FedOrbit: Adaptive Personalized Federated Learning for Non-IID LEO Satellite Constellations

2026-08-10

AI Summary

This work addresses the challenges of federated learning in low Earth orbit (LEO) satellite constellations, where non-IID data distributions and irregular ground station visibility can lead to aggregation failure or imbalanced personalization. To overcome these issues, the authors propose FedOrbit, a novel framework that uniquely integrates orbital geometry into federated learning design. FedOrbit enables orbit-level continuous training via inter-satellite links and introduces several key mechanisms: class-aware hierarchical aggregation, quality-weighted feature aggregation, backhaul-rate damping, and adaptive feature decomposition based on inter-orbit class similarity. Evaluated across three remote sensing benchmarks under two non-IID partitioning schemes, FedOrbit achieves state-of-the-art accuracy in five out of six settings—outperforming the strongest baseline by up to 16.1% (Dirichlet) and 8.6% (pathological)—with the remaining setting showing a marginal gap (<0.9%). It also substantially reduces performance disparity across orbits.

Abstract

Federated learning (FL) in Low Earth Orbit (LEO) satellite constellations is affected by non-IID data and irregular ground-station visibility, both driven by orbital geometry. Global aggregation performs poorly when orbit-level class distributions are disjoint, while strong personalisation can be excessive when these distributions overlap. We present FedOrbit, which combines continuous orbit-level training over inter-satellite links, class-aware hierarchical aggregation, quality-weighted feature aggregation with return-rate dampening, and adaptive feature decomposition based on inter-orbit class similarity. Across three remote-sensing benchmarks and two non-IID partitions, FedOrbit achieves the highest accuracy in five of six settings and is within $0.9$ percentage points of the best result in the sixth. The gains over the strongest baseline reach $16.1$ percentage points under Dirichlet partitioning and $8.6$ under pathological partitioning, with the smallest per-orbit accuracy spread in five of six settings.

Problem

• Federated Learning
• Non-IID
• LEO Satellite Constellations
• Orbital Geometry
• Personalization

Innovation

• Federated Learning
• LEO Satellite Constellations
• Non-IID Data
• Adaptive Personalization
• Hierarchical Aggregation