Allen AI's OlmoEarth Studio now exports satellite-image embeddings for custom downstream analysis

The Allen Institute for AI has added an embedding-export feature to OlmoEarth Studio, its platform for building Earth-observation models. Users can now compute and export embedding vectors — compact numerical representations of a piece of Earth-observation data — that the underlying OlmoEarth foundation model produces. The source code, model weights and the paper describing them are all publicly available.

The idea behind an embedding is that locations with similar surface characteristics — two irrigated agricultural parcels, say — end up with similar vectors, while very different areas land far apart in the same numerical space. Users pick an area of interest, a time range from 1 to 12 monthly periods, an encoder variant, a resolution and imagery source (Sentinel-2 L2A, Sentinel-1 RTC, or both), and get back a Cloud-Optimized GeoTIFF with one band per embedding dimension. Vectors are stored as 8-bit integers to keep files light.

One demonstrated use is similarity search — picking a query pixel, extracting its embedding, and computing cosine similarity against every other pixel to produce a heat-map of where a landscape looks most like the query. In one example, a query near the Merced urban centre in California lit up urban fabric and road corridors while agricultural parcels stayed dark, with no training data required.

A second use is few-shot land-cover mapping. To test the approach, Allen AI labelled just 60 pixels — 20 each for mangrove, water and 'other' — over the coastal Ca Mau region of Vietnam, trained a linear classifier on those pixels, and predicted every pixel in the region. The classifier reached a weighted F1 score of 0.84, and the team says accuracy barely changes when the label count grows from 30 to 300, because the embeddings themselves are doing most of the work.

Because embeddings are computed on demand rather than pulled from a pre-computed global archive, they reflect the exact conditions being studied — including monthly snapshots for seasonal dynamics. Custom-export is available to Studio users; instructions for computing embeddings from the publicly available OlmoEarth models are also published.