OlmoEarth v1.3: State-of-the-art embeddings for the whole planet
huggingface ·
Allen AI has released OlmoEarth v1.3, a state-of-the-art satellite embedding model that generates precomputed annual embeddings for the entire planet at 10-meter resolution. This release includes both 64-dimensional and 128-dimensional vectors derived from Sentinel-1, Sentinel-2, and Landsat data, allowing users to analyze planetary changes without requiring GPU access or advanced technical skills. The new OlmoEarth Embeddings Explorer provides a browser-based interface for interactive map exploration, similarity searches, and lightweight classifier training directly in the browser. In evaluations across 27 tasks on six continents, OlmoEarth v1.3 outperformed competitors like AlphaEarth Foundations and Tessera v2, demonstrating strong generalization capabilities even in unseen locations. The model is also significantly more efficient than previous versions, requiring only 53% of the compute time and 50% of the carbon emissions of its predecessor, v1.2. To illustrate the utility of these embeddings, the article highlights how they can track illegal gold mining activity in Peru’s Madre de Dios region over several years. All model weights, pretraining data, code, and global embeddings are fully open-sourced under the OlmoEarth Artifact License. Users can access these resources through Hugging Face, Source Cooperative, or by fine-tuning the base model via OlmoEarth Studio for specific custom requirements.
huggingface