HomeArticle

The Earth version of ChatGPT has emerged spectacularly. Google AI compresses the human planet in 64 dimensions, and the 10 - meter - level "God's perspective" can be opened instantly.

新智元2025-07-31 16:54
The AI version of the "virtual satellite" makes its debut with 64-dimensional ultra-high precision.

Google DeepMind Unveils "God's Perspective": The New AlphaEarth Foundations Launches with a Bang, Offering 10-Meter Resolution for an Unprecedented Digital Portrait of Earth. Netizens Exclaim: Isn't This the "ChatGPT of Earth"?

Just now, the Alpha family has a new addition!

This time, Google DeepMind has opened a "God's perspective" for humanity. The new AlphaEarth Foundations is mapping the entire planet with astonishing detail.

It integrates petabytes of Earth observation data and can generate a unified data representation.

Specifically, AlphaEarth Foundations condenses the information of each 10x10-meter grid on Earth into efficient data with a total of 64 dimensions.

The 10-meter resolution is sufficient for you to see every corner of the Earth clearly.

Google has condensed the essence of one year's multi-source satellite data into each 10-meter square pixel.

The uniqueness of AlphaEarth Foundations lies in its powerful "feature learning" ability.

Through complex embedding techniques, the model can extract key features from optical, radar, and 3D data, easily distinguishing between beaches and deserts, forests and farmlands.

This ability enables it to outperform other AI and traditional methods in terms of performance, reducing the error rate by 24%.

On the same day, the Google team also released a 63-page comprehensive technical report.

Paper link: https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/alphaearth-foundations.pdf

AlphaEarth Foundations is like a "virtual satellite," providing humanity with a window to peek into the Earth's pulse.

It enables scientists to analyze the Earth's dynamics more quickly and efficiently, monitor crop health, track deforestation, and address global issues such as climate change.

A netizen praised that Google has taken an important step in building the "ChatGPT of Earth."

The question arises: Why do we need an AI-powered Earth model?

The AI-Powered "Virtual Satellite" Debuts with 64-Dimensional Ultra-High Precision

Every day, satellites capture every inch of change on Earth from space, generating a vast amount of images and observation data.

This data provides scientists and decision-makers with an almost real-time panoramic view of the Earth.

In the past 15 years, the Earth observation images and geospatial data available on the Earth Engine platform have completely changed the way we perceive the Earth.

However, its complexity, multi-modality, and high refresh rate have also given rise to a new challenge: How to connect heterogeneous datasets and utilize them efficiently?

The birth of AlphaEarth Foundations has become the key to solving this problem.

This is an AI model known as a "virtual satellite" that can integrate a vast amount of global observation data into a computer system and easily handle a unified digital representation (i.e., "embedding").

Ultimately, it maps all the land and coastal waters on Earth with unprecedented precision.

AlphaEarth Foundations not only provides scientists with a more complete and consistent picture of the Earth's evolution but also helps them make more informed decisions in areas such as food security, deforestation, urban expansion, and water resource management.

Working Principle

By addressing the two major challenges of "data overload" and "information inconsistency," AlphaEarth Foundations offers a new perspective for us to understand the Earth.

First, it integrates a vast amount of information from dozens of different public sources, including optical satellite images, radar, 3D laser mapping, and climate simulations.

After integrating all the information, it analyzes the global land and coastal waters with a clear ultra-high precision of 10x10 meters, tracking the Earth's changes over time.

Second, it makes this data easy to use.

The key innovation of this system is the ability to create a highly compact digital summary for each block area.

Compared with other AI systems, the storage space for these summaries is reduced by 16 times, significantly reducing the cost of planetary-scale analysis.

This breakthrough enables scientists to do something that has been impossible until now: create detailed and consistent world maps on demand.

The working principle of AlphaEarth Foundations: Extract non-uniformly sampled frames from video sequences to index positions at any time point. This helps the model build a continuous view of the location while interpreting a large amount of measurement data.

Whether monitoring crop health, tracking deforestation, or observing new buildings, they no longer need to rely solely on a single satellite passing overhead.

Now, they have a new type of "geospatial data foundation."

After rigorous testing, AlphaEarth Foundations has also demonstrated unparalleled accuracy.

It performs excellently in various tasks over different time periods, including identifying land use and estimating surface attributes.

Crucially, even when labeled data is scarce, its error rate is 24% lower on average than other models, demonstrating excellent learning efficiency.

The process of decomposing the global embedding field into individual embeddings (from left to right). Each embedding contains 64 components, which map to coordinate points on a 64-dimensional sphere.

In the generated map data, three of the 64 dimensions of the AlphaEarth Foundations embedding are assigned red, green, and blue colors respectively, enabling us to visualize the rich details of our world.

In Ecuador, the model can penetrate persistent clouds to present detailed views of farmlands at different stages of development.

In other places, it clearly maps the complex surface of Antarctica - an area that is extremely difficult to image due to irregular satellite imaging.

In addition, it also reveals the invisible differences in agricultural land use in Canada.

Next, we will break down the power of the dataset generated by AlphaEarth Foundations in detail.

10x10 Pixels, Condensing One Year's Data

The Google team used AlphaEarth Foundations to generate a globally pre - computed embedding dataset with a 10 - meter resolution for each year from 2017 to the present.

These "embedded" images may seem like ordinary Earth Engine image collections, but they integrate AI - enabled feature extraction capabilities into each pixel.

What information is contained in the "embedded" vector?

· Multi - source measurement data

The embedded vector learns from multiple data sources to capture the semantic information of surface attributes.

For example, the embedding of a pixel not only reflects its spectral characteristics but also includes information about the surrounding environment, seasonal changes (such as vegetation phenology and snow cover), as well as terrain and climate features.

· Spatiotemporal context

AlphaEarth Foundations was trained on over 3 billion independent image frames sampled from more than 5 million locations globally.

The model treats the satellite images of a location changing over time as continuous frames in a video.

Thus, it can learn across space, time, and measurement modes to generate embeddings that capture both spatial context and temporal trajectories.

This means that each embedded vector in the satellite embedding dataset provides a highly compact and semantically rich representation of the condition of each 10 - meter pixel (100 square meters) on the Earth's land surface.

The embedding of each 10 - meter pixel also captures information about its surrounding area.

Therefore, even if some areas (such as the asphalt pavement of a parking lot and a highway) look very similar in isolation, their embedded vectors will be very different.

· Viewing the Earth in 64 Dimensions: Coordinates and Bands

The images in the satellite embedding dataset have 64 bands - but they are different from classical optical reflectance or radar echoes.

On the contrary, the 64 "bands" of a single pixel in the AlphaEarth Foundations embedding represent a 64 - dimensional coordinate on a 64 - dimensional "sphere."

These coordinates are generated by deep learning, are highly interpretable mathematically, but are not direct physical measurements. Instead, they are a compact representation of the high - dimensional measurement space.

The "satellite embedding" is essentially a coordinate point on the surface of a 64 - dimensional "sphere."

With the satellite embedding dataset, scientists can perform "similarity searches."

By simply selecting a target pixel,