From the source
Ai2 is releasing OlmoEarth v1.3, a satellite embedding model it says produces one 128-dimensional vector for every 10-meter pixel on Earth, and sharing annual embeddings for the whole planet.
- The post says users previously had to fine-tune in OlmoEarth Studio, which needs GPU access and imagery storage, while embeddings compress satellite data into features that can be reused cheaply.
- Ai2 evaluated the embeddings against AlphaEarth Foundations and Tessera v2 across 27 evaluations spanning six continents; it says the 128-dimensional version scores highest on 20 of 27 and the 64-dimensional version, at half the size, leads on 18.
- Ai2 says OlmoEarth v1.3 was its cheapest model to train, taking 53% of the computing time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions.
- Precomputed embeddings sit on Source Cooperative, and the OlmoEarth Embeddings Explorer lets users pan a false-color map, flip between years, find similar places, train a small classifier and export a GeoTIFF without an account.
- The archive holds 10-meter-resolution embeddings at roughly 150 TB per year; at launch 2025, 2024 and 2023 are global, and 2017 to 2022 cover select countries.
