Resource-Aware Federated Mixture-of-Experts with Adaptive Pruning for Onboard Learning in LEO Satellite Constellations

Mohamed Shaaban; Mohamed Elmahallawy; Marius Bernahrndt; Tobias Hecking

cs.LGSep 25, 2026

Resource-Aware Federated Mixture-of-Experts with Adaptive Pruning for Onboard Learning in LEO Satellite Constellations

Authors: Mohamed Shaaban, Mohamed Elmahallawy, Marius Bernahrndt, Tobias Hecking

Organizations: School of Engineering and Applied Science, Washington State University, Richland, WA 99354, USA · Institute for Software Technology, German Aerospace Center (DLR), Porz, 51147 Cologne, Germany

Abstract

Low-Earth-orbit (LEO) satellites are increasingly expected to perform onboard learning for applications such as disaster response and environmental monitoring. However, conventional federated learning (FL) is ill-suited to onboard satellite learning, as it assumes computational, memory, and communication resources beyond the capabilities of resource-constrained LEO platforms, often necessitating the transmission of raw imagery to ground stations. We present COSMIC-FL, a resource-aware FL framework for efficient onboard learning in LEO satellite constellations. COSMIC-FL introduces two complementary Mixture-of-Experts (MoE) architectures: a Sliced design that shares backbone representations while activating task-specific channel subsets, and a Modular design that employs lightweight gating to route inputs to physically separated expert networks. A semantic class-to-expert mapping enables each satellite to train, update, and communicate only the expert paths relevant to its local data. To further improve efficiency, COSMIC-FL integrates staged optimization with three structured pruning strategies: server-side pruning, client-side fixed-ratio pruning with mean-vote aggregation, and adaptive client-side per-layer pruning based on aggregated importance and a MAD-based gap criterion. Combined with semantic expert routing, these techniques jointly adapt computation and model sparsity to both data semantics and layer importance, yielding a favourable accuracy--efficiency trade-off for heterogeneous space platforms. Experiments on six image classification benchmarks under highly non-i.i.d. settings show that COSMIC-FL maintains competitive accuracy while reducing communication, computation, and energy consumption by up to 80% over SOTA FL methods. We further validate COSMIC-FL on an NVIDIA Jetson AGX Orin, confirming its efficiency gains under realistic embedded deployment constraints.

Figures & tables

Figure 1: Overview of the four stages of the COSMIC-FL framework.

Method (# Satellites)	CIFAR-10	AID	EuroSAT	RESISC45	UC Merced
FedAvg (10)	88.50	79.39	97.41	83.73	83.81
FedAvg (50)	82.03	62.88	94.48	72.78	56.51
FedAvg (100)	75.46	53.24	91.07	60.40	49.21
FedProx (10)	86.99	78.80	97.30	83.02	83.17
FedProx (50)	81.55	64.80	94.56	73.67	55.87
FedProx (100)	74.61	53.43	90.22	62.82	48.89

Table 1: Accuracy (%) under IID settings with different numbers of satellites.

Method	$\alpha{=}0.1$	$\alpha{=}0.3$	$\alpha{=}1.0$	$\alpha{=}10.0$
FedAvg	81.33	90.89	93.74	94.41
FedProx	83.00	90.63	93.37	94.48
CentMoE	86.89	94.07	95.07	95.26
PM-MoE	11.11	92.70	95.22	95.26
FedDropout ( $r{=}0.75$ )	64.72	75.13	79.19	79.94
HeteroFL ( $r{=}0.75$ )	74.98	81.70	86.17	89.54

Table 2: Impact of data heterogeneity (Dirichlet-

\alpha

) on test accuracy (EuroSAT, 50 clients). Lower

\alpha

= more non-i.i.d (see Supplementary Material F for more results).

Figure 2: Inference cost (GFLOPs per sample) by method and dataset (AID, EuroSAT, RESISC45). Methods: CentMoE, COSMIC-FL (Modular), COSMIC-FL (Sliced), and their server-, client-, and adaptive-pruning variants.

Model	Prune	Server		Client		Adaptive
Model	Prune	P	Acc	P	Acc	P	Acc
COSMIC-FL (Sliced)	0	12.4	87.56	12.4	87.56	12.4	87.56
	10	11.5	83.74	11.5	82.07	11.5	83.15
	30	8.2	81.15	8.2	81.04	8.4	81.93
	50	5.6	80.67	5.6	82.04	5.8	82.22
	70	3.5	80.59	3.5	79.67	4.0	80.89

Table 3: Pruning strategy and ratio analysis for COSMIC-FL (Sliced and Modular) on EuroSAT with 50 satellites. Results report retained parameters (P, millions) and accuracy (%). Additional results are provided in Supplementary Material F.

