Paper deep dive
ORBITALIF: An Efficient Spiking Federated Learning Framework for Onboard Cloud Removal
Bohan Zhang, Chenyu Xu, Yijie Mao, Yuanming Shi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/26/2026, 5:31:41 AM
Summary
The paper introduces OrbitALIF, a novel framework for onboard cloud removal in Low-Earth Orbit (LEO) satellite constellations. It combines a decentralized Satellite Federated Learning (SFL) strategy with a compact 2.30 M-parameter Spiking Neural Network (SNN) backbone. Key architectural components include an Adaptive Gated Fusion Module (AGFM) and a Spectral-Spatial Hybrid Attention Module (SHAM). The framework enables both onboard training and inference, achieving competitive cloud removal quality while consuming only 0.287 mJ per inference, representing a 72.3x (98.6%) energy reduction compared to equivalent Artificial Neural Networks (ANNs).
Entities (9)
Relation Signals (8)
OrbitALIF → runson → LEO Satellites
confidence 95% · onboard cloud removal framework... for LEO satellite constellations
OrbitALIF → uses → Spiking Neural Network
confidence 95% · OrbitALIF performs both onboard training and inference using a compact 2.30 M-parameter spiking neural network (SNN) backbone
OrbitALIF → contains → Spectral-Spatial Hybrid Attention Module
confidence 92% · backbone with an adaptive gated fusion module (AGFM) and a spectral-spatial hybrid attention module (SHAM)
OrbitALIF → contains → Adaptive Gated Fusion Module
confidence 92% · backbone with an adaptive gated fusion module (AGFM)
OrbitALIF → employs → Satellite Federated Learning
confidence 90% · we propose a novel satellite federated learning framework... named orbital attention leaky integrate-and-fire (OrbitALIF)
OrbitALIF → evaluatedon → CUHK-CR
confidence 90% · We evaluate OrbitALIF on the public CUHK-CR [7] cloud removal benchmark
OrbitALIF → outperforms → Artificial Neural Network
confidence 90% · 72.3 times (98.6%) energy reduction versus an equivalent artificial neural network (ANN)
Satellite Federated Learning → uses → Inter-Satellite Links
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Low-earth-orbit (LEO) satellites enable high-resolution, large-scale Earth observation for applications such as disaster monitoring and environmental surveillance. However, cloud coverage often obscures the Earth's surface, and conventional cloud-removal pipelines that download cloudy images to ground stations for processing suffer from limited contact windows, constrained satellite-to-ground bandwidth, and high latency. In this work, we propose a novel satellite federated learning framework for cloud removal across LEO constellations, named orbital attention leaky integrate-and-fire (OrbitALIF). OrbitALIF performs both onboard training and inference using a compact 2.30,M-parameter spiking neural network (SNN) backbone with an adaptive gated fusion module (AGFM) and a spectral-spatial hybrid attention module (SHAM), combined with a decentralized federated learning strategy that shares model weights via inter-satellite links. Our experiments show that OrbitALIF achieves competitive cloud removal quality while consuming only 0.287,mJ per inference on neuromorphic hardware, a 72.3 times (98.6%) energy reduction versus an equivalent artificial neural network (ANN).
Tags
Links
- Source: https://arxiv.org/abs/2608.24073v1
- Canonical: https://arxiv.org/abs/2608.24073v1
Trouble viewing inline? Open PDF directly →
Full Text
35,368 characters extracted from source content.
