In 2023, NASA and IBM Research openly released Prithvi, described by NASA as its first open-source geospatial AI foundation model for Earth-observation data. Development began in January 2023, and the model was published on Hugging Face in the summer, with NASA’s announcement and the public release coming in early August. NASA and IBM called it the largest geospatial foundation model available on Hugging Face at the time.
Prithvi is a temporal vision transformer pre-trained on NASA’s Harmonized Landsat and Sentinel-2 (HLS) dataset, which stitches together surface-reflectance imagery from US satellites into a consistent record covering the same ground every few days at 30-meter resolution. A foundation model is pre-trained once on a large pool of unlabeled data and then fine-tuned for many downstream tasks, so Prithvi can be adapted to jobs like tracking land use, mapping floods and burn scars, and estimating crop yields without training a fresh model from scratch each time.
The release reflected a deliberate open-science strategy: by giving the pretrained weights away, NASA and IBM aimed to let researchers and agencies build their own applications on top. Collaborators included Clark University, the European Space Agency, the US Geological Survey, and Oak Ridge National Laboratory, and the team later followed with larger Prithvi versions trained on global data.