A groundbreaking advancement in artificial intelligence has seen the Prithvi Geospatial AI foundation model successfully deployed aboard two in-orbit platforms, marking it as the first geospatial foundation model to operate in space. This significant achievement, a collaborative effort involving NASA and IBM, was spearheaded by a research team from Adelaide University and the SmartSat Cooperative Research Center in South Australia, unlocking new capabilities for Earth observation.

The deployment involved uploading a compressed version of the Prithvi model to the South Australian government’s Kanyini satellite and the Thales Alenia Space IMAGIN-e payload, which is affixed to the International Space Station. Researchers rigorously tested the model’s performance in critical areas such as flood and cloud detection across these distinct orbiting environments. Trained on an extensive dataset comprising 13 years of global geospatial information from NASA’s Landsat and ESA’s Sentinel-2 satellites, Prithvi is designed to facilitate a wide array of Earth observation tasks, including mapping flood plains, monitoring natural disasters, predicting crop yields, and identifying burn scars from wildfires. The open-source nature of the model proved invaluable, with lead researcher Dr. Andrew Du noting that its availability saved considerable time and effort in development.

Foundation models are a class of AI trained on massive amounts of unlabeled data, enabling them to discern complex patterns that might elude human analysis before being fine-tuned for specific applications with smaller, labeled datasets. Deploying such a model in orbit offers substantial advantages. Processing and analyzing data directly in space allows researchers to gain insights more rapidly and efficiently. Furthermore, this approach overcomes the bandwidth limitations often faced by active satellites; instead of uploading an entire updated model for new tasks, only a compact ‘decoder package’ is required, significantly reducing data transfer needs. Kevin Murphy, Chief Science Data Officer at NASA, underscored the importance of open-source initiatives like Prithvi, stating that sharing these tools accelerates scientific and technological development.

Looking ahead, the successful orbital demonstration of Prithvi is seen as an early indicator of how foundation models could transform Earth observation and broader space science. Beyond advanced data analysis, there is potential for these models, including large language models (a type of foundation model), to enable operators to interact with satellites using natural language, allowing for conversational queries about onboard data or system status. The team behind Prithvi continues to develop open-source foundation models, with a heliophysics model named Surya already released, and future plans for models in planetary science, astrophysics, and biological and physical sciences, promising a new era of intelligent space exploration.