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

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

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.

Introduction

Low Earth Orbit (LEO) satellite constellations are now an important platform for global Earth observation, and on-board processing makes it possible to train and run machine learning models directly on the satellite. Transmitting raw observation data to a ground station for centralised training is impractical given the intermittent nature of satellite-to-ground links, data privacy requirements, and bandwidth constraints. Federated learning (FL) reduces the need for raw-data transfer: each satellite trains locally on its own observations and exchanges only model parameters with a ground station (GS).

FL in LEO constellations differs fundamentally from the terrestrial setting. In a Walker Delta constellation, each orbit follows a distinct ground track and observes a different geographic region, producing orbit-level class distributions that are inherently non-IID. Orbit-to-GS visibility is governed by the same orbital 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. 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 model in every round; (iii) class-aware hierarchical aggregation, routing classifier updates by per-class data ownership; (iv) quality-weighted feature aggregation with return-rate dampening, limiting the influence of an orbit returning from a long absence; (v) adaptive feature decomposition, blending each orbit's personal feature extractor toward the global one at a rate set by inter-orbit class similarity. We evaluate FedOrbit on three remote-sensing classification benchmarks under both pathological and Dirichlet (α = 0.5) partitions, with the largest gains on the partition where existing methods perform worst.

Related Work

FL under non-IID data: FedAvg degrades under heterogeneous distributions; FedProx, SCAFFOLD, and FedNova use proximal regularisation, control variates, or local-step normalisation. These methods produce a single global model but do not model the coupling between local data and visibility-dependent participation.

Personalised FL: Ditto adds an l2 penalty toward the global model; APFL interpolates local and global models; other methods use regularised inner loops, meta-learned initialisations, or split representations. These methods do not account for bias in the global reference model caused by visibility-dependent participation.

FL in non-terrestrial networks: Prior LEO FL work has addressed scheduling under irregular visibility, semi-asynchronous aggregation, aerial parameter servers, decentralised intra-orbit aggregation, over-the-air computation, system heterogeneity, and decentralised personalisation. The common focus is communication architecture, convergence speed, and scheduling. To our knowledge, prior work does not jointly address orbit-level non-IID data and visibility-dependent participation or adapt personalisation to inter-orbit similarity.

System Model

LEO Constellation and Visibility Model: A Walker Delta constellation with L = 5 orbits, Nl = 4 satellites per orbit (20 satellites total), altitude h = 550 km, inclination i = 53°, GS at latitude ϕg = 51°N, minimum elevation αe = 10°, and round duration ∆t = 5 min. The GS acts as the parameter server. Each satellite follows a circular Keplerian orbit with mean motion no = √(µ/(RE + h)3), where µ = 3.986×105 km3/s2 is Earth's gravitational parameter and RE = 6371 km is Earth's mean radius. The right ascension of the ascending node (RAAN) of orbit l is omegal = 2πl/L for l = 0,..., L−1, giving equally spaced orbits. The GS position in the Earth-centred inertial (ECI) frame rotates with Earth at angular velocity ωE = 7.292×10−5 rad/s. Satellite k in orbit l is visible from the GS when the elevation angle εk(t) ≥ αe. An orbit is treated as visible at round t if at least one of its satellites satisfies the visibility condition. A binary matrix V ∈ 0,1(T×L) records visibility, with Vt,l = 1 if orbit l is visible at round t and 0 otherwise; the set of visible orbits at round t is V(t) = l: Vt,l = 1. All evaluated methods use the same V. For T = 400, approximately 47% of rounds have no visible orbit, the per-orbit visibility shares are 15.2%, 10.2%, 8.5%, 5.2%, and 14.0%, and the longest contiguous interval with no visible orbit is τ max = 191 rounds.

Hierarchical Federated Learning Architecture: The architecture has satellite, orbit, and GS tiers. Local training: Satellite k holds a dataset Dk of size nk with local objective Fk(w) = (1/nk) Σ x∈Dk f(x; w), and runs E epochs of minibatch SGD with learning rate ηt = η0 γ t, producing wkt = LocalSGD(wt, Dk, E, ηt). Intra-orbit aggregation: Satellites in orbit l combine their updates over short-range ISLs by sample-weighted averaging: wlt = Σ k∈Sl (nk / Σ j∈Sl nj) wkt, where Sl is the set of satellites in orbit l. Inter-orbit aggregation at the GS: Baselines apply FedAvg across the visible orbits: wt+1 = Σ l∈V(t) (nl / Σ j∈V(t) nj) wlt, with nl = Σ k∈Sl nk. FedProx, APFL, and Ditto use the same weighting and differ from FedAvg only in their local objectives.