	EuroSAT				RESISC45
Method	Acc	Lat.	Mem	Pow	Acc	Lat.	Mem	Pow
FedAvg	94.3	3.66	38.2	4.90	72.1	3.66	38.3	4.96
FedProx	94.6	3.65	38.2	4.99	72.1	3.66	38.3	5.04
FedDropout	79.7	4.52	39.1	4.34	42.8	4.53	40.0	4.59
HeteroFL	89.2	4.54	26.9	4.12	58.4	4.54	26.9	4.12
PM-MoE	96.0	22.9	190.3	4.47	68.6	22.9	193.8	4.47

Table 4: Jetson AGX Orin inference performance on EuroSAT and RESISC45. Reported metrics include classification accuracy (%), latency (ms), peak allocated GPU memory (MB), and average board power (W).

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Impl.	Target	$z{=}1$	$z{=}2$	$z{=}3$	$z{=}10$	$z{=}20$
EuroSAT	Modular	30%	20.44 / 78.30	20.44 / 78.30	20.44 / 78.30	20.44 / 78.30	16.14 / 77.63
EuroSAT	Modular	50%	35.42 / 79.11	35.42 / 79.11	35.42 / 79.11	35.42 / 79.11	24.54 / 77.78
EuroSAT	Modular	70%	49.97 / 78.93	49.97 / 78.93	49.97 / 78.93	47.01 / 79.00	33.54 / 78.33
EuroSAT	Sliced	30%	34.52 / 80.19	34.52 / 80.19	34.52 / 80.19	29.79 / 82.22	18.00 / 79.81
EuroSAT	Sliced	50%	53.30 / 79.52	53.30 / 79.52	53.30 / 79.52	46.47 / 81.63	29.94 / 77.37
EuroSAT	Sliced	70%	65.05 / 76.93	65.05 / 76.93	65.05 / 76.93	56.29 / 79.56	39.15 / 79.96

Appendix

Table 5: MAD-threshold sensitivity of adaptive pruning for the reported threshold sweep. Each cell reports achieved parameter pruning and test accuracy.

Num. Experts	EuroSAT				RESISC45
	Modular		Sliced		Modular		Sliced
	Params (M)	Acc. (%)	Params (M)	Acc. (%)	Params (M)	Acc. (%)	Params (M)	Acc. (%)
3	19.8	88.19	7.4	87.96	19.9	65.82	7.4	63.67
5	33.0	73.56	9.0	73.85	33.2	55.56	9.0	40.27
7	46.3	67.74	10.6	73.44	46.5	49.07	10.7	37.98
Best	19.8	88.19	7.4	87.96	19.9	65.82	7.4	63.67

Appendix

Table 6: Impact of the number of experts on COSMIC-FL modular and sliced variants evaluated on EuroSAT and RESISC45 with 50 clients. Results report parameter counts (M) and classification accuracy (%).

	Modular			Sliced
Num. Experts	Params (M)	Acc. ${}_{\text{short}}$ (%)	Acc. ${}_{\text{long}}$ (%)	Params (M)	Acc. ${}_{\text{short}}$ (%)	Acc. ${}_{\text{long}}$ (%)
3	17.6	21.00	29.05	7.6	22.30	29.15
5	29.3	18.37	25.21	9.2	15.46	19.31
7	41.0	15.52	21.59	10.9	10.22	13.20
10	58.6	14.03	19.96	13.4	9.13	12.72

Appendix

Table 7: Impact of the number of experts on COSMIC-FL modular and sliced variants evaluated on CIFAR-100 (100 classes) under a short and a long training budget. Results report parameter counts (M) and classification accuracy (%).

	$\alpha{=}0.1$		$\alpha{=}0.3$		$\alpha{=}1.0$		$\alpha{=}10$
Stage	Acc.	$\delta$	Acc.	$\delta$	Acc.	$\delta$	Acc.	$\delta$
Gate pretraining (init.)	8.07	–	8.22	–	8.52	–	8.48	–
+ Expert specialization	53.15	+45.08	73.48	+65.26	75.85	+67.33	76.37	+67.89
+ Gate fine-tuning	57.04	+3.89	76.52	+3.04	80.19	+4.34	80.07	+3.70
+ Head fine-tuning	67.67	+10.63	83.26	+6.74	86.52	+6.33	87.78	+7.71

Appendix

Table 8: Test accuracy (%) after each training stage of COSMIC-FL (Sliced) on EuroSAT with 50 clients under varying Dirichlet heterogeneity.

