Imagine a nature club studying a lake near its town. The members know where that lake is. What they want to find is other places with similar surroundings, without inspecting every corner of a satellite map.
That is the appeal of searching by example: point to a place and ask the computer to look for similar patterns.
On October 22, Ai2 released OlmoEarth v1.3 alongside a browser-based Embeddings Explorer. Ai2 says the explorer lets people select a location, find similar places, compare years and build simple maps from labeled examples, without signing in. The interesting part is what makes that possible: the satellite observations have already been turned into reusable numerical summaries.
A patch of land, written as numbers
An embedding is a list of numbers that a model uses to represent patterns in its input. Here, the input includes radar and optical satellite observations across a year.
The technical report describes how OlmoEarth combines those observations into 128 numbers for each output pixel on a 10-meter grid. A smaller, 64-number version is also available.
Think of it as a compact description that a computer can compare with other descriptions. It is not a written entry saying “lake with trees.” Similarity comes from the learned representation, which may capture patterns that are difficult to describe in a search box.
For our imaginary club, that changes the starting point. Instead of guessing the right words, members could use a known place to generate candidates, then inspect what those candidates actually contain. A match would be a lead to investigate, not a field survey.
The expensive preparation is already done
The practical advance is reuse. Ai2 has run the model over large areas and published the resulting embeddings. A new project can work with those summaries instead of repeatedly processing the original imagery through the full model.
The dataset documentation describes fitting a small classifier—a program that assigns categories—from labeled points. In the club’s example, members might provide examples of the kinds of surroundings they want to distinguish. Their labels would define the question; the embeddings would supply information for making the comparison.
The explorer brings a version of that workflow into the browser. Ai2 also describes an option to view the underlying Sentinel-2 imagery, which is useful when a colorful result needs a closer look.
Check the year before reading the map
These embeddings summarize years, not live conditions. At launch, the dataset lists global coverage for 2023, 2024 and 2025 within its mapped extent; 2017–2022 cover selected countries. The extent covers land and much of the coastal zone, excludes Antarctica and has some data gaps. A year appearing in the archive does not mean every place has data for it.
That matters for interpretation. A difference between annual summaries does not supply the exact day a change happened or establish its cause. The club would need other evidence to answer those questions.
To explore the idea, start with a familiar place in the OlmoEarth Embeddings Explorer, check its available year and compare a suggested match with the satellite imagery. The useful outcome is a better next question: what do these places share, and what would you need to check on the ground?