Expand or collapse full text
OrbitALIF: An Efficient Spiking Federated Learning Framework for Onboard Cloud RemovalThanks: This work was supported by the National Natural Science Foundation of China under Grant 62571331, the work of Yuanming Shi was supported by the National Natural Science Foundation of China under Grant 62522117 and Grant 62271318. (Corresponding author: Yijie Mao) Bohan Zhang, Chenyu Xu, Yijie Mao, Yuanming Shi Affiliation: School of Information Science and Technology, ShanghaiTech University, Shanghai, China Email: zhangbh2023, xuchy2024, maoyj, shiym@shanghaitech.edu.cn Affiliation: Abstract Low-earth-orbit (LEO) satellites enable high-resolution, large-scale Earth observation for applications such as disaster monitoring and environmental surveillance. However, cloud coverage often obscures the Earth’s surface, and conventional cloud-removal pipelines that download cloudy images to ground stations for processing suffer from limited contact windows, constrained satellite-to-ground bandwidth, and high latency. In this work, we propose a novel satellite federated learning framework for cloud removal across LEO constellations, named orbital attention leaky integrate-and-fire (OrbitALIF). OrbitALIF performs both onboard training and inference using a compact 2.302.30 M-parameter spiking neural network (SNN) backbone with an adaptive gated fusion module (AGFM) and a spectral-spatial hybrid attention module (SHAM), combined with a decentralized federated learning strategy that shares model weights via inter-satellite links. Our experiments show that OrbitALIF achieves competitive cloud removal quality while consuming only 0.2870.287 mJ per inference on neuromorphic hardware, a 72.3×72.3× (98.6%98.6\%) energy reduction versus an equivalent artificial neural network (ANN). Index Terms: Spiking neural network, cloud removal, satellite federated learning, LEO constellation, neuromorphic computing. †aftertitle: I Introduction Low-earth-orbit (LEO) satellites have become an essential infrastructure for Earth observation, supporting time-sensitive applications such as disaster response [1], environmental monitoring [2], and large-scale remote-sensing analytics. However, heavy cloud coverage can severely degrade the quality of the captured imagery. Cloud removal, which aims to recover the original surface from a cloud-contaminated observation, is therefore a key preprocessing step for remote-sensing tasks. Early cloud removal methods employed basic convolutional networks but only for thin cloud [3]. To alleviate the ambiguity caused by cloud occlusion, subsequent methods incorporated additional observations. SAR-assisted approaches such as DSen2-CR [4] exploit co-registered cross-sensor information, while multi-temporal methods leverage repeated acquisitions of the same scene. Although effective, these approaches introduce additional requirements for auxiliary sensors, image registration, or repeated revisits. Other methods therefore remove the cloud directly from optical observation through powerful representation model such as MemoryNet [5]. Recently, diffusion-enhanced methods further improve cloud removal quality through iterative denoising [6, 7], but at the cost of substantial computational complexity and inference latency. Consequently, existing cloud removal solutions face two fundamental deployment challenges. First, performing cloud removal on the ground requires cloud-contaminated images to be transmitted through bandwidth-constrained links, resulting in high latency. Second, sophisticated cloud removal models are often too computationally and energy intensive for practical onboard training and inference. To address these challenges, satellite federated learning (SFL) has been proposed, leveraging the entire satellite constellation as a collaborative computing platform [8]. By performing local training on each satellite and sharing model parameters via inter-satellite links (ISLs) [9, 10], SFL enables access to richer and more diverse training data without transmitting raw imagery to the ground, while also preserving data security and privacy. However, prior SFL efforts have been limited to classification and detection tasks. The more challenging task of dense pixel-level cloud removal, which is a per-pixel regression problem under physically non-independent and identically distributed (non-IID) conditions, remains largely unexplored. Moreover, SFL alone does not reduce the energy cost of onboard inference, which can exceed the power budgets of LEO satellites. This calls for an inherently energy-efficient model architecture. Spiking neural networks (SNNs) offer a solution through event-driven inference [11], consuming power only when spikes are emitted, and recent