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

FedOrbit: Adaptive Personalized Federated Learning for Non-IID LEO Satellite Constellations 1st Satwat Bashir,

2nd Tasos Dagiuklas,

3rd Muddesar Iqbal

arXiv:2608.09687v1 [cs.LG] 10 Aug 2026

School of Engineering London South Bank University, London, UK {bashis11, tdagiuklas, m.iqbal}@lsbu.ac.uk

geometry, producing imbalanced participation. In the constellation considered in this work, approximately 47% of training rounds have no orbit visible to the GS, and the least-visited orbit participates in fewer than one-third as many rounds as the most-visited orbit [8]. As the class distribution of an orbit is linked to its ground track and its visibility schedule is determined by orbital geometry, class contribution and timing are systematically correlated rather than independent. This correlation between data heterogeneity and participation is the defining property of the LEO FL setting. This paper presents FedOrbit, a personalised FL framework that addresses the correlation between data and participation in LEO constellations. The contributions are: (i) identification of two failure modes induced by this correlation, namely global-model collapse on pathological partitioning and over-personalisation on Dirichlet partitioning; (ii) continuous orbit-level training over inter-satellite links, so every orbit refines the I. I NTRODUCTION model in every round; (iii) class-aware hierarchical Low Earth Orbit (LEO) satellite constellations are now aggregation, routing classifier updates by per-class data an important platform for global Earth observation [1], ownership; (iv) quality-weighted feature aggregation [2], and on-board processing makes it possible to with return-rate dampening, limiting the influence of train and run machine learning models directly on the an orbit returning from a long absence; (v) adaptive satellite [3], [4]. Transmitting raw observation data to feature decomposition, blending each orbit’s personal a ground station for centralised training is impractical feature extractor toward the global one at a rate set given the intermittent nature of satellite-to-ground links, by inter-orbit class similarity. We evaluate FedOrbit on data privacy requirements, and bandwidth constraints. three remote-sensing classification benchmarks under Federated learning (FL) [5] reduces the need for raw- both pathological and Dirichlet (α = 0.5) partitions, data transfer: each satellite trains locally on its own with the largest gains on the partition where existing observations and exchanges only model parameters with methods perform worst. a ground station (GS). II. R ELATED W ORK FL in LEO constellations differs fundamentally from FL under non-IID data. FedAvg [5] degrades the terrestrial setting. In a Walker Delta constellation, under heterogeneous distributions [9]; FedProx [10], each orbit follows a distinct ground track and observes SCAFFOLD [11], and FedNova [12] use proximal rega different geographic region, producing orbit-level ularisation, control variates, or local-step normalisation. class distributions that are inherently non-IID [6], [7]. These methods produce a single global model but do Orbit-to-GS visibility is governed by the same orbital not model the coupling between local data and visibilitydependent participation. This work has received funding from the European Commission Personalised FL. Ditto [13] adds an ℓ2 penalty toward under the Horizon Europe MSCA programme (HORIZON-MSCA2024-SE-01-01), Grant Agreement No. 101236523 (AeroNet project). the global model; APFL [14] interpolates local and

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, classaware 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. Index Terms—Federated learning, LEO satellite constellations, personalised aggregation, non-IID data, nonterrestrial networks.

global models; other methods use regularised inner loops [15], meta-learned initialisations [16], or split representations [17]. These methods do not account for bias in the global reference model caused by visibilitydependent participation. FL in non-terrestrial networks. Prior LEO FL work has addressed scheduling under irregular visibility [8], semi-asynchronous aggregation [18], aerial parameter servers [19], [20], decentralised intra-orbit aggregation [21], over-the-air computation [22], system heterogeneity [23], and decentralised personalisation [24]. The common focus is communication architecture, convergence speed, and scheduling. To our knowledge, prior work does not jointly address orbit-level nonIID data and visibility-dependent participation or adapt personalisation to inter-orbit similarity.

Fig. 1. FedOrbit architecture. Satellites train locally, aggregate within each orbit over ISLs, and downlink the orbit-level model to the GS during visibility windows.

