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

OlmoEarth Studio Enables Custom Embedding Exports for Geospatial Analysis

The Allen Institute for AI (AI2) has announced that its OlmoEarth Studio platform now supports the computation and export of custom embedding vectors. These embeddings are compact numerical representations generated by the open-source OlmoEarth foundation models, designed for Earth observation data. This development matters because it provides a more accessible and cost-effective method for performing complex downstream analysis, moving beyond pre-computed data archives to on-demand, user-defined vector generation.

The new feature allows users to tailor embedding exports to their specific needs directly through the Studio UI or API. The source code and model weights are publicly available, offering transparency into how the vectors are generated. Key customization parameters include:

• Encoder Variants: Nano (128-dim), Tiny (192-dim), and Base (768-dim) models are available, balancing performance with computational cost.
• Imagery Sources: Users can select Sentinel-2 L2A, Sentinel-1 RTC, or a combination of both.
• Spatiotemporal Range: Any area of interest can be defined by a polygon, with temporal resolution ranging from monthly to annual composites.
• Output Format: The results are delivered as a Cloud-Optimized GeoTIFF (COG) with signed 8-bit integer vectors, which are lightweight and compatible with standard geospatial tools like GDAL and rasterio.

This feature significantly lowers the barrier to entry for sophisticated geospatial AI applications. Instead of requiring extensive labeled data for supervised fine-tuning, developers and researchers can now leverage the rich, pre-trained representations of OlmoEarth for tasks like similarity search, few-shot segmentation, and change detection. The provided examples demonstrate that a simple linear classifier trained on just 60 labeled pixels can produce a highly accurate land-cover map, highlighting the power of the underlying embeddings. This approach enables rapid, scalable analysis, allowing users to quickly explore landscape structures and temporal shifts without the need for model training expertise or significant computational resources.

By offering direct access to model embeddings, AI2's OlmoEarth is moving the geospatial AI market beyond static data products and towards a more flexible, developer-centric model where the underlying learned representations become the primary asset for custom downstream analysis.