Hierarchical Aggregation for Distributed Federated Learning in Large-Scale Satellite Networks
Chengbiao Fu, Haotong Wang, Jun Du, Xiangwang Hou, Jintao Wang · 2025
Low Earth Orbit (LEO) satellites are increasingly deployed for a wide range of Earth observation missions. As the scale of LEO satellite constellations continues to expand, new challenges arise in communication efficiency, coordination, and scalable model aggregation. These challenges are particularly critical in the context of distributed machine learning for next-generation wireless networks. Traditional Federated Learning (FL) approaches typically rely on parameter server aggregation at ground stations. This centralized strategy often leads to high latency and slow model convergence, especially in dynamic satellite environments. To address these limitations, we propose a cluster-based hierarchical FL framework tailored for large-scale, time-varying LEO satellite networks. The proposed framework dynamically selects cluster heads based on network centrality and forms local inter-satellite clusters within a limited hop range. By leveraging the topological proximity among neighboring satellites, intra-cluster weighted aggregation is performed to accelerate local model convergence. Ground stations then connect to visible satellites within a constrained elevation angle to execute global aggregation across clusters, enabling a temporally driven and spatially distributed learning process. Moreover, the framework incorporates the Age of Information (AoI) as a key metric to assess the freshness of model parameters, ensuring that the most up-to-date models are prioritized during dissemination and aggregation. Simulation results validate the effectiveness of the proposed method in improving both convergence speed and model accuracy under realistic satellite network dynamics.