H-FedSpace: Hierarchical Federated Learning for Uplink-Constrained LEO Satellite Networks
Low Earth orbit (LEO) satellite federated learning is constrained by short ground-station contacts, intermittent inter-satellite links (ISLs), and stale updates. Under these conditions, the central issue is not whether communication can be reduced in general, but whether uplink traffic to the ground can be reduced without materially weakening learning quality. This paper compares FedSpace and H-FedSpace under one shared implementation, one cached feature benchmark, and one common communication accounting scheme. The results show that H-FedSpace cuts logged uplink traffic by 80.0%, from 133.03 MB to 26.61 MB, while preserving final accuracy at 1.0000. It also achieves a much lower final loss, decreasing from 0.0838 to 0.0014, and records zero logged staleness at evaluation checkpoints. At the same time, these gains are accompanied by 133.03 MB of ISL traffic, a 20.0% increase in total logged communication, and longer runtime. The evidence therefore supports H-FedSpace in the narrower operating regime where satellite-to-ground uplink is the binding resource. The interpretation remains cautious because, in the current implementation, H-FedSpace performs local training on all 12 satellites in each round, whereas FedSpace trains only a randomly selected satellite per round.