AI BriefingOlmoEarth Embedding Export Support
Key point
OlmoEarth Studio now supports on-demand generation and export of Earth observation data embeddings.
Details
Allen Institute for AI has added embedding generation and export capabilities for the OlmoEarth foundation model for Earth observation to OlmoEarth Studio.
Users can generate embeddings on-demand via the Studio UI or API by specifying the following conditions:
- Area of interest and time period (1–12 months)
- Encoder variants: Nano (128 dimensions, 1.4 million parameters), Tiny (192 dimensions, 6.2 million parameters), Base (768 dimensions, 89 million parameters)
- Spatial resolution: 10·20·40·80m
- Data sources: Sentinel-2 L2A, Sentinel-1 RTC, or a combination of both
Results are provided as Cloud-Optimized GeoTIFF (COG) files including per-dimension bands, with embedding values quantized to signed int8 for storage efficiency. The value -128 is reserved for nodata, and floating-point vectors can be restored using dequantize_embeddings if needed.
Generated embeddings can be used for similarity search, few-label segmentation, land cover analysis, unsupervised exploration, and seasonal/time-series change detection. OlmoEarth's source code, model weights, research paper, and custom embedding computation methods have also been released. The custom embedding feature in Studio is currently available in an access-restricted form.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NASA-IBM Releases Open-Source Multimodal Geospatial AI Model 'Lunar Foundation' for Lunar Exploration