B. Hierarchical Federated Learning Architecture

The architecture has satellite, orbit, and GS tiers (Fig. 1) and is shared by all methods. The following FedAvg-based procedure is used by the baselines. III. S YSTEM M ODEL FedOrbit modifies both aggregation tiers and trains nonvisible orbits over inter-satellite links, as described in A. LEO Constellation and Visibility Model Section V. We consider a Walker Delta constellation with Local training. Satellite k holds a dataset parameters following prior FL-over-LEO work [8]: Dk of size nk with local objective Fk (w) = L = 5 orbits, Nl = 4 satellites per orbit (20 satellites (1/nk ) P x∈Dk f (x; w), and runs E epochs of minitotal), altitude h = 550 km, inclination i = 53◦ , GS batch SGD with learning rate ηt = η0 γ t , producing at latitude ϕg = 51◦ N, minimum elevation αe = 10◦ ,
wkt = LocalSGD wt , Dk , E, ηt . (2) and round duration ∆t = 5 min. The GS acts as the parameter server. Intra-orbit aggregation. Satellites in orbit l combine Each satellite follows p a circular Keplerian orbit with their updates over short-range ISLs by sample-weighted mean motion no = µ/(RE + h)3 , where µ = 3.986× averaging: 105 km3 /s2 is Earth’s gravitational parameter and RE = X n P k 6371 km is Earth’s mean radius. The right ascension of wlt = wkt , (3) n j j∈Sl the ascending node (RAAN) of orbit l is Ωl = 2πl/L for k∈Sl l = 0, . . . , L − 1, giving equally spaced orbits. The GS where S is the set of satellites in orbit l. l position in the Earth-centred inertial (ECI) frame rotates Inter-orbit aggregation at the GS. Baselines apply with Earth at angular velocity ωE = 7.292 × 10−5 rad/s. FedAvg [5] across the visible orbits: Satellite k in orbit l is visible from the GS when X n   P l wt+1 = wlt , (4) (rk (t) − rg (t)) · r̂g (t) n j j∈V(t) l∈V(t) εk (t) = arcsin ≥ αe , (1) ∥rk (t) − rg (t)∥ P with nl = k∈Sl nk . FedProx, APFL, and Ditto use the where rk (t) and rg (t) are the satellite and GS positions same weighting and differ from FedAvg only in their in the ECI frame and r̂g (t) is the local zenith unit vector local objectives. at the GS. An orbit is treated as visible at round t if at least one of its satellites satisfies (1), since the orbit-level C. Non-IID Data Distribution Model Each orbit collects imagery along its ground track, model is what is uploaded to the GS. A binary matrix T ×L producing a class distribution Pl that reflects the terrain V ∈ {0, 1} records visibility, with Vt,l = 1 if orbit it overflies. Because satellites in the same orbit share l is visible at round t and 0 otherwise; the set of visible the same ground track, heterogeneity is orbit-specific orbits at round t is V(t) = {l : Vt,l = 1}. All evaluated rather than satellite-specific. Two partitioning methods methods use the same V. For T = 400, approximately 47% of rounds have are considered to represent contrasting levels of interno visible orbit, the per-orbit visibility shares are orbit overlap: Pathological partitioning. Each orbit is assigned a 15.2%, 10.2%, 8.5%, 5.2%, and 14.0%, and the longest max disjoint subset of classes, modelling orbits that observe contiguous interval with no visible orbit is τ = 191 rounds.

