A team of researchers from Adelaide University and the SmartSat Cooperative Research Center (CRC) has successfully uploaded and demonstrated NASA and IBM's open-source Prithvi Geospatial AI foundation model aboard two in-orbit platforms. This marks the first deployment of a geospatial foundation model in orbit.
Platforms and Deployment
The Prithvi model was deployed on two separate platforms:
- The South Australian government's Kanyini satellite
- The Thales Alenia Space IMAGIN-e payload aboard the International Space Station
A compressed version of the model was uploaded, enabling onboard data analysis before transmission to ground.
Model Capabilities and Testing
The model was tested for flood and cloud detection performance across different orbiting platforms and computing environments. Prithvi was chosen for its ability to generalize across Earth observation tasks and its open-source availability.
Demonstrated applications included:
- Burn scar prediction from the Gifford Fire (August 17, 2025)
- Flood extent mapping from Hurricane Helene (October 4, 2024)
"Having the model openly available reduced the time and effort required for the project."
— Dr. Andrew Du, project lead researcher and AI engineer at SmartSat CRC
Training and Development
Developed by IBM and NASA's Interagency Implementation and Advanced Concepts Team (IMPACT), Prithvi was trained on 13 years of data from the Harmonized Landsat and Sentinel-2 dataset. This dataset combines global geospatial data from NASA's Landsat and the European Space Agency's Sentinel-2 satellites. The model can be fine-tuned for tasks including flood plain mapping, disaster monitoring, and crop yield prediction.
Significance
Deploying foundation models in orbit allows advanced data analysis to be performed before data is transmitted to Earth, reducing bandwidth requirements. The model's flexibility enables new tasks to be implemented via small decoder packages rather than uploading entire new models.
"Sharing open-source AI models accelerates scientific and technological development."
— Kevin Murphy, NASA Chief Science Data Officer
Future Work
Researchers have suggested that foundation models could eventually allow natural language interaction with satellites, enabling operators to ask questions about onboard data or system status and receive conversational responses. NASA continues developing open-source foundation models for other scientific domains, including heliophysics (Surya, released in 2025), planetary science, astrophysics, and biological and physical sciences.