work has demonstrated their feasibility for low-level vision tasks such as deraining [12]. Yet conventional binary spikes lack sufficient expressiveness for dense per-pixel regression and aggressively reducing computation degrades cloud removal quality. In this work, we propose the first onboard cloud removal framework named orbital attention leaky integrate-and-fire (OrbitALIF) that jointly supports onboard training and onboard inference for cloud removal over a LEO satellite constellation.By employing a decentralized SFL framework and an efficient SNN backbone, the proposed OrbitALIF significantly reduces per-satellite energy consumption while maintaining competitive testing accuracy. The main contributions of this paper are summarized as follows: • We propose OrbitALIF an onboard SNN-based federated cloud-removal framework for LEO satellite constellations. To the best of our knowledge, this is the first work to go beyond classification tasks and combine onboard SNN inference with cross-orbit SFL for the low-level vision task of cloud removal. • We design a lightweight SNN backbone for onboard cloud removal. Compared with existing cloud removal neural networks, the proposed backbone introduces two key architectural novelties. First, an adaptive gated fusion module (AGFM) replaces the rigid skip connections of a standard U-Net with a learnable gating mechanism. Second, a spectral-spatial hybrid attention module (SHAM) is introduced with two branches: a temporal attention branch that weights per-time-step importance, and a spectral attention branch that reweights frequency components via the 2-D fast Fourier transform (FFT) while preserving spatial phase structure. The complete model contains only 2.302.30 M parameters, making it suitable for resource-constrained satellite deployment. • We evaluate OrbitALIF on the public CUHK-CR [7] cloud removal benchmark under hierarchical inter-satellite links. OrbitALIF achieves performance comparable to an ANN U-Net model of the same architecture with a 72.3×72.3× reduction in inference energy, demonstrating the promise of spiking computation for onboard cloud removal. I System Model I-A Satellite Communication Model In this work, we consider a constellation consisting of N orbital planes, each containing K satellites. We denote the j-th satellite within the i-th orbital plane by si,js_i,j. Satellites are assumed to communicate with one another via ISLs. Two decision metrics are introduced to determine whether a pair of satellites can establish communication: 1) Line-of-sight (LoS) visibility: Let si,js_i,j and si′,j′s_i ,j be an arbitrary satellite pair (hereafter denoted simply as s1s_1 and s2s_2 when no ambiguity arises), with altitudes h1h_1 and h2h_2 above the Earth’s surface, respectively. The horizon half-angle for satellite sks_k is ρk=arccos(rE/(rE+hk)) _k= (r_E/(r_E+h_k) ), which determines the angular radius of the cone within which the satellite can establish LoS with other spacecraft. Let θkv _k^v and θkh _k^h denote the latitude and longitude of satellite sks_k, respectively. The geocentric angle φ between the two satellites is given by the spherical law of cosines: φ=arccos[sinθ1vsinθ2v+cosθ1vcosθ2vcos(θ1h−θ2h)], = [ _1^v _2^v+ _1^v _2^v ( _1^h- _2^h) ], (1) The two satellites are mutually LoS visible if and only if their geocentric angle does not exceed the sum of their individual horizon half-angles, i.e., φ≤ρ1+ρ2 ≤ _1+ _2. Fig. 1: System model of the LEO satellite constellation. 2) Doppler reliability: The Doppler shift captures the dynamic dependency of ISL reliability. A clear LoS is insufficient to ensure link usability under high relative velocity. The Doppler shift between two satellites is fd=ψfc/cf_d=ψ f_c/c, where ψ is the relative speed, fcf_c is the carrier frequency, and c is the speed of light. The two satellites are considered Doppler reliable if and only if fd≤fmaxf_d≤ f_ . Once both metrics are satisfied, the satellite pair is considered credible, enabling successful information transmission between the two satellites. I-B Satellite Federated Learning Model Assuming each satellite si,js_i,j holds a local dataset i,jD_i,j, all satellites collaboratively learn a global model by solving the following empirical risk minimization (ERM) problem: min∈ℝdf()=1NK∑i=1N∑j=1Kfi,j(), _w ^d f(w)= 1NK _i=1^N _j=1^Kf_i,j(w), (2) where fi,j()=1|i,j|∑∈i,jℓ(,),f_i,j(w)= 1|D_i,j| _z _i,j (w;z), (3) and w denotes the model parameters with dimension d, ℓ(⋅,) (·;z) is the loss on sample z. The operator |⋅||·| denotes the cardinality of the set. Inspired by [13], we adopt a constellation-aware federated learning architecture that combines intra-plane ring all-reduce with inter-plane gossip averaging, ensuring efficient and stable