disjoint land-cover categories. Within each orbit, satellite where τl (t) is the number of consecutive non-visible allocations follow a Dirichlet(α = 0.5) split. rounds prior to t, se ∈ (0, 1) scales the catch-up, and Dirichlet partitioning (α = 0.5). Inter-orbit class Emax caps the load. Each of the Rintra intra-orbit rounds proportions are drawn from Dirichlet(0.5), giving over- uses El (t) local epochs and is followed by an intra-orbit lapping but skewed orbit distributions. Intra-orbit alloca- aggregation step (Section V-B). tion also follows Dirichlet(0.5). This models the partial Non-visible orbits (ISL relay). For l ∈ / V(t), each overlap that arises when orbits cover geographically satellite in the orbit runs Edark epochs of local SGD proximate but distinct regions. on the orbit-level model state, after which an intra-orbit aggregation produces an updated orbit model. This step IV. P ROBLEM F ORMULATION uses only intra-orbit ISLs and does not consume the GS In the Walker Delta constellation of Section III, the uplink. E dark is kept small (two epochs) so non-visible class distribution Pl is linked to the ground track of orbit refinements do not overwrite the heavier visible updates. l, while the visibility schedule {Vt,l }Tt=1 is determined by orbital geometry through (1). The standard FL B. Class-Aware Hierarchical Aggregation assumption that client participation is independent of When a class is concentrated on a rarely visible orbit, local data [5], [10], [11] therefore does not hold, and the global model cannot serve as an unbiased reference ordinary averaging dilutes the corresponding classifier row with updates from orbits that do not hold the for personalisation. class. FedOrbit weights aggregation by per-class data e l }L Let {w l=1 denote the orbit-level inference models, Fl the expected loss on Pl , and Accl the corresponding ownership at both the intra-orbit and inter-orbit tiers. accuracy. FedOrbit is guided by the following fairnessClass strength and class affinity. For satellite k ∈ Sl constrained objective: and class c ∈ {1, . . . , C}, min L1

e l} {w

L X l=1

e l) Fl (w

s.t.

max Accl − min Accl ≤ δ. l

sk,c = {x ∈ Dk : y(x) = c} ,

(7)

l

(5) Equation (5) expresses the design objective; the fairness term is not explicitly enforced during optimisation. The empirical spread is reported in Section VII. The formulation imposes four design requirements: (R1) reduce contribution imbalance caused by V; (R2) weight classifier updates by class ownership; (R3) adapt personalisation to inter-orbit heterogeneity without manual retuning; and (R4) keep GS-uplink traffic comparable to FedAvg, with additional work performed on-board or within an orbit. V. P ROPOSED M ETHOD : F ED O RBIT FedOrbit modifies on-board scheduling, intra-orbit and inter-orbit aggregation, and per-orbit inference. The model is split into a feature extractor and a classifier, w = [wf , wc ], since the feature extractor encodes content shared across the constellation while each classifier row is tied to a specific class.

and the class affinity of orbit l for class c is P sk,c l al,c = PL k∈S . P j=1 k∈Sj sk,c

(8)

Both quantities are computed once at initialisation and remain fixed throughout training. Intra-orbit aggregation. The classifier row for class c is averaged across the orbit’s satellites by class strength, and the feature extractor by the magnitudes of the persatellite updates ∆wkf = wkf − wf : X s P k,c wlc [c, :] = wkc [c, :], (9) s j∈Sl j,c k∈Sl

wlf =

X

∆wkf wkf , f j∈Sl ∆wj

P k∈Sl

(10)

where | · | is the elementwise absolute value, and the ratio in (10) is elementwise. Inter-orbit aggregation of the classifier. At the GS, the classifier row for class c is averaged across all L orbits with weights given by class affinity times a staleness factor: PL al,c ρσl (t) wlc [c, :] wc [c, :] ← l=1 , (11) PL σl (t) l=1 al,c ρ

A. Continuous Orbit-Level Training with ISL Relay Under standard FL, the least-visited orbit has almost no influence on the trained model because it participates in only a small fraction of rounds. FedOrbit trains every orbit in every round, using two procedures depending on visibility. Visible orbits. For l ∈ V(t), the number of local epochs grows with the length of the preceding nonwith ρ ∈ (0, 1] and σl (t) the number of rounds since visible interval: orbit l last updated its classifier; σl (t) is small because
El (t) = min Emax , Ebase + se τl (t) , (6) every orbit trains in every round.