\delta

= gain over the preceding stage.

Panel of Figure 3 — Figure 3: Test accuracy during the expert-specialization phase with 50 rounds and 50 clients on EuroSAT, AID, and RESISC45. Curves compare FedAvg, FedProx, CentMoE, PM-MoE, and the COSMIC-FL variants.

	Modular		Sliced
Metric	Sparse	Normal	Sparse	Normal
10 clients
Val. accuracy (%)	76.22	75.78	80.41	78.85
Expert participation rate (%)	60.0	100	56.7	100
Experts trained / client / round	1.80	3.00	1.70	3.00
Params transferred / client / round (M)	11.89	19.81	4.19	7.39

Appendix

Table 9: Sparse expert participation vs. full (normal) expert training for COSMIC-FL (EuroSAT, Dirichlet

\alpha{=}0.1

, 3 experts, 15 rotation rounds) across 10, 50, and 100 clients. With sparse participation, a client only trains experts for which it holds sufficient local samples, so it does not have to train the whole expert set each round. As the constellation scales, each client holds fewer samples, so fewer experts pass the ownership threshold: the communication saving grows (40%

\to

89% for Modular, 43%

\to

88% for Sliced) while validation accuracy is maintained or improved.

Method	EuroSAT				RESISC45
	$\alpha{=}0.1$	$\alpha{=}0.3$	$\alpha{=}1.0$	$\alpha{=}10.0$	$\alpha{=}0.1$	$\alpha{=}0.3$	$\alpha{=}1.0$	$\alpha{=}10.0$
FedAvg	81.33	90.89	93.74	94.41	57.69	66.64	69.24	72.2
FedProx	83.00	90.63	93.37	94.48	62.27	68.42	70.98	73.27
CentMoE	86.89	94.07	95.07	95.26	61.49	66.60	66.96	67.36
PM-MoE	11.11	92.70	95.22	95.26	29.91	45.44	65.91	71.56
FedDropout ( $r{=}0.75$ )	64.72	75.13	79.19	79.94	33.28	36.84	41.09	43.73

Appendix

Table 10: Impact of data heterogeneity (Dirichlet-

\alpha

) on test accuracy (EuroSAT & RESISC45, 50 clients). Lower

\alpha

= more non-i.i.d.

Panel of Figure 4 — Figure 4: Per-client parameters trained and uploaded per round (millions) on EuroSAT, AID, and RESISC45. Each phase communicates only the parameters being optimized. CentMoE and PM-MoE train full models, while COSMIC-FL (Sliced/Modular) separates expert, gating, and head components.

	Payload / client / round (M)			Total uplink (G)
Method	Rotation	Gate	Head	50 clients	100 clients
FedAvg (full model)	5.82	–	–	14.5	29.1
FedProx (full model)	5.82	–	–	14.5	29.1
CentMoE	7.40	–	–	18.5	37.0
PM-MoE †	5.02	–	–	58.0	195.4
FedDropout ( $r{=}0.75$ )	3.27	–	–	8.2	16.4

Appendix

Table 11: Total uplink communication across all training phases on EuroSAT, computed from measured per-client per-round payloads under the reference schedule (

T_{E}{=}50

T_{G}{=}20

T_{H}{=}20

). Sparse-expert variants scale rotation communication by the measured expert-participation rate

\rho

(Table 9 ). “–” denotes methods without separate gate or head phases. † PM-MoE communication is derived from the implementation, where Phase 2 aggregates the full model including body, heads, and gate, causing client-dependent scaling.

Method (# Satellites)	CIFAR-10	AID	EuroSAT	RESISC45	UC Merced
FedDropout ( $r{=}0.25$ , 10)	31.53	14.33	33.72	12.42	5.08
FedDropout ( $r{=}0.25$ , 50)	21.75	6.30	19.59	6.62	7.62
FedDropout ( $r{=}0.25$ , 100)	14.41	5.56	18.24	4.47	4.76
FedDropout ( $r{=}0.5$ , 10)	65.01	37.44	66.33	35.04	13.65
FedDropout ( $r{=}0.5$ , 50)	43.92	13.46	59.98	16.42	9.52
FedDropout ( $r{=}0.5$ , 100)	38.71	6.90	47.13	9.41	8.57

Appendix

Table 12: Full capacity-fraction sweep for FedDropout and HeteroFL: accuracy (%) under IID settings with different numbers of satellites.

