OlmoEarth v1.3: open satellite embeddings for all of Earth

What changed

The Allen Institute for AI (Ai2) released OlmoEarth v1.3 (Hugging Face blog). It reads a year of Sentinel-1 radar plus Sentinel-2 and Landsat multispectral time series and produces one 128-dimensional vector for every 10 meter pixel on Earth. The first 64 dimensions also work on their own as a smaller embedding.

Ai2 also precomputed annual embeddings for the entire planet and published them on Source Cooperative, roughly 150 TB per year. At launch, 2023, 2024 and 2025 are global, while 2017 to 2022 cover select countries.

New as well is the OlmoEarth Embeddings Explorer, a browser map that needs no account. You can flip between years, find similar places, train a small classifier from a few labeled points and export the result as a GeoTIFF.

How good is it?

Ai2 compared it with Google's AlphaEarth Foundations and Tessera v2 across 27 evaluations on six continents. The 128-dimensional version ranked first on 20 of them and leads overall. The 64-dimensional version, at half the size, won 18. For the evaluation Ai2 also introduced PLANTEUR, a new benchmark for tropical crop types.

Training got cheaper too: v1.3 used 53% of the compute time of v1.2 and 18% of the first version. Ai2 estimates the final pretraining run would cost about $2,720 in rented H100 time. For fine-tuning it matches v1.2.

Why it matters

Until now, the main way to use OlmoEarth was fine-tuning: training a separate model per task, which needed GPUs and the ability to download and store satellite imagery. With ready-made embeddings, a crop map or a change analysis shrinks to a small classifier you can run locally.

Ai2's example is Peru's Madre de Dios region: comparing embeddings across years shows, with no model training, where illegal gold mining spread between 2017 and 2020 and where the land is still changing through 2025.

For developers this is an open alternative to AlphaEarth Foundations, with weights, pretraining data and code available. Every embedding also ships with open provenance tables listing the satellite scenes it was built from.

Try it

The release includes the OlmoEarth v1.3 Base model (allenai/OlmoEarth-v1_3-Base on Hugging Face), the global embeddings, the Embeddings Explorer and the technical report. OlmoEarth Studio runs the model over a custom time window, and a tutorial covers generating embeddings on your own compute. It uses the OlmoEarth Artifact License, the same as earlier versions.