LunarFL: Federated Terrain Learning Across Distributed Planetary Surface Assets Under Intermittent, Thermally-Gated Connectivity

As the Artemis and Moon-to-Mars programs advance toward a sustained lunar presence, semi-autonomous surface assets must perceive terrain onboard while sharing a scarce, thermally constrained link budget. Federated learning (FL) fits this setting: assets improve a shared perception model by exchanging updates rather than raw imagery. Existing FL studies, however, assume near-persistent connectivity and evaluate on synthetically partitioned data. This paper reports what is, to the author’s knowledge, the first evaluation of federated terrain segmentation on real planetary rover imagery under a physically motivated model of intermittent, thermally-gated connectivity. A naturally non-independent-and-identically-distributed (non-IID) three-client federation is built from the AI4Mars Curiosity dataset and the synthetic Artificial Lunar Landscape dataset; its heterogeneity is measured, not assumed (mean pairwise Jensen-Shannon divergence 0.42). A SegFormer-B0 model is trained with FedAvg while an aggregation probability pagg and a thermal duty cycle gate every round. Over 113 runs with up to five seeds per condition, accuracy declines monotonically with aggregation scarcity: federation recovers 88.8% of the gap between local-only (0.363 mIoU) and centralized (0.897) training at pagg = 1, 78.6% at 0.3, and 5.9% at 0.05. The deficit scales as paggb with b = −0.650 (95% CI [−0.755, −0.545]), steeper than the optimization-limited 1/√pagg reference and consistent with heterogeneity-driven client drift. Neither a proximal term nor server momentum recovers the scarce regime, and the temporal structure of outages, not the mean rate alone, governs predictability. Uplink volume is reported alongside accuracy at every operating point.