OlmoEarth Studio launches custom embedding exports for Earth observation
OlmoEarth Studio has rolled out a new feature that lets users create and export embedding vectors from Earth-observation data, using their open-source foundation models. The source code and model weights are open for public examination. These embeddings can be used for a variety of tasks, such as similarity searches and segmentation. Users can define their area of interest, time period, encoder type, resolution, and imagery sources through the Studio UI or API. The platform produces lightweight Cloud-Optimized GeoTIFFs (COGs), with each embedding dimension stored in its own band. Monthly embeddings highlight seasonal variations, and custom exports work with QGIS, GDAL, and rasterio. A tutorial and Colab notebook are also provided, though performance may vary with input quality.
Key facts
- OlmoEarth Studio now allows computing and exporting embedding vectors for Earth observation data.
- Embeddings are generated by open-source OlmoEarth foundation models; source code and weights are public.
- Embeddings support similarity search, segmentation, and unsupervised exploration.
- Exported files are Cloud-Optimized GeoTIFFs (COGs), lightweight and easy to share.
- Users can configure area of interest, time span, encoder variant, resolution, and imagery sources.
- Encoder variants: Nano (128-dim), Tiny (192-dim), Base (768-dim).
- Spatial resolutions: 10, 20, 40, or 80 meters per pixel.
- Imagery sources: Sentinel-2 L2A, Sentinel-1 RTC, or both.
- Vectors stored as int8, values -127 to +127, -128 for nodata.
- Examples include similarity search near Merced, California; few-shot segmentation in Ca Mau, Vietnam; change detection of Park Fire in Butte County, California; PCA exploration of Flevoland, Netherlands.
- Few-shot segmentation achieved weighted F1 = 0.84 with only 60 labeled pixels.
- Supervised fine-tuning (SFT) is available for higher performance.
- Sentinel-2 L2A imagery from ESA accessed via Microsoft Planetary Computer; ESA WorldCover 2021 v200 used.
- Global clustering visualization uses OlmoEarth-v1-Base with 1.1M pretraining samples.
Entities
Institutions
- OlmoEarth Studio
- Microsoft Planetary Computer
- European Space Agency
Locations
- Merced, California
- California
- Ca Mau, Vietnam
- Vietnam
- Butte County, California
- Flevoland, Netherlands
- Netherlands