Low-Earth orbit (LEO) satellites have been prosperously deployed for various Earth observation missions due to its capability of collecting a large amount of image or sensor data. However, traditionally, the data training process is performed in the terrestrial cloud server, which leads to a high transmission overhead. With the recent development of LEO, it is more imperative to provide ultra-dense LEO constellation with enhanced on-board computation capability. Benefited from it, we have proposed a collaborative federated learning over LEO satellite constellation (FedLEO). We allocate the entire process on LEOs with low payload inter-satellite transmissions, whilst the low-delay terrestrial gateway server (GS) only takes care for initial signal controlling. The GS initially selects an LEO server, whereas its LEO clients are all determined by clustering mechanism and communication capability through the optical inter-satellite links (ISLs). The re-clustering of changing LEO server will be executed once with low communication quality of FedLEO. In the simulations, we have numerically analyzed the proposed FedLEO under practical Walker-based LEO constellation configurations along with MNIST training dataset for classification mission. The proposed FedLEO outperforms the conventional centralized and distributed architectures with higher classification accuracy as well as comparably lower latency of joint communication and computing.

翻译：低地球轨道（LEO）卫星因其能够收集大量图像或传感器数据的能力，已被广泛部署于各类地球观测任务。然而，传统的数据训练过程在地面云服务器上执行，导致较高的传输开销。随着近期LEO技术的发展，为超密集LEO星座提供增强的星载计算能力变得更为迫切。借此优势，我们提出了一种基于LEO卫星星座的协同联邦学习框架（FedLEO）。我们将整个处理过程分配至各LEO卫星上，仅涉及低有效载荷的星间传输，而低延迟的地面网关服务器（GS）仅负责初始信号控制。GS首先选定一颗LEO服务器，而其LEO客户端则全部通过聚类机制以及经由光学星间链路（ISL）评估的通信能力来确定。当FedLEO的通信质量下降时，将对变化的LEO服务器执行一次重新聚类。在仿真中，我们基于实际Walker构型的LEO星座配置，结合用于分类任务的MNIST训练数据集，对所提出的FedLEO进行了数值分析。与传统的集中式和分布式架构相比，所提出的FedLEO在分类精度上表现更优，同时联合通信与计算的延迟也相对更低。