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 chosen time range.
- Three model sizes are available: Nano (1.4 million parameters), Tiny (6.2 million) and Base (89 million), trading speed for detail.
- In a test using only 60 labeled pixels, the system mapped mangrove forests in Vietnam with a weighted F1 accuracy score of 0.84 out of 1.0.
- The underlying model weights and source code are publicly available, so anyone can inspect or build on the work.
- The Park Fire burn scar in Butte County, California, appeared in a change-detection test using just two months of satellite data and no manual labels.
If you've ever looked at a satellite photo and wondered which parts are forest and which burned last summer, you're 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, and it's a real step forward, not a press-release promise.
OlmoEarth Studio, the Allen Institute's platform for interpreting satellite imagery, now lets users compute and export what the team calls embeddings. An embedding is a list of numbers, a numerical fingerprint, that a trained AI model assigns to each patch of land it examines. Similar landscapes get similar numbers; different ones land far apart in that numerical space.
Think of it like a wine 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, reading satellite pixels and producing a score sheet for every location.
What can you actually do with these numbers?
Quite a lot, 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, 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 needed.
More powerful is what the team calls few-shot segmentation, meaning you label a tiny handful of examples and the model handles 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 tagged numbers into a logistic regression, a decades-old statistical classifier, and let it predict every other pixel. The result covered the whole region with a weighted F1 score of 0.84. Scaling the labeled sample to 300 pixels barely moved that number, because the embeddings are already doing the heavy lifting.
| 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. No labels, no manual review: the Park Fire burn scar from summer 2024 showed up immediately.
Who is this for?
Right now it's aimed at researchers and analysts who work with satellite data: conservation groups, urban planners, climate scientists, agricultural agencies. Because the model weights and code are public, smaller teams with coding skills can run the open-source version themselves. We first covered OlmoEarth on 28 July 2026 when the platform ran a wildfire-risk model across all of North America in about 30 hours; the embedding export is the next practical layer on top of that infrastructure.
The Studio version, which handles satellite data acquisition automatically, requires contacting Allen AI for access. Spatial resolution runs from 10 to 80 meters per pixel, and the time window can be as short as one month or as long as a year.
A fair caution: the examples shown are curated. Real-world results will vary by region and the quality of satellite coverage. The 0.84 accuracy figure comes from a single, well-documented coastal site; your terrain may be messier.
Takeaway: If your work touches land or environmental monitoring, bookmark 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.