model aggregation under the physical constraints of the satellite network. The learning procedure at each global iteration r proceeds in the following four steps, as shown in Fig 1: Step 1 — Model Distribution. At the beginning of iteration r, each orbital plane i holds a local model copy irw_i^r. Within each plane, the K satellites are organized as a logical ring topology based on their stable relative positions along the orbit. In this ring, each satellite si,js_i,j has a unique predecessor si,j−1s_i,j-1 and a unique successor si,j+1s_i,j+1 (indices modulo K). The satellite that participated in the previous inter-plane aggregation disseminates irw_i^r to all other satellites si,js_i,j in the same orbit by forwarding it along this ring, using the reliable intra-plane ISLs. After this step, every satellite in orbital plane i holds irw_i^r as its initial model for the current round, and the ring topology used for distribution also serves as the communication structure for subsequent aggregation. Step 2 — Local Training. Each satellite si,js_i,j updates its local model with a mini-batch optimizer AdamW [14].The learning-rate schedule depends on the global round and is shared by all satellites, whereas the optimizer moments remain local. We denote the weights after local training by i,jr+12w_i,j^r+ 12. Step 3 — Intra-Plane Model Aggregation (Ring All-Reduce). Using the same ring topology established in Step 1, the K satellites perform ring all-reduce for intra-plane aggregation. Each satellite si,js_i,j sends its local update i,jr+12w_i,j^r+ 12 to its successor si,j+1s_i,j+1 and receives an update from its predecessor si,j−1s_i,j-1. After K such communication rounds, every satellite in the plane obtains the exact average of all local models within that orbit: ir+12=1K∑j=1Ki,jr+12,∀i∈1,…,N.w_i^r+ 12= 1K _j=1^Kw_i,j^r+ 12, ∀ i∈\1,…,N\. (4) At this point, each orbit i has a unified intra-plane model ir+12w_i^r+ 12. Step 4 — Inter-Plane Model Aggregation (Gossip Averaging). Unlike the stable intra-plane ring topology, inter-plane ISLs are intermittent due to varying inclinations, altitudes, and relative velocities, making it impractical to achieve global aggregation across all N orbits in a single round. Instead, a gossip averaging scheme is adopted: each orbital plane i communicates only with its immediate neighbors on a pre-defined inter-plane connectivity graph =(,ℰ)G=(V,E), where =1,…,NV=\1,…,N\ and ei,i′∈ℰe_i,i only if there exists at least one credible pair between orbits i and i′i (i.e., satisfying both LoS visibility and Doppler reliability constraints). In each gossip round, orbit i exchanges its model ir+12w_i^r+ 12 with its neighbors i=i′∣ei,i′∈ℰN_i=\i e_i,i \ and updates [13]: ir+1=(1−α)ir+12+α|i|∑i′∈i′r+12,w_i^r+1=(1-α)w_i^r+ 12+ α|N_i| _i _iw_i ^r+ 12, (5) where α∈(0,1)α∈(0,1) is the gossip mixing weight. Through repeated gossip rounds across iterations, model information gradually diffuses across the entire constellation, reaching consensus over time. This avoids the high communication overhead of global aggregation while remaining robust to the intermittent nature of inter-plane ISLs. I Proposed Orbital Attention Leaky Integrate-and-fire (OrbitALIF) Framework I-A Motivation and Architectural Overview Fig. 2: The core Encoder—Decoder architecture of the proposed OrbitALIF framework. The system model in Section I reveals three constraints for onboard cloud removal: tight power budgets, intermittent ISLs, and heterogeneous data across orbital planes. These call for a lightweight yet expressive backbone that extracts rich features under limited computation and generalizes across diverse conditions. To meet these constraints, we propose the orbital attention leaky integrate-and-fire (OrbitALIF) framework, as shown in Fig 2, which features a U-shaped architecture based on the widely adopted U-Net [3]. The input cloudy image ∈ℝ3×H×Wc ^3× H× W is embedded and replicated along a synthetic time axis with T time steps, enabling the network to process both spatial and temporal information simultaneously. Building upon this design, we introduce two key modules tailored for onboard cloud removal: • Adaptive Gated Fusion Module (AGFM). In a standard U-Net, encoder features are directly concatenated or summed with decoder features at the same resolution level. However, when clouds are dense or observations from different satellites vary significantly, the encoder features may contain information that is not perfectly aligned with the decoder’s current representation. AGFM replaces this rigid connection with a learnable gating mechanism, allowing the network to adaptively decide how much information to fuse at each spatial location. • Spectral-Spatial Hybrid Attention Module (SHAM). Inspired