Non-IID Data Distribution Model: Each orbit collects imagery along its ground track, producing a class distribution Pl that reflects the terrain it overflies. Because satellites in the same orbit share the same ground track, heterogeneity is orbit-specific rather than satellite-specific. Two partitioning methods are considered:

• Pathological partitioning: Each orbit is assigned a disjoint subset of classes, modelling orbits that observe disjoint land-cover categories. Within each orbit, satellite allocations follow a Dirichlet(α = 0.5) split.

• Dirichlet partitioning (α = 0.5): Inter-orbit class proportions are drawn from Dirichlet(0.5), giving overlapping but skewed orbit distributions. Intra-orbit allocation also follows Dirichlet(0.5). This models the partial overlap that arises when orbits cover geographically proximate but distinct regions.

Problem Formulation

In the Walker Delta constellation, the class distribution Pl is linked to the ground track of orbit l, while the visibility schedule Vt,l is determined by orbital geometry. The standard FL assumption that client participation is independent of local data therefore does not hold, and the global model cannot serve as an unbiased reference for personalisation.

Let w̃l denote the orbit-level inference models, Fl the expected loss on Pl, and Accl the corresponding accuracy. FedOrbit is guided by the following fairness-constrained objective:

min w̃l Σ l=1 L Fl(w̃l) s.t. max l Accl − min l Accl ≤ δ.

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.

Proposed Method: FedOrbit

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.

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 non-visible interval: El(t) = min(Emax, Ebase + se τl(t)), where τl(t) is the number of consecutive non-visible rounds prior to t, se ∈ (0, 1) scales the catch-up, and Emax caps the load. Each of the Rintra intra-orbit rounds uses El(t) local epochs and is followed by an intra-orbit aggregation step.

• Non-visible orbits (ISL relay): For l ∉ V(t), each satellite in the orbit runs Edark epochs of local SGD on the orbit-level model state, after which an intra-orbit aggregation produces an updated orbit model. This step uses only intra-orbit ISLs and does not consume the GS uplink. Edark is kept small (two epochs) so non-visible orbit refinements do not overwrite the heavier visible updates.

Class-Aware Hierarchical Aggregation: When a class is concentrated on a rarely visible orbit, ordinary averaging dilutes the corresponding classifier row with updates from orbits that do not hold the class. FedOrbit weights aggregation by per-class data ownership at both the intra-orbit and inter-orbit tiers.

• Class strength and class affinity: For satellite k ∈ Sl and class c, sk,c = x ∈ Dk: y(x) = c, and the class affinity of orbit l for class c is al,c = Σ k∈Sl sk,c / Σ j=1 L Σ k∈Sj sk,c. 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 per-satellite updates ∆wkf = wkf − wf: wlc[c,:] = Σ k∈Sl (sk,c / Σ j∈Sl sj,c) wkc[c,:], and wlf = Σ k∈Sl (∆wkf / Σ j∈Sl ∆wjf) wkf, where · is the elementwise absolute value, and the ratio 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: wc[c,:] ← Σ l=1 L (al,c ρ σl(t) wlc[c,:]) / Σ l=1 L al,c ρ σl(t), with ρ ∈ (0, 1] and σl(t) the number of rounds since orbit l last updated its classifier; σl(t) is small because every orbit trains in every round.

Quality-Weighted Feature Aggregation and Return-Rate Dampening: Continuous training produces orbit-level updates with different training budgets: a visible orbit returning from a long non-visible interval performs up to Emax · Rintra epochs, while a non-visible orbit performs only Edark. FedOrbit weights inter-orbit feature aggregation by the amount of training each orbit performed: w̄f(t) = Σ l=1 L ql(t) wlf / Σ l=1 L ql(t), where ql(t) = El(t) Rintra for l ∈ V(t), and ql(t) = Edark for l ∉ V(t). Under the configuration used in the experiments (Ebase = 5, Emax = 10, Rintra = 2, Edark = 2), a visible orbit contributes a weight of 10 to 20 and a non-visible orbit a weight of 2.

A returning orbit can train at the global learning rate for up to Emax epochs in each intra-orbit round, 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 in proportion to the length of the non-visible interval that just ended: ηl(t) = ηt / (1 + κ τl(t)/τ max), where τ max = 191 rounds is the longest non-visible interval and κ ≥ 0 is a hyperparameter. For κ = 0.5, an orbit returning after the longest interval uses ηl(t) = ηt/1.5, two thirds of the global value; an orbit visible in the previous round (τl(t) = 0) is not dampened. The rule is applied only on the first round after re-entry.

Adaptive Feature Decomposition: The aggregation rules above produce a single global [w̄f(t), wc(t)]. A single global model is appropriate when class distributions overlap and suboptimal when they are disjoint. FedOrbit adapts between these two cases using a single coefficient computed from the data. For each orbit l, let pl ∈ Δ(C−1) be the normalised class histogram aggregated over its satellites. The Jensen-Shannon distance between orbits i and j is JS(pi, pj) = √(½ DKL(pi∥m) + ½ DKL(pj∥m)), with m = ½(pi + pj) and DKL using log2, so JS ∈ [0, 1]. The inter-orbit similarity coefficient is β = (2/(L(L−1))) Σ i<j (1 − JS(pi, pj))2. The coefficient is computed once and fixed.

