What changed
Hugging Face announced the release of OlmoEarth v1.3, a satellite embedding model designed to provide state-of-the-art features for planetary analysis. The release includes precomputed annual embeddings for the entire planet at 10-meter resolution, available in both 64- and 128-dimensional versions. A new browser-based tool, the OlmoEarth Embeddings Explorer, allows users to explore and customize maps based on these embeddings without requiring sign-in or advanced technical knowledge.
What this means in practice
The OlmoEarth v1.3 model processes satellite data from radar (Sentinel-1) and multispectral (Sentinel-2 and Landsat) sources to generate embeddings for every 10-meter pixel on Earth. These embeddings can be used for tasks such as crop mapping, land-cover analysis, and change detection. The embeddings are precomputed and shared on Source Cooperative, making them accessible for lightweight modeling tasks. Users can train small classifiers on top of these embeddings to create maps locally, without needing GPU access or extensive data storage.
Limitations and availability
The embeddings are available for the entire planet, with annual data covering 2025, 2024, and 2023 globally, and 2017 to 2022 for select countries. The 128-dimensional version of the embeddings scores highest in 20 of 27 evaluations, while the 64-dimensional version leads in 18 of 27. The model was trained with 53% of the computing time and 18% of the carbon emissions compared to its predecessor, v1.2. The OlmoEarth Embeddings Explorer allows users to compare embeddings across years and analyze changes in landscapes, such as illegal gold mining in Peru's Madre de Dios region. The model weights, data, and code are all open under the OlmoEarth Artifact License.
