Hugging Face’s OlmoEarth Studio now exports custom Earth-observation embeddings for downstream analysis

Hugging Face’s OlmoEarth Studio now lets users compute and export embedding vectors—compact numerical representations of Earth-observation data—produced by the open-source OlmoEarth foundation models. The embeddings are delivered as Cloud-Optimized GeoTIFFs (COGs), which are lightweight and designed for easy sharing and downstream analysis.

Users can tailor embeddings by selecting an area of interest, time span (1–12 monthly periods), encoder variant (Nano, Tiny, or Base), spatial resolution (10m to 80m per pixel), and imagery sources (Sentinel-2 L2A, Sentinel-1 RTC, or both). The exported COGs contain one band per embedding dimension, stored as signed 8-bit integers (int8) ranging from -127 to +127, with -128 reserved for nodata.

OlmoEarth embeddings support several downstream tasks, including similarity search, few-shot segmentation, and change detection. For similarity search, users can extract an embedding for a query pixel and compute cosine similarity across the raster to identify locations with similar surface characteristics. In few-shot segmentation, a simple linear classifier trained on as few as 60 labeled pixels achieved a weighted F1 of 0.84 for land-cover mapping in a coastal mangrove region. Change detection is enabled by generating embeddings at monthly or annual intervals and comparing them to spot shifts in surface conditions.

The embeddings are computed on demand, ensuring they reflect the exact conditions specified by the user rather than relying on pre-computed global archives. This allows for dynamic analysis, such as capturing seasonal dynamics with monthly embeddings rather than annual snapshots.

OlmoEarth Studio also supports supervised fine-tuning for applications requiring higher performance. Access to custom-computed embeddings is available to users of OlmoEarth Studio, with instructions for computing embeddings using the publicly available OlmoEarth models provided in the documentation.