Earth observation using small satellites operating in low Earth orbit (LEO) has received increasing interest over the past years. With high-resolution sensors amassing up to terabytes of data per day in large satellite constellations, onboard processing using neural network models is considered to manage downlink communication capacity. A question frequently overlooked, however, is how to update these models using the - often orders of magnitude smaller - uplink capacity. Yet, model updates are crucial to allow for flexible long-term missions and commercial Earth-observation-as-a-service models. In this paper, we propose an efficient communication protocol for model updates based on incremental transmission of prioritized parameters and vector quantization. The evaluation results show that our approach significantly outperforms existing incremental model update transmission schemes.
