OlmoEarth v1.3 brings global satellite embeddings and a tool for exploring maps

The authors have released OlmoEarth v1.3, precomputed annual satellite embeddings covering the entire planet, and the browser-based tool OlmoEarth Embeddings Explorer. The data is available on Source Cooperative; the model weights, pretraining data and pretraining code have also been released.

The model processes annual time series of radar data from Sentinel-1 and multispectral data from Sentinel-2 and Landsat. For each pixel at a resolution of 10 meters, it produces embeddings: a 128-dimensional numerical representation of the characteristics of the area. The first 64 dimensions can be used independently as a smaller variant. A small classifier can be trained locally on this data, for example to map crops or land cover.

According to the authors, the 128-dimensional variant achieved the best result in 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2; the 64-dimensional variant led in 18 evaluations. The authors estimate the cost of the final pretraining run alone for OlmoEarth v1.3 Base at approximately 2 720 dollars for H100 rental. This is the cost of reproducing that run, not the price of using the tool.

OlmoEarth Embeddings Explorer works without an account or installation. It lets users switch between years, verify displayed changes using Sentinel-2 imagery, search for similar locations or train a small classifier using a few labeled points. The resulting map can be exported in GeoTIFF format.