by [15], this module enhances the backbone by jointly attending to the temporal and frequency domains. Different time steps of the spiking rollout carry unequal information as the membrane state accumulates, and cloud artifacts occupy characteristic spatial-frequency ranges. SHAM learns to emphasize informative components in both domains. The final prediction is formed by a global residual connection: ^=+Conv3×3(1T∑t=1T[t]), y=c+Conv_3× 3\! ( 1T _t=1^Ts[t] ), (6) where ^∈ℝ3×H×W y ^3× H× W is the restored cloud-free image, [t]∈ℝC×H×Ws[t] ^C× H× W is the processed spike feature map at time step t with C channels, and Conv3×3Conv_3× 3 maps the temporally averaged features to 3 RGB channels. The global residual formulation lets the network learn only the cloud correction, reducing the optimization burden and stabilizing training. The complete model contains only 2.302.30 M parameters, making it suitable for resource-constrained satellite deployment. The two proposed modules are detailed in the following subsections. For further details regarding the backbone architecture, readers are referred to [12]. I-B Adaptive Gated Fusion Module (AGFM) In a standard U-Net, the skip connection at each level directly concatenates the encoder feature map with the corresponding decoder feature map. This design assumes that encoder features at a given resolution are always beneficial to the decoder at the same resolution. In practice, however, encoder and decoder features can have substantial semantic discrepancies since the encoder captures increasingly abstract representations, while the decoder reconstructs fine-grained spatial details. Under the non-IID data distributions inherent to satellite constellations, these discrepancies become more pronounced, and indiscriminate fusion can introduce noise rather than useful information. AGFM addresses this problem by inserting a lightweight learnable gate between each encoder–decoder skip connection. Specifically, let ∈ℝC×H×We ^C× H× W be the encoder feature and ∈ℝC×H×Wd ^C× H× W be the decoder feature at the same resolution level. The gated fusion is performed as: =(1−)⊙+⊙,f=(1-g) +g , (7) where ⊙ denotes element-wise multiplication, and the gate g is produced by applying a 1×11× 1 convolution to the concatenated features, followed by a sigmoid activation. Unlike a single scalar gate, g is both per-pixel and per-channel: it can suppress the encoder branch at heavily clouded pixels, where the contaminated encoder feature may hurt the output, while preserving it at clear pixels. AGFM offers three advantages for onboard deployment: minimal parameter overhead, improved representational capacity with negligible inference cost, and better generalization under heterogeneous orbital data distributions. I-C Spectral-Spatial Hybrid Attention Module (SHAM) While the U-Net backbone processes spatial features effectively, cloud removal demands reasoning across two additional dimensions: time and frequency. In the temporal domain, different time steps contribute unequally to the final prediction and some steps may carry more useful cloud information than others. In the frequency domain, cloud artifacts exhibit characteristic spectral signatures. Inspired by [15], SHAM jointly models information from the temporal and frequency domains. Temporal attention branch. Given the input feature tensor ∈ℝT×C×H×Wx ^T× C× H× W, this branch computes the temporal attention vector as =σ(mC[FC(λ1AvgPool()+λ2MaxPool())]),a=σ\! (m_C\! [FC\! ( _1AvgPool(x)+ _2MaxPool(x) ) ] ), (8) where λ1 _1 and λ2 _2 are learnable coefficients initialized to 0.50.5. Two pooling operations aggregate spatial dimensions while fully connected layer operates along time dimension. The resulting features are averaged over channels by mC(⋅)m_C(·) and passed through the sigmoid function σ(⋅)σ(·), producing ∈(0,1)Ta∈(0,1)^T. Each element ata_t rescales the feature map at simulation time step t. Spectral attention branch. This branch operates in the frequency domain to emphasize informative spectral bands while preserving spatial structure. It first computes the temporally averaged feature ¯=1T∑t[t] x= 1T _tx[t], then applies a 2-D real FFT to obtain the amplitude spectrum |ℱ(¯)||F( x)| and the phase ∠ℱ(¯) ( x). The amplitude is reweighted via a lightweight 1×11× 1 multilayer perceptron (MLP), while the phase is preserved unchanged. Phase encodes the spatial location of edges and structures, modifying it would distort image geometry. After inverse FFT, the result is passed through a 7×77× 7 convolution to produce a spatial attention map s∈ℝH×WA_s ^H× W. The final output of SHAM is given by SHAM()=+σ(λs)⋅(+s+s⊙)⊙,SHAM(x)\;=\;x+σ( _s)·(a+A_s+A_s ) , (9) where ∈ℝTa ^T is broadcast along the channel and spatial