C. Quality-Weighted Feature Aggregation and Return- Algorithm 1 FedOrbit: one communication round. Require: V; {sk,c }, {al,c } from (7), (8); β from (15); Rate Dampening τ max Continuous training produces orbit-level updates with f,p 1: State: wc (t), w̄f (t), {wl (t)}, {τl , σl }; compute different training budgets: a visible orbit returning from V(t) a long non-visible interval performs up to Emax · Rintra 2: for all l ∈ V(t) ▷ visible do epochs, while a non-visible orbit performs only Edark . 3: Set E (t) via (6); η (t) via (13); init from (17) l l FedOrbit weights inter-orbit feature aggregation by the 4: for r = 1, . . . , R do intra amount of training each orbit performed: Each k ∈ Sl runs El (t) epochs at ηl (t); 5: PL f aggregate via (9), (10) f l=1 ql (t) wl w̄ (t) = PL , (12a) 6: end for q (t) ( l=1 l 7: ql = El (t)Rintra ; reset τl , σl El (t) Rintra , l ∈ V(t), 8: end for (12b) ql (t) = Edark , l∈ / V(t). 9: for all l ∈ / V(t) ▷ non-visible do 10: Init from (17); each k runs Edark epochs at ηt ; Under the configuration used in our experiments aggregate via (9),(10) (Ebase = 5, Emax = 10, Rintra = 2, Edark = 2), 11: ql = Edark ; τl , σl += 1 a visible orbit contributes a weight of 10 to 20 and a 12: end for non-visible orbit a weight of 2. f,p 13: GS: w̄f (t) ← (12); wc (t+1) ← (11); wl (t+1) ← A returning orbit can train at the global learning (16) rate for up to Emax epochs in each intra-orbit round, e l (t + 1)} via (17) 14: return {w producing a feature shift that other orbits’ classifier rows have not adapted to. To attenuate this, FedOrbit dampens the local rate of any orbit re-entering visibility Each orbit l maintains a personal feature extractor in proportion to the length of the non-visible interval wlf,p , initialised at wlf,p (0) = wf (0) and updated by an that just ended: exponential blend toward the global quality-weighted ηt average: , (13) ηl (t) = 1 + κ τl (t)/τ max wlf,p (t + 1) = (1 − β) wlf,p (t) + β w̄f (t). (16) where τ max = 191 rounds is the longest non-visible interval and κ ≥ 0 is a hyperparameter. For κ = 0.5, The same β is used for all orbits and rounds. The an orbit returning after the longest interval uses ηl (t) = orbit’s inference model is the pair of its personal feature ηt /1.5, two thirds of the global value; an orbit visible extractor and the global classifier:   in the previous round (τl (t) = 0) is not dampened. The e l (t) = wlf,p (t), wc (t) . w (17) rule is applied only on the first round after re-entry. A larger β moves the personal extractor toward the D. Adaptive Feature Decomposition global one faster; a smaller β preserves orbit-specific The aggregation rules above produce a single global information. Both partitions are handled by the same [w̄f (t), wc (t)]. A single global model is appropriate procedure with the same hyperparameters. when class distributions overlap and suboptimal when they are disjoint. FedOrbit adapts between these two E. Communication and Computation Cost cases using a single coefficient computed from the data. Each visible orbit uploads one orbit-level model For each orbit l, let pl ∈ ∆C−1 be the normalised to the GS, as in FedAvg. FedOrbit adds intra-orbit class aggregated over its satellites, pl,c = ISL exchanges and local training during non-visible P histogram P ′ k sk,c / c′ ,k sk,c . The Jensen-Shannon distance be- periods. Both operations are performed in parallel tween orbits i and j is across orbits. For a model with P parameters, an q FP32 model requires approximately 4P bytes; training JS(pi , pj ) = 12 DKL (pi ∥m) + 21 DKL (pj ∥m), (14) also requires memory for gradients, momentum, and activations. The sub-million-parameter LeNet models with m = 12 (pi + pj ) and DKL using log2 , so JS ∈ used here therefore require tens of megabytes of working [0, 1]. The inter-orbit similarity coefficient is memory. A practical deployment would require an onX
board multicore processor supported by an FPGA, GPU, 2 β= 1 − JS(pi , pj ) . (15) or AI accelerator. L(L − 1) i<j At the GS, aggregation and personal-model storage The coefficient is computed once and fixed. grow linearly with L, while the one-time similarity

TABLE I T RAINING H YPERPARAMETERS Optimiser Weight decay Grad. clip Batch η0 γ

SGD, mom. 0.9 5×10−4 1.0 64 0.01 0.998

Ebase Emax Edark se Rintra ρ, κ

TABLE II P ERSONALISED ACCURACY ( TOP ) AND PER - ORBIT SPREAD ( BOTTOM ) 5 10 2 0.7 2 0.95, 0.5

calculation in (15) has complexity O(L2 C). The visibility matrix V supports time-varying connectivity, but larger and dynamically reconfigured constellations have not been evaluated. These settings and hardware-based validation are left for future work.

