A recent arXiv paper describes the development and in-orbit demonstration of an onboard artificial intelligence system for marine anomaly detection on the PhiSat-2 satellite. The authors present their experience moving from simulation-based development to actual deployment in space, offering a concrete case study of onboard AI in Earth observation.
The motivation is straightforward: satellites typically downlink large volumes of raw data, which can be slow and bandwidth-hungry. By processing data directly on the satellite, AI can identify relevant events in real time and only send back the most useful information. This improves responsiveness and makes more efficient use of limited communication links.
The paper details the practical lessons learned during this process, from building and testing the model in simulation to operating it in orbit. While the abstract is brief, it highlights the value of validating AI systems in realistic conditions before launch, and the potential for onboard processing to become a standard tool in future Earth observation missions.