Ai2 releases reusable satellite features for local mapping
Ai2 released OlmoEarth v1.3 alongside annual satellite embeddings and a browser explorer. The model turns a year of radar and multispectral observations into compact representations of land. Users can reuse those features for crop and land-cover mapping, compare locations and train small classifiers without repeatedly processing the original imagery.
Artificial Intelligence··Night
A year of satellite observations becomes reusable features
Ai2, the US nonprofit artificial-intelligence institute, released OlmoEarth v1.3 with precomputed annual satellite embeddings and a browser explorer. The embeddings are numerical representations that can be reused in different mapping tasks. The model combines Sentinel-1 radar observations with multispectral imagery from Sentinel-2 and Landsat. It reads a year of observations and represents each land pixel measuring 10 meters with 128 numbers. The first 64 dimensions can also be used independently. Users can work with these compact features instead of downloading and processing the original satellite scenes for every new crop or land-cover classification task. Internally, the model combines a vision transformer, an attention-pooling stage that aggregates land-patch information and a linear layer reducing the representation.[1], [2]
The browser builds maps from labeled points
The explorer lets users search for similar places, compare years and label points to train a small classifier. Results can be exported as GeoTIFF, a format that preserves a map’s geographical coordinates. Ai2 says the explorer needs neither an account nor a graphics processor. Earlier OlmoEarth Studio workflows involved task-specific fine-tuning, downloaded imagery and GPU access. Studio remains available for custom dates, modalities, resolution or model adaptation. The reusable features are distributed through Source Cooperative, a data-sharing platform; a small classifier can run locally on top of the downloaded embeddings.[1], [2]
Annual archives retain their satellite provenance
Ai2 describes global archives for 2023 through 2025; the Japanese coverage notes that some 2023 regions are still being filled. Coverage for 2017 through 2022 is limited to selected countries. Each annual archive occupies about 150 terabytes. Provenance tables identify the satellite scenes used for each tile measuring 20 kilometers and each year. Ai2 released model weights, pretraining data and pretraining code under the OlmoEarth Artifact License. Its mapping examples include smallholder crops in Kenya, brick kilns west of Dhaka in Bangladesh and fire damage in Indonesia. PLANTEUR adds a crop-type benchmark focused on tropical farming.[1], [2]