Method Accuracy FedAvg FedProx APFL Ditto FedOrbit Spread FedAvg FedProx APFL Ditto FedOrbit

EuroSAT So2Sat RESISC45 Dir. Path. Dir. Path. Dir. Path. 68.7 68.5 60.9 58.5 76.4

20.0 20.3 74.7 95.8 94.9

38.9 39.1 39.9 35.7 50.9

16.9 16.9 40.7 84.0 86.9

37.3 38.1 32.3 36.3 54.2

14.8 15.4 54.4 77.7 86.3

28.4 31.0 40.2 49.0 15.8

78.1 78.2 53.9 9.7 12.5

26.1 27.9 31.6 32.6 24.7

68.2 68.4 70.2 17.6 15.8

10.0 9.9 12.1 11.6 8.7

53.3 55.4 38.4 17.6 8.9

VI. E XPERIMENTAL S ETUP P Constellation and visibility. All experiments use is the per-orbit mean L1 l Accl , averaged over the last the Walker Delta configuration of Section III-A. The ten evaluated rounds; the secondary metric is the pervisibility matrix V is precomputed via (1) and used orbit spread maxl Accl − minl Accl . All methods were by every method. Each simulation uses T = 200 implemented in PyTorch using the same partitions and communication rounds for EuroSAT and So2Sat LCZ42 visibility matrices in each repetition. The hyperparameters in Table I are fixed across and T = 400 for RESISC45. the three datasets and both partitions. Adaptation to Datasets. EuroSAT [25]: Sentinel-2 land use, heterogeneity is provided by β in (15), not by dataset10 classes, 64×64 RGB, 27,000 samples. So2Sat specific tuning. A broader sensitivity analysis of κ, ρ, LCZ42 [26]: Sentinel-2 Local Climate Zone, 10 and s is left for future work. urban classes, 32×32, 10 bands, approximately e 352,000 training and 24,000 test samples. NWPUVII. R ESULTS AND A NALYSIS RESISC45 [27]: aerial scenes, 45 classes, 32×32 RGB, 700 images per class, 80/20 split. Table II reports the per-orbit personalised accuracy and Non-IID partitioning. Both partitions in Section III-C the per-orbit accuracy spread on all three datasets under are evaluated. The test samples are distributed among both partitions. FedOrbit achieves the highest accuracy orbits using the corresponding partitioning rule. in five of six settings and is within 0.9 percentage points Model and training. The on-board model is LeNet of the best result in the sixth. (two 5×5 convolutional layers of width 6 and 16 with Pathological partition. With disjoint inter-orbit classes, 2×2 max-pooling, followed by fully connected layers FedAvg and FedProx achieve only 14.8% to 20.0% fc1, fc2, fc3 of widths 120, 84, C). Input channels because the global classifier serves only the classes are 3 for EuroSAT and RESISC45 and 10 for So2Sat. on frequently visible orbits. APFL improves through The feature/classifier split is between fc1 and fc2, and interpolation but is regularised toward the same biased the class-affinity weighting (11) is applied to the rows anchor. Ditto is the strongest baseline; FedOrbit gains of fc3. Training hyperparameters (Table I) are identical 8.6 percentage points over Ditto on RESISC45 and 2.9 across methods within each dataset. on So2Sat, while trailing Ditto by 0.9 percentage points Baselines. FedAvg [5] averages visible-orbit mod- on EuroSAT. els (4); FedProx [10] adds a proximal term (µ = Dirichlet partition. A different pattern is observed. 0.01); APFL [14] interpolates the FedAvg global and Ditto is 37 to 48 percentage points lower than under a local orbit model with an adaptively learned mixing pathological partitioning, and APFL also degrades coefficient; Ditto [13] trains a per-orbit personal model because strong personalisation is less beneficial when with an ℓ2 proximal term (λDitto = 0.1) toward a orbit distributions overlap. FedOrbit gains 7.7, 11.0, FedAvg global trained in parallel. The optimiser, batch and 16.1 percentage points over the strongest baseline size, initial learning rate, and scheduler are shared across on EuroSAT, So2Sat, and RESISC45. The adaptive methods. feature decomposition provides a consistent explanation: Evaluation. For each orbit l, Accl is the accuracy β (15) is moderate under Dirichlet partitioning (≈ 0.44) e l on orbit l’s test partition (the global model for and small under pathological partitioning (≈ 0.11). of w FedAvg and FedProx, the personalised model for APFL The personal feature extractor therefore tracks the and Ditto, and (17) for FedOrbit). The primary metric global average more quickly under Dirichlet partitioning