dimensions, scaling each time step t by [t]t[a_t]_t across all channels and pixels. Meanwhile, s∈ℝH×WA_s ^H× W is broadcast along the time and channel dimensions, scaling each spatial position (h,w)(h,w) by s[h,w]A_s[h,w] across all time steps and channels. λs∈ℝ _s is a learnable scalar initialized to zero. σ(⋅)σ(·) denotes the sigmoid activation function.The residual formulation ensures training stability. SHAM is well-suited for onboard training and inference for two reasons. First, the 2-D FFT operation is highly optimized on modern processors and adds low latency. Second, both the temporal MLP and the spectral 1×11× 1 MLP introduce negligible parameter overhead, preserving the 2.302.30 M-parameter budget of the full model. By adaptively weighting time steps and frequency components, SHAM enables the network to extract richer representations from the same computational budget, making it particularly effective for the diverse cloud patterns encountered across satellite orbits. I-D Training Objective The entire OrbitALIF network is trained with a combined Charbonnier–SSIM loss [16]. Specifically, the Charbonnier component penalizes pixel-wise reconstruction errors while remaining smooth near zero, whereas the SSIM component encourages consistency in local luminance and spatial structure, helping mitigate the over-smoothed textures. Their combination therefore provides complementary reconstruction constraints for preserving both fidelity and structural detail. The training objective is defined as ℒ(^,)=1|Ω|∑p∈Ω(y^p−yp)2+ε2+λ(1−SSIM(^,)),L( y,y)= 1| | _p∈ ( y_p-y_p)^2+ ^2+λ (1-SSIM( y,y) ), (10) where the first term is Charbonnier loss and the second is SSIM loss, SSIM(⋅,⋅)∈[−1,1]SSIM(·,·)∈[-1,1]. Ω is the set of all pixels in a figure, y^p y_p and ypy_p are the predicted and ground-truth values at pixel p, λ=0.1λ\!=\!0.1 balances the two terms. IV Experiments IV-A Centralized Evaluation and Ablation We first conduct centralized training on the CUHK-CR [7] dataset. All models are trained for 400400 epochs using AdamW (β1=0.9 _1=0.9, β2=0.999 _2=0.999, ϵ=10−8ε=10^-8, weight decay 00).Batch size is 44 for training and 11 for testing. The learning rate is linearly warmed up over the first three epochs and then annealed by a cosine schedule from 10−310^-3 to 10−710^-7. TABLE I: Centralized cloud removal on CUHK-CR1[7]. Ops. denotes MACs for ANN models and SOPs for OrbitALIF. w/o means “without”. Method Params (M) Ops. (G) PSNR (dB) CVAE [6] 15.56 37.13 24.252 MemoryNet [5] 3.64 548.65 26.073 OrbitALIF w/o SHAM 2.22 3.5 24.854 OrbitALIF w/o AGFM 2.30 3.7 24.774 OrbitALIF (ours) 2.30 3.7 25.374 Table I compares OrbitALIF against two state-of-the-art cloud-removal baselines. The first, CVAE [6], employs a Vision Transformer (ViT)-based encoder that maps the cloudy input to a latent distribution. The second, MemoryNet [5], introduces a three-granularity memory augment layer combined with contrastive learning to preserve and align deep features across multi-stage construction. OrbitALIF achieves 25.37425.374 dB PSNR with only 2.302.30 M parameters and 3.73.7 G SOPs—the lowest operation count among all methods. CVAE requires 6.8×6.8× more parameters and 10.0×10.0× more operations yet yields 1.1221.122 dB lower PSNR. MemoryNet attains the highest raw PSNR (26.07326.073 dB) but at over 148×148× the computational cost, making it impractical for satellite deployment. Ablation results further confirm the contribution of both proposed modules: removing SHAM drops PSNR to 24.85424.854 dB (with 2.222.22 M parameters and 3.53.5 G SOPs) and removing AGFM drops it to 24.77424.774 dB (with 2.302.30 M parameters and 3.73.7 G SOPs), both below 2525 dB. Since AGFM only introduces a lightweight 1×11× 1 gating convolution at each skip connection, its parameter and operation overhead is negligible. IV-B Decentralized Federated Learning and Energy Evaluation We further evaluate OrbitALIF under the decentralized federated learning setup described in Section I, using a 5×105×10 Walker-Star constellation (N=5N=5 orbital planes, K=10K=10 per plane, 5050 clients total). Client datasets are partitioned from CUHK-CR1 and CR2 via Dirichlet sampling (β=0.1β=0.1) over source labels, with at least 55 samples per client. The FL protocol runs for R=200R=200 global rounds,each comprising Eintra=2E_intra=2 intra-plane ring synchronizations, with Elocal=2E_local=2 local AdamW per client. To isolate the benefits of each design component, we construct two baseline architectures. • OrbitANN: Retains the full multi-domain architecture of OrbitALIF (AGFM, SHAM) but replaces all LIF neurons with ReLU. This baseline quantifies the energy–accuracy trade-off introduced by spiking computation: the ANN variant avoids spiking sparsity