AI2 released OlmoEarth 1.3 models and precomputed satellite embeddings covering the globe. They represent spatial information at 10-meter resolution in 128 or 64 dimensions, with model weights, code, and data also made available. Users can test a small classifier in a browser explorer or download data for their own analysis, reducing the burden of training a satellite imagery model from scratch. The team reported that the 128-dimensional model performed best on 20 of 27 evaluations. Those are the provider’s evaluations; separate validation is needed when the region, time period, or analysis target changes.
Back to Tech
Tech
OlmoEarth 1.3 releases global satellite embeddings
Precomputed spatial representations can connect to a small classifier to get regional analysis started.