An AI map of Earth released on 22 October 2026 offers a new way to compare what the same stretch of land looked like in different years. Built by the Allen Institute for AI, or Ai2, the OlmoEarth Embeddings Explorer turns long sequences of satellite observations into patterns people can examine in a browser. You do not need to train a model or own a powerful computer simply to begin exploring.
The arresting example is an area of the Peruvian Amazon where gold mining has repeatedly disturbed forest. Ai2 shows how its maps highlight changing zones before and after a government crackdown. The colours are not literal photographs of excavators or proven evidence of a particular illegal act. They are signals that the surface changed, which can guide further investigation.
What an AI map of Earth actually sees
The system combines different types of satellite observations, including radar and measurements of light reflected from the ground. A model compresses these readings into a numerical description of each patch of land. Researchers call this an embedding, but a useful way to picture it is a fingerprint for a location at a particular time. Comparing fingerprints helps reveal where the underlying conditions have shifted.
Ai2 says it is publishing annual representations at ten-metre resolution and making a browser-based explorer available without a sign-in. A user can move around the map, select years, inspect the underlying satellite imagery and look for other places with similar patterns. That makes the system useful for initial questions about land use rather than a replacement for specialised fieldwork.
The project covers the whole planet for some years, with earlier historical coverage limited to selected countries. Users should check the dates available for the specific location they are investigating. A map with global-looking colours does not automatically mean every place has a continuous comparable record stretching back indefinitely.
A forest can change for more than one reason
In the Peruvian mining example, changes in the map's pattern coincide with shifts in disturbed ground and recovering vegetation described by environmental monitoring groups. That is a starting point for questions about where mining activity moved and whether previously cleared land began to recover. It is not, by itself, an answer to who caused a change or whether a particular activity was lawful.
A field, forest or coastal area may change with rainfall, seasonal cultivation, fire, flooding, construction or natural regrowth. Satellite observations also vary with atmospheric conditions, sensor characteristics and the dates on which images were captured. The underlying imagery and local knowledge remain essential checks when a map highlights something surprising.
LiveAIWire has previously covered AI systems detecting fishing activity from satellites. Both stories share the promise of watching remote areas more closely, but this release is different: it makes general-purpose land representations and an interactive comparison tool available for many kinds of mapping work rather than targeting one type of vessel or event.
Could it tell farmers what is growing?
One possible use is classifying crops in places where reliable maps are scarce. Ai2 says teams can add a relatively small amount of labelled information to the precomputed patterns and build specialised maps without repeating the most expensive model-training work. That can make experimentation more accessible to local researchers and organisations with limited computing resources.
There is, however, an important geographic test. In a separate release about the PLANTEUR crop-mapping benchmark, Ai2 describes how systems tested on tropical farming regions perform worse than on an established French benchmark. That finding is a reminder that success in familiar environments does not guarantee equivalent performance elsewhere. The institute's own research presents both the new capability and the need to test it carefully.
The same caution applies to identifying damaged forest, tracking urban growth or surveying land after a storm. A colourful result is a useful pointer, but the question being asked determines which errors matter. Missing a small field might be tolerable for an overview and unacceptable for a compensation or enforcement decision.
Opening the data changes who can investigate
The release includes model weights, data and code, as well as prepared map representations. Publishing these components allows other researchers to examine how the system behaves and build more focused tools. Ai2 says the explorer supports labelling sample locations and exporting a result for use elsewhere. That broadens the audience beyond well-funded groups able to process vast satellite archives from scratch.
Open access does not remove the work of interpretation. Anyone considering an important claim from the map still needs to check dates, the original observations, the relevant geography and the limits of the classification. The strongest result is one that survives comparison with independent evidence on the ground.
LiveAIWire has also discussed satellite-based crop monitoring and Google's separate work on AI-supported geographic disaster maps. This is a new Ai2 product with its own data release and evaluation, not proof those earlier systems were wrong or obsolete. The real advance is the combination of accessible browsing and reusable global representations.
For a curious reader, the most sensible next step is to inspect a familiar landscape and then switch to the source imagery. The interesting part is not merely seeing a coloured map. It is discovering which changes the model makes easier to notice and which questions still need a human answer.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.