Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis

August 12, 2026 | Source: Hugging Face Blog

Tags: OlmoEarth, Allen Institute, AI2, Earth observation, satellite imagery, embeddings, geospatial AI

Allen Institute's OlmoEarth Studio adds embedding export for Earth observation data — offering three open-source encoder variants (Nano 128-dim, Tiny 192-dim, Base 768-dim) from publicly available model weights, enabling similarity search, segmentation, and unsupervised analysis of Sentinel satellite imagery.

Details

OlmoEarth is AI2's open-source foundation model platform for Earth observation. The new embedding export feature lets users generate compact numerical representations of Sentinel-2 L2A and Sentinel-1 RTC imagery, delivered as Cloud-Optimized GeoTIFFs with one band per embedding dimension, stored as signed 8-bit integers (int8, range -127 to +127). Three encoder variants cover different capability/cost tradeoffs: Nano (128-dim, 1.4M params), Tiny (192-dim, 6.2M params), and Base (768-dim, 89M params). Users configure area of interest, time span (1-12 months), spatial resolution (10-80m/pixel), and imagery source via the Studio UI or API. Embeddings are computed on-demand rather than from a pre-built global archive, ensuring results reflect the exact requested configuration. Locations with similar surface characteristics cluster together in embedding space; dissimilar locations land far apart. OlmoEarth embeddings support downstream tasks including similarity search, segmentation, and unsupervised exploration. All source code, model weights, and the research paper are publicly available. For higher-performance applications, Studio also supports supervised fine-tuning.