but executes all operations at full precision, incurring the dense computational cost of the multi-domain design. • OrbitUNet: A plain U-Net [17] that serves as a traditional ANN baseline for cloud removal. It uses standard convolutional encoder–decoder skip connections without any attention or frequency-domain modules, matching the backbone capacity to that of OrbitALIF. (a) (b) Fig. 3: Simulation results under the decentralized federated learning setup. (a) PSNR convergence curves. (b) Per-inference energy consumption. Energy is estimated by attaching forward hooks to record the multiply-accumulate operations (MACs) for ANN models and the spike-driven synaptic operations (SOPs) for OrbitALIF at each layer during inference, then multiplying the operation count and input spike rate at each layer: SOP=MAC×active rateSOP=MAC×active rate. ANN consumption is computed at 4.64.6 pJ/MAC (45 nm CMOS [18]); SNN consumption uses 7777 fJ/SOP, consistent with spiking neuromorphic processor [19] and prior works [20]. We first examine the PSNR convergence curves in Fig. 3(a). OrbitANN (22.67922.679 dB) outperforms OrbitUNet (22.01222.012 dB) by +0.667+0.667 dB, reflecting the attribution of two key proposed modules. First, AGFM replaces the rigid skip connections of a standard U-Net with a learnable gating mechanism, enabling the network to adaptively balance encoder and decoder features at each resolution level. Second, SHAM jointly attends to both the temporal and frequency domains, allowing the network to emphasize informative time steps and frequency bands. Meanwhile, OrbitALIF (21.88621.886 dB) trails OrbitANN by 0.7930.793 dB, reflecting the accuracy–efficiency trade-off introduced by spiking sparsity. Turning to energy consumption in Fig. 3(b), OrbitALIF consumes only 0.2870.287 mJ per inference on neuromorphic hardware (7777 fJ/SOP), a 72.3×72.3× (98.6%98.6\%) reduction compared with OrbitANN at 20.7520.75 mJ (4.64.6 pJ/MAC). Notably, OrbitANN consumes more energy than OrbitUNet, because the multi-domain architecture (AGFM, SHAM) is designed and optimized for SNN rather than ANN execution. When the same architecture is run as an ANN, all operations execute unconditionally, leading to inflated energy consumption. Despite this architectural mismatch, OrbitALIF still achieves an approximately 11×11× energy reduction even when compared with the OrbitUNet baseline. Figs. 3(a) and 3(b) demonstrate that the proposed spiking architecture delivers competitive cloud removal quality with dramatically lower energy consumption, validating that the brain-inspired design is well-suited for the tight power budgets of LEO satellites. Fig. 4 further illustrates the visual quality of OrbitALIF under decentralized training. Despite operating under the resource-constrained federated setting with only 2.302.30 M parameters, the predicted images effectively recover cloud-free surface details, confirming that the proposed framework achieves practical cloud removal performance while remaining feasible for onboard satellite deployment. Fig. 4: The cloud removal quality of OrbitALIF under decentralized training. V Conclusion We proposed OrbitALIF, the first framework to perform both onboard training and onboard inference for optical cloud removal on a LEO satellite constellation. Rather than compressing an existing ground-side model, we jointly designed the neuron activation, construction network, and federated aggregation protocol from the ground up for the constraints of Walker-Star operation. Our evaluation reveals two key findings. First, the brain-inspired spiking architecture achieves substantial energy savings, yielding a 72.3×72.3× reduction in energy consumption relative to its ANN counterpart (OrbitANN), at the cost of merely 0.7930.793 dB PSNR degradation. Second, the proposed SNN backbone outperforms the traditional U-Net (OrbitUNet) by +0.667+0.667 dB in its ANN variant, confirming the effectiveness of the AGFM and SHAM modules for cloud removal. Furthermore, under centralized evaluation, the resulting spiking architecture OrbitALIF achieves a competitive PSNR of 25.37425.374 dB while requiring only 2.302.30 M parameters and 3.73.7 G operations. These results collectively indicate that the lightweight design is well-suited for resource-constrained satellite deployment. Future work includes extending the framework to multi-spectral and SAR-assisted inputs, and validating the energy bracket with hardware-measured power on Loihi-2, Speck, or Akida substrates. References [1] S. Voigt et al. (2016) Global trends in satellite-based emergency mapping. Science 353 (6296), p. 247–252. External Links: Document Cited by: §I. [2] X. X. Zhu, D. Tuia, L. Mou, G. Xia, L. Zhang, F. Xu, and F. Fraundorfer (2017) Deