Prithvi becomes the first geospatial foundation model to rotate directly into orbit

The NASA and IBM open model, trained on thirteen years of Landsat and Sentinel-2 images, was carried on the Australian Kanyini satellite and on a payload installed on board the ISS.

Until now, an observation satellite photographs, transmits, and the analysis is done on the ground. Prithvi reverses this chain: a compressed version of the geospatial foundation model developed jointly by NASA and IBM has been uploaded and then executed aboard two platforms in orbit, a first for this type of model.

The model was trained over thirteen years of imagery from NASA's Landsat and the European Space Agency's Sentinel-2 satellites. The tests focused on flood detection and cloud cover reconnaissance on board the South Australian government's Kanyini satellite and Thales Alenia Space's IMAGIN-e payload, installed on the International Space Station. The work involves the University of Adelaide, ESA's Φ-lab, the SmartSat research centre and NASA's Marshall Space Flight Center.

The operational interest lies in the format: the model can be specialized on a new task — mapping burned areas, forecasting agricultural yields — without having to send all the weights back to the satellite. Sorting out the useful images before transmission, rather than after, changes what can be done with a limited downlink.