and retains orbit-specific information for longer under pathological partitioning. Per-orbit fairness. The spread block of Table II operationalises the fairness term in (5). FedOrbit has the smallest spread in five of six settings and is within 3 percentage points of the smallest in the sixth (EuroSAT pathological). Under pathological partitioning, the global baselines leave spreads of 53 to 78 percentage points, confirming poor accuracy for low-visibility orbits. Component interpretation. As the mechanisms operate jointly, Table II does not isolate their individual effects. Fixed hyperparameters avoid dataset-specific tuning, while the change in β is consistent with stronger sharing under Dirichlet partitioning and greater orbit-specific retention under pathological partitioning. A controlled ablation is left for future work. VIII. C ONCLUSION FedOrbit addresses the coupling between orbit-level data heterogeneity and visibility-dependent participation in LEO constellations. It combines continuous orbit-level training, class-aware hierarchical aggregation, qualityweighted feature aggregation with return-rate dampening, and adaptive feature decomposition. Across three remotesensing datasets, FedOrbit achieves the highest accuracy in five of six settings and is within 0.9 percentage points of the best result in the sixth. It also produces the smallest per-orbit accuracy spread in five of six settings. Future work will evaluate individual component contributions, hardware deployment, and larger dynamic constellations. R EFERENCES [1] O. Kodheli et al., “Satellite communications in the new space era: A survey and future challenges,” IEEE Commun. Surveys Tuts., vol. 23, no. 1, pp. 70–109, 2021. 1st Quart. [2] I. del Portillo, B. G. Cameron, and E. F. Crawley, “A technical comparison of three low earth orbit satellite constellation systems to provide global broadband,” Acta Astronautica, vol. 159, pp. 123–135, 2019. [3] G. Giuffrida et al., “CloudScout: A deep neural network for onboard cloud detection on hyperspectral images,” Remote Sens., vol. 12, no. 14, p. 2205, 2020. [4] G. Mateo-Garcia et al., “Towards global flood mapping onboard low cost satellites with machine learning,” Sci. Rep., vol. 11, p. 7249, 2021. [5] H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. Agüera y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Proc. 20th Int. Conf. Artif. Intell. Statist. (AISTATS), vol. 54, pp. 1273–1282, 2017. [6] B. Matthiesen, N. Razmi, I. Leyva-Mayorga, A. Dekorsy, and P. Popovski, “Federated learning in satellite constellations,” IEEE Netw., vol. 37, no. 5, pp. 32–39, 2023. [7] I. Leyva-Mayorga et al., “LEO small-satellite constellations for 5G and beyond-5G communications,” IEEE Access, vol. 8, pp. 184955–184964, 2020. [8] N. Razmi, B. Matthiesen, A. Dekorsy, and P. Popovski, “Groundassisted federated learning in LEO satellite constellations,” IEEE Wireless Commun. Lett., vol. 11, pp. 717–721, Apr. 2022. [9] Q. Li, Y. Diao, Q. Chen, and B. He, “Federated learning on non-IID data silos: An experimental study,” in Proc. 38th IEEE Int. Conf. Data Eng. (ICDE), pp. 965–978, 2022.