Model	Pruning	EuroSAT						RESISC45
		Server		Client		Adaptive		Server		Client		Adaptive
		P	Acc	P	Acc	P	Acc	P	Acc	P	Acc	P	Acc
COSMIC-FL (Sliced)	0	12.4	87.56	12.4	87.56	12.4	87.56	12.5	64.07	12.5	64.07	12.5	64.07
	10	11.5	83.74	11.5	82.07	11.5	83.15	11.5	46.62	11.5	46.58	11.5	46.40
	30	8.2	81.15	8.2	81.04	8.4	81.93	8.3	44.91	8.3	44.31	8.5	46.96
	50	5.6	80.67	5.6	82.04	5.8	82.22	5.7	44.09	5.7	45.33	6.3	43.27

Appendix

Table 13: Impact of pruning ratio and strategy (server-side, client-side, adaptive) on COSMIC-FL (Sliced and Modular) on EuroSAT and RESISC45 with 50 satellites. Results report parameters (P, millions) and test accuracy (%).

Dataset	Impl.	Variant	Acc. (%)	Ach. pruning (%)	Latency (ms)	Throughput (img/s)	Peak memory (MB)
EuroSAT	Modular	Dense	86.22	0.00	10.59	94.43	83.45
EuroSAT	Modular	Adaptive	79.11	35.42	11.21	89.17	61.52
EuroSAT	Sliced	Dense	86.26	0.00	23.23	43.05	105.05
EuroSAT	Sliced	Adaptive	79.52	53.30	21.37	46.80	42.28
RESISC45	Modular	Dense	63.60	0.00	10.70	93.50	84.78
RESISC45	Modular	Adaptive	58.87	39.81	11.23	89.07	55.01

Appendix

Table 14: Dense and adaptively pruned proposed models on the Jetson AGX Orin. Values are averaged over two independent runs.

Dataset	Impl.	Memory	Latency	Throughput	Power
EuroSAT	Modular	$-26.28\%$	$+5.90\%$	$-5.57\%$	$-3.96\%$
EuroSAT	Sliced	$-59.75\%$	$-7.98\%$	$+8.70\%$	$-10.75\%$
RESISC45	Modular	$-35.11\%$	$+4.96\%$	$-4.73\%$	$-6.20\%$
RESISC45	Sliced	$-57.74\%$	$-7.36\%$	$+7.94\%$	$-11.11\%$

Appendix

Table 15: Relative hardware effect of adaptive pruning with respect to the corresponding dense implementation.

Explore similar work

FedOrbit: Adaptive Personalized Federated Learning for Non-IID LEO Satellite Constellations
Aug 10, 2026Satwat Bashir, Tasos Dagiuklas, Muddesar IqbalHierarchical Federated Learning Federated Learning Aggregation
School of Engineering London South Bank University, London, UK
ORBITALIF: An Efficient Spiking Federated Learning Framework for Onboard Cloud Removal
Aug 25, 2026Bohan Zhang, Chenyu Xu, Yijie Mao +1Remote Sensing Federated Learning
School of Information Science and Technology, ShanghaiTech University, Shanghai, China
MLFCIL: A Multi-Level Forgetting Mitigation Framework for Federated Class-Incremental Learning in LEO Satellites
Mar 14, 2026Heng Zhang, Xiaohong Deng, Sijing Duan +4Federated Continual Learning Federated Learning
School of Information Engineering, Gannan University of Science and Technology, Ganzhou 341000, China · Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China · Department of Electrical and Computer Engineering, Queen University, Kingston, ON K7L 3N6, Canada +2

Aug 10, 2026J/K move · Enter open · S save

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

Satwat Bashir, Tasos Dagiuklas, Muddesar Iqbal

School of Engineering London South Bank University, London, UK

Hierarchical Federated Learning Federated Learning Aggregation Non-IID Federated Learning LEO Satellite Networks Federated Learning

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

ORBITALIF: An Efficient Spiking Federated Learning Framework for Onboard Cloud Removal

MLFCIL: A Multi-Level Forgetting Mitigation Framework for Federated Class-Incremental Learning in LEO Satellites