OlmoEarth v1.3: State-of-the-art embeddings for the whole planet

Hugging Face has launched OlmoEarth v1.3, a satellite embedding model designed to compress radar and multispectral satellite data into reusable vectors for tasks like land cover mapping and change analysis. The model processes one year of Sentinel-1, Sentinel-2, and Landsat imagery to produce 128-dimensional embeddings for every 10-meter pixel on Earth, with a lighter 64-dimensional option available. Precomputed embeddings for 2023, 2024, and 2025 are now accessible globally, while select countries cover 2017–2022, hosted on Source Cooperative for community use.

A new browser-based tool, the OlmoEarth Embeddings Explorer, allows users to visualize and analyze planetary embeddings without requiring sign-ins or advanced technical skills. The tool enables comparisons between years to track changes, identification of similar locations, and on-the-fly training of small classifiers using labeled points. Embeddings can be exported as GeoTIFF files for further analysis, and the underlying satellite imagery is accessible for verification.

OlmoEarth v1.3 was evaluated against AlphaEarth Foundations and Tessera v2 across 27 benchmarks spanning six continents, with the 128-dimensional embeddings outperforming competitors in 20 evaluations and the 64-dimensional version leading in 18. The model’s efficiency improved significantly, requiring 53% of the compute time and 18% of the carbon emissions of v1.2, while maintaining strong fine-tuning performance. Training costs for the base model are estimated at $2,720 of rented H100 time, with all weights, data, and code released openly.

The release includes a technical report and an interactive story demonstrating practical applications, such as tracking illegal gold mining in Peru’s Madre de Dios region. The embeddings reveal shifts in mining activity between 2017–2020 and 2020–2025, showing the impact of enforcement and ongoing environmental recovery. Users can trace embeddings back to source imagery via provenance tables, ensuring transparency and reproducibility in analysis.