Allen Institute releases OlmoEarth v1.3 satellite embedding model
OlmoEarth v1.3 provides high-performance satellite data embeddings and precomputed global maps for efficient environmental monitoring.
Summary
The Allen Institute for AI has released OlmoEarth v1.3, a state-of-the-art satellite embedding model designed to compress complex satellite imagery into usable features. The release includes precomputed annual embeddings for the entire planet at 10-meter resolution and a new browser-based Embeddings Explorer tool. The model, weights, pretraining data, and code are all fully open.
Technical details of OlmoEarth v1.3
- Uses a vision transformer with an Attention Pooler to produce 64- or 128-dimensional vectors per 10-meter pixel.
- Processes Sentinel-1 (radar) and Sentinel-2/Landsat (multispectral) satellite time series.
- Outperforms AlphaEarth Foundations and Tessera v2 in 20 out of 27 evaluations.
- Includes the PLANTEUR benchmark for tropical crop-type classification.
- Training efficiency: v1.3 required only 53% of the compute time and 50% of the carbon emissions compared to v1.2.
Why it matters
By providing precomputed global embeddings, the project allows researchers to perform complex tasks like crop mapping or change detection using lightweight local classifiers instead of massive GPU resources.
Original reference: huggingface.co →