[10] T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V. Smith, “Federated optimization in heterogeneous networks,” in Proc. Mach. Learn. Syst. (MLSys), pp. 429–450, 2020. [11] S. P. Karimireddy, S. Kale, M. Mohri, S. J. Reddi, S. U. Stich, and A. T. Suresh, “SCAFFOLD: Stochastic controlled averaging for federated learning,” in Proc. 37th Int. Conf. Mach. Learn. (ICML), pp. 5132–5143, 2020. [12] J. Wang, Q. Liu, H. Liang, G. Joshi, and H. V. Poor, “Tackling the objective inconsistency problem in heterogeneous federated optimization,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, pp. 7611–7623, 2020. [13] T. Li, S. Hu, A. Beirami, and V. Smith, “Ditto: Fair and robust federated learning through personalization,” in Proc. Int. Conf. Mach. Learn. (ICML), pp. 6357–6368, 2021. [14] Y. Deng, M. M. Kamani, and M. Mahdavi, “Adaptive personalized federated learning,” arXiv preprint arXiv:2003.13461, 2020. [15] C. T. Dinh, N. H. Tran, and T. D. Nguyen, “Personalized federated learning with moreau envelopes,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, pp. 21394–21405, 2020. [16] A. Fallah, A. Mokhtari, and A. Ozdaglar, “Personalized federated learning with theoretical guarantees: A model-agnostic metalearning approach,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, pp. 3557–3568, 2020. [17] L. Collins, H. Hassani, A. Mokhtari, and S. Shakkottai, “Exploiting shared representations for personalized federated learning,” in Proc. 38th Int. Conf. Mach. Learn. (ICML), pp. 2089– 2099, 2021. [18] J. So, K. Hsieh, B. Arzani, S. Noghabi, S. Avestimehr, and R. Chandra, “FedSpace: An efficient federated learning framework at satellites and ground stations,” arXiv preprint arXiv:2202.01267, 2022. [19] M. Elmahallawy and T. Luo, “FedHAP: Fast federated learning for LEO constellations using collaborative HAPs,” in Proc. IEEE Int. Conf. Wireless Commun. Signal Process. (WCSP), pp. 1–6, 2022. [20] M. Elmahallawy and T. Luo, “AsyncFLEO: Asynchronous federated learning for LEO satellite constellations with highaltitude platforms,” in Proc. IEEE Int. Conf. Big Data, pp. 5478– 5487, Dec. 2022. [21] Z. Zhai, Q. Wu, S. Yu, R. Li, F. Zhang, and X. Chen, “FedLEO: An offloading-assisted decentralized federated learning framework for low earth orbit satellite networks,” IEEE Trans. Mobile Comput., vol. 23, pp. 5260–5279, May 2024. [22] Y. Huang, X. Li, M. Zhao, H. Li, and M. Peng, “Asynchronous federated learning via over-the-air computation in LEO satellite networks,” IEEE Trans. Wireless Commun., vol. 23, pp. 19885– 19897, Dec. 2024. [23] Z. Lin, Z. Chen, Z. Fang, X. Chen, X. Wang, and Y. Gao, “FedSN: A federated learning framework over heterogeneous LEO satellite networks,” IEEE Trans. Mobile Comput., vol. 24, pp. 1293–1307, Mar. 2025. [24] L. Zhao et al., “ALANINE: A novel decentralized personalized federated learning for heterogeneous LEO satellite constellation,” arXiv preprint arXiv:2411.07752, 2024. [25] P. Helber, B. Bischke, A. Dengel, and D. Borth, “EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 12, no. 7, pp. 2217–2226, 2019. [26] X. X. Zhu, J. Hu, C. Qiu, Y. Shi, J. Kang, L. Mou, H. Bagheri, M. Häberle, Y. Hua, R. Huang, L. Hughes, H. Li, Y. Sun, G. Zhang, S. Han, M. Schmitt, and Y. Wang, “So2Sat LCZ42: A benchmark data set for the classification of global local climate zones,” IEEE Geosci. Remote Sens. Mag., vol. 8, no. 3, pp. 76–89, 2020. [27] G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classification: Benchmark and state of the art,” Proc. IEEE, vol. 105, pp. 1865–1883, Oct. 2017.

Related documents

• Adaptive Beam Hopping and Power Control for Dual-Layer Over-the-Air Online Federated Learning in LEO Satellite Networks#925563
• Communication-Efficient Agnostic Federated Learning via Faster Convergence and Compression#1131811
• CoeF-SFL: Preserving Collaborative Server-Client Learning with Enhanced Communication Efficiency#1119433
• Metaplasticity as adaptive gradient preconditioning for incremental learning#609072
• Spatial and Temporal Correlation of Interference in a Narrow Multibeam LEO Satellite Random Access Network#1110295