A Satellite AI Model You Can Actually Use: Allen Institute Opens Up OlmoEarth Embeddings
Allen AI's OlmoEarth Studio can now export numerical 'fingerprints' of any patch of Earth, letting researchers map forests, detect wildfires and track land change with almost no labeled training data.
Key points
- Allen Institute for AI's OlmoEarth Studio now lets users export embeddings, compact numerical summaries of satellite imagery, for any area and time range they choose.
- Three model sizes are available: Nano (1.4 million parameters), Tiny (6.2 million), and Base (89 million), each trading speed for detail.
- In a test using only 60 labeled pixels, the system mapped mangrove forests in Vietnam with an F1 accuracy score of 0.84 out of 1.0.
- The underlying model weights and source code are publicly available, meaning anyone can inspect or build on the work.
- The Park Fire burn scar in Butte County, California, showed up in a change-detection test using just two months of satellite data and no manual labels.
If you have ever looked at a satellite photo and wondered which parts are forest, which are farmland, and which burned last summer, you are thinking like an Earth-observation researcher. Until recently, answering those questions required expensive datasets and months of expert labeling. A new tool from the Allen Institute for AI wants to change that.
OlmoEarth Studio, the Allen Institute's platform for building models that interpret satellite imagery, now lets users compute and export what the team calls embeddings. An embedding is a list of numbers, a kind of numerical fingerprint, that a trained AI model assigns to each patch of land it looks at. Places that look alike on the ground get similar numbers. Places that differ land far apart in that numerical space.
Think of it like a wine-tasting score sheet. Instead of describing a wine in prose, you record dozens of measurements: acidity, tannin, sugar. Two similar wines end up with similar score sheets even if you never tasted them side by side. OlmoEarth does the same for land: it reads satellite pixels and produces a score sheet for every location.
What can you actually do with these numbers?
Quite a lot, it turns out, even with almost no training data.
The simplest use is similarity search: pick one pixel, ask the model which other pixels look most like it, and get a heat map back. In a test over California's Merced region, an urban query lit up road corridors across the image while farmland stayed dark. No labels were needed.
More powerful is what the team calls few-shot segmentation, meaning you label a tiny handful of examples and the model figures out the rest. Researchers labeled just 60 pixels across a coastal mangrove region in Ca Mau, Vietnam: 20 tagged as mangrove, 20 as water, 20 as other. They fed those 60 tagged numbers into a basic text-book classifier (logistic regression, a decades-old statistical tool) and let it predict every other pixel. The result covered the whole region with a weighted F1 score of 0.84. Increasing the labeled sample to 300 pixels barely moved that number, because the embeddings are already doing the hard work.
| Model variant | Dimensions | Parameters | Typical use |
|---|---|---|---|
| Nano | 128 | 1.4 million | Fast, low-storage tasks |
| Tiny | 192 | 6.2 million | Balanced performance |
| Base | 768 | 89 million | Highest detail, more compute |
Change detection is perhaps the most striking demo. The team pulled monthly satellite embeddings for the same part of Butte County, California in September 2023 and September 2024, then measured how much each pixel's fingerprint shifted. The Park Fire burn scar from summer 2024 appeared immediately, no labels, no manual review.
Who is this for?
Right now, the tool is aimed at researchers and analysts who work with satellite data: conservation groups, urban planners, climate scientists, agricultural agencies. But because the model weights and code are public, smaller teams with coding skills can run the open-source version themselves, as first reported by Hugging Face in its model coverage.
The Studio version, which handles the satellite data acquisition automatically, requires contacting Allen AI for access. Custom spatial resolution runs from 10 meters per pixel down to 80 meters, and the time window can be as short as one month or as long as a year.
A fair word of caution: the examples shown are curated. Real-world results will vary by region, season and the quality of satellite coverage. The mangrove accuracy figure of 0.84 is strong, but it comes from a single, well-documented site. Your landscape may be messier.
Takeaway: If your work touches land, forests or environmental monitoring, it is worth bookmarking the OlmoEarth documentation and the technical report. The open weights mean you can experiment for free before deciding whether the managed Studio service fits your budget.
