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

OlmoEarth Studio, a platform for constructing Earth observation models, now enables users to generate and export embedding vectors from its open-source OlmoEarth foundation models. These compact numerical representations of Earth-observation data prove useful for various downstream tasks, including similarity search, segmentation, and unsupervised exploration.

The embeddings closely reflect locations with similar surface characteristics, while dissimilar locations result in widely separated vectors. In addition to their strong performance in benchmarking and independent evaluations, the exported Cloud-Optimized GeoTIFFs (COGs) are lightweight and easy to share.

Users can customize their embeddings by selecting the area of interest, time range, encoder variant, resolution, and imagery sources through the Studio's user interface or API. For enhanced performance, the platform also supports supervised fine-tuning (SFT).

To compute custom embeddings, users can follow the same workflow as any other prediction in Studio. They configure a model, run it, and then download the results. Several parameters allow users to tailor the output, such as visualization options and the COG format, which contains one band per embedding dimension.

The embeddings are stored as signed 8-bit integers (int8) with values ranging from -128 to +127, with -128 reserved for nodata. To recover floating-point vectors, users can use the dequantize_embeddings function in olmoearth_pretrain.

Unlike pre-computed global archives, Studio computes embeddings on demand, ensuring they accurately reflect the specific conditions of interest. This allows users to generate monthly embeddings to capture seasonal dynamics, rather than solely relying on annual snapshots.

By extracting an embedding from a specific query pixel and computing cosine similarity against every other pixel, users can create a heatmap that showcases areas within the landscape most and least similar to the query. This technique was demonstrated using OlmoEarth-v1-Tiny (192-dim) embeddings at 40-meter resolution with Sentinel-2 L2A composites.