Each orbit l maintains a personal feature extractor wlf,p, initialised at wlf,p(0) = wf(0) and updated by an exponential blend toward the global quality-weighted average: wlf,p(t + 1) = (1 − β) wlf,p(t) + β w̄f(t). The same β is used for all orbits and rounds. The orbit's inference model is the pair of its personal feature extractor and the global classifier: w̃l(t) = (wlf,p(t), wc(t)). A larger β moves the personal extractor toward the global one faster; a smaller β preserves orbit-specific information. Both partitions are handled by the same procedure with the same hyperparameters.

Communication and Computation Cost: Each visible orbit uploads one orbit-level model to the GS, as in FedAvg. FedOrbit adds intra-orbit ISL exchanges and local training during non-visible periods. Both operations are performed in parallel across orbits. For a model with P parameters, an FP32 model requires approximately 4P bytes; training also requires memory for gradients, momentum, and activations. The sub-million-parameter LeNet models used here require tens of megabytes of working memory. A practical deployment would require an on-board multicore processor supported by an FPGA, GPU, or AI accelerator. At the GS, aggregation and personal-model storage grow linearly with L, while the one-time similarity calculation in (15) has complexity O(L2C). The visibility matrix V supports time-varying connectivity, but larger and dynamically reconfigured constellations have not been evaluated.

Experimental Setup

Constellation and visibility: All experiments use the Walker Delta configuration described above. The visibility matrix V is precomputed via (1) and used by every method. Each simulation uses T = 200 communication rounds for EuroSAT and So2Sat LCZ42 and T = 400 for RESISC45.

Datasets: EuroSAT (Sentinel-2 land use, 10 classes, 64×64 RGB, 27,000 samples); So2Sat LCZ42 (Sentinel-2 Local Climate Zone, 10 urban classes, 32×32, 10 bands, approximately 352,000 training and 24,000 test samples); NWPU-RESISC45 (aerial scenes, 45 classes, 32×32 RGB, 700 images per class, 80/20 split).

Non-IID partitioning: Both partitions are evaluated. The test samples are distributed among orbits using the corresponding partitioning rule.

Model and training: The on-board model is LeNet (two 5×5 convolutional layers of width 6 and 16 with 2×2 max-pooling, followed by fully connected layers fc1, fc2, fc3 of widths 120, 84, C). Input channels are 3 for EuroSAT and RESISC45 and 10 for So2Sat. The feature/classifier split …

Improvements to AI Systems Based on FedOrbit:

1. Adaptive Personalization via Inter-Client Similarity: Implement a dynamic blending coefficient (β) that automatically adjusts the degree of personalization based on measured data distribution overlap between clients. This prevents over-personalization when distributions overlap and under-personalization when they are disjoint, eliminating manual tuning for heterogeneous environments.

2. Visibility-Aware Continuous Training: Enable all clients to train in every round, even when disconnected from the central server, using local relay mechanisms (e.g., inter-satellite links). This reduces contribution imbalance and ensures low-visibility clients still influence the global model, improving fairness and accuracy for underrepresented data sources.

3. Class-Aware Hierarchical Aggregation: Replace sample-count-based averaging with per-class data ownership weighting at both local and global aggregation tiers. This prevents classifier dilution when classes are concentrated on rarely participating clients, improving accuracy on rare or niche classes.

4. Quality-Weighted Model Aggregation with Staleness Dampening: Weight client updates by the actual training effort performed (e.g., epochs) and dampen the learning rate of clients returning from long absences. This prevents stale or over-trained updates from destabilizing the global model, improving convergence stability in asynchronous or intermittently connected settings.

5. Fairness-Constrained Optimization: Incorporate a fairness constraint (e.g., minimizing per-client accuracy spread) directly into the training objective. This produces models with more equitable performance across all clients, which is critical for applications where uniform service quality is required (e.g., global Earth observation, healthcare across institutions).

What the Improved AI System Can Do:

• Operate reliably in intermittently connected networks (e.g., satellites, IoT, mobile devices) by training continuously and aggregating fairly, even when some clients are rarely online.

• Automatically adapt to varying data heterogeneity without manual hyperparameter tuning, switching between global sharing and local specialization as needed.

• Achieve higher accuracy on rare or client-specific classes by weighting updates by class ownership, rather than treating all samples equally.

• Provide more equitable performance across all clients, reducing accuracy disparities caused by uneven participation or data imbalance.

• Handle dynamic, non-stationary environments (e.g., changing visibility schedules, shifting data distributions) with robust aggregation and staleness management.

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.