What Ai2 released on October 22
Ai2 released OlmoEarth v1.3, a satellite-embedding model, along with precomputed annual embeddings and a browser-based Embeddings Explorer. It uses radar and multispectral imagery from Sentinel-1, Sentinel-2, and Landsat. Each 10-meter pixel gets a 128-number vector; Ai2 also publishes the first 64 dimensions as a smaller option. At launch, 2023–2025 are global, while 2017–2022 cover selected countries. Ai2 says the archive is about 150 TB per year. A small vector per pixel still adds up at planetary scale.
A vector describes a pixel; it does not name it
An embedding is a list of numbers arranged to preserve useful similarities. Here, it summarizes a pixel’s annual satellite imagery. Nearby vectors can signal that two places look similar to the model, but the vector does not itself say “forest,” “crop,” “flood,” or “healthy.” A separate model needs labeled examples to learn those categories, and someone still needs to check whether its predictions match the real place. Our beginner lesson explains embeddings with text; the same vector idea applies to images.
Precomputed features reduce repeated work, not every task
Ai2 says its ready-made embeddings let teams begin a crop, land-cover, or change map with a small classifier instead of processing the source imagery from scratch for each task. Inference: reusing these vectors can save repeated feature extraction when a project needs the same annual inputs. It does not remove the need for labeled ground truth or a task-specific check. A project that needs different dates, sensors, or resolution may need to generate new embeddings; Ai2 points those users to OlmoEarth Studio.
Check years and geography before relying on a map
The published archive is annual: it is well suited to year-to-year questions, but may blur a short event or a change within a season. That is an inference from the annual summaries, not a limitation we measured. Ai2 says 2023–2025 coverage is global and earlier years cover selected countries. Source Cooperative publishes a coverage mask and warns that Antarctica is excluded; a nodata value can also mean a pixel is outside coverage or lacked imagery in that year's 30-day Sentinel-2 mosaics. Check the actual year and coverage for your area before designing a workflow around it. Also review the OlmoEarth Artifact License before use: Ai2's posted terms restrict military/defense and human-surveillance or policing uses, and extractive activities such as mining, drilling, and deforestation. This summarizes Ai2's terms; it is not legal advice.
Try a small check with fake land-cover data
- Create a tiny table of invented map tiles with fake vectors, dates, regions, and labels such as “forest” or “field.” Keep the example data clearly separate from real satellite observations.
- Train a simple classifier on some tiles from one region and year. Hold back tiles from another region or year so the check asks whether the pattern transfers, not only whether it memorized the examples.
- Compare each prediction with the held-out mock label and inspect every mismatch. A nearby vector is a similarity clue, not proof that the tile has the same land cover.
- For a real project, use documented labels, inspect the source imagery, record the year, area, and sensor inputs, and test the exact regions and dates that matter. This exercise does not run OlmoEarth or use its dataset.
A feature layer is not a verified map
Ai2 presents OlmoEarth as a high-performing satellite embedding model and reports its own evaluation results. Those are the publisher’s results, not an independent validation of your classifier, region, or decision. We did not download the weights or data, use the Explorer, or reproduce a benchmark. Treat the vectors as reusable features and validate any labels or decisions built on top of them.