Decentralised Semi-supervised Onboard Learning for Scene Classification in Low-Earth Orbit

A new generation of satellites is currently bringing hardware suitable for machine learning (ML) onboard spacecraft into Earth orbit. Recent works [1] explored the possibility to train ML models in a distributed manner onboard satellite constellations. Distributed onboard training brings the potential to reduce communication requirements, operational cost and time, and improve autonomy by sharing ML models, trained close to the sensors, instead of the collected data. While previous missions have demonstrated the ability to perform inference onboard spacecraft for data processing [2], training onboard presents additional challenges. Convincingly addressing operational constraints is crucial, as the computational cost of training is significantly higher, and the lack of labeled examples during the mission can often be prohibitive. In this work, we investigate the training of an ML model onboard a satellite constellation for scene classification.