Bringing Federated Learning to Space

TL;DR

This paper tackles the bottleneck of data downlink in dense LEO satellite constellations by enabling on-board collaborative learning via federated learning. It introduces a space-ification framework that adapts terrestrial FL algorithms (FedAvg, FedProx, FedBuff) to orbital constraints and augments them with scheduling and intra-cluster communication. A comprehensive simulation study across 768 constellation configurations demonstrates that space-ified FL achieves over 80% accuracy on FEMNIST, with training times for a 100-satellite constellation reduced by up to 9x. These results provide a practical path for scalable, autonomous, data-driven satellite operations with distributed onboard learning.

Abstract

As Low Earth Orbit (LEO) satellite constellations rapidly expand to hundreds and thousands of spacecraft, the need for distributed on-board machine learning becomes critical to address downlink bandwidth limitations. Federated learning (FL) offers a promising framework to conduct collaborative model training across satellite networks. Realizing its benefits in space naturally requires addressing space-specific constraints, from intermittent connectivity to dynamics imposed by orbital motion. This work presents the first systematic feasibility analysis of adapting off-the-shelf FL algorithms for satellite constellation deployment. We introduce a comprehensive "space-ification" framework that adapts terrestrial algorithms (FedAvg, FedProx, FedBuff) to operate under orbital constraints, producing an orbital-ready suite of FL algorithms. We then evaluate these space-ified methods through extensive parameter sweeps across 768 constellation configurations that vary cluster sizes (1-10), satellites per cluster (1-10), and ground station networks (1-13). Our analysis demonstrates that space-adapted FL algorithms efficiently scale to constellations of up to 100 satellites, achieving performance close to the centralized ideal. Multi-month training cycles can be reduced to days, corresponding to a 9x speedup through orbital scheduling and local coordination within satellite clusters. These results provide actionable insights for future mission designers, enabling distributed on-board learning for more autonomous, resilient, and data-driven satellite operations.

Figures (10)

• Figure 1: Scheduling client selection in FL rounds reduces total aggregation time by prioritizing satellites with shorter combined initial contact and revisit times. Satellites that make initial contact later can still be selected if their revisit times are faster, reducing idle time by nearly half compared to selecting the first satellites to contact a ground station. The figures above show simulations of a Walker-Star constellation with 4 clusters of 6 satellites each and access to 3 ground stations.
• Figure 2: Visual representation of intra- and inter-cluster links (Walker-star topology). Intra-cluster communication occurs between adjacent satellites of the same cluster, and may be persistent. In contrast, inter-cluster communication occurs across neighboring clusters, typically over a finite communication window in which the satellites are within range.
• Figure 3: Ground station facilities used for parameter sweep simulations, with 13 communication locations inspired from the International Ground Station (IGS) network wulder_global_2016. The different locations are set to be in: Sioux Falls (US), Sanya (China), Johannesburg (South Africa), Cordoba (Argentina), Tromso (Norway), Kashi (China), Beijing (China), Neustrelitz (Germany), Parepare (Indonesia), Alice Springs (Australia), Fairbanks (US), Prince Albert (Canada), and Shadnagar (India).
• Figure 4: An example set of heatmaps for FedAvg with space-ification, outlining performances in accuracy, FL round durations, and idle time. Even with FedAvg, we see that with large enough constellation sizes and ground station networks, enough opportunities for access windows can be made to reach convergence. Proper comparisons against performance of other algorithms can be found in the larger span of heatmaps in \ref{['fig:accs', 'fig:base_durations', 'fig:idleavg_normal']}, highlighting the importance of the augmentations on FL algorithm performance.
• Figure 5: Heatmaps depicting the maximum accuracy reached in the training of each satellite simulations, testing configuration parameters of varying #s of clusters, #s of satellites per cluster, and #s of ground stations available to connect in the network. Multiple FL algorithms were implemented and tested, specifically FedAvg, FedProx, and FedBuff, and measured against versions of the same algorithm but with scheduled and Intra SL enabled communications. All algorithms if provided enough aggregation opportunities could reach more than 80% of accuracy. This was typically possible by optimizing for existing access windows through scheduling, or through the addition of more communication points through larger ground station networks or ISL enabled communications.