FedHRA: A Joint Optimization Framework for Fast Convergent Decentralized Federated Learning in LEO Satellite Networks

Low Earth Orbit (LEO) satellites are playing an important role in earth observation, providing valuable images for training machine learning (ML) models used in tasks such as environmental monitoring and pattern recognition. However, due to unstable communication links and limited downlink bandwidth, it is economically impractical to transmit all raw images to ground stations (GSs) for model training. Federated learning (FL), a privacy-preserving distributed machine learning method, can reduce the communication overhead by exchanging model parameters. Generally, FL needs a fixed central server to aggregate the global model, which is challenging in LEO satellite networks, given the dynamic nature of satellite orbits. To overcome this, we propose a decentralized federated learning (DFL) framework that enables efficient model aggregation through satellite collaboration. Specifically, the proposed framework, named FedHRA, is based on model-agnostic meta-learning (MAML), which jointly optimizes hyperparameters and resource allocation to mitigate straggler effect and address statistical heterogeneity. Extensive numerical results on MNIST and CIFAR-10 datasets demonstrate that FedHRA achieves shorter learning time and higher model accuracy compared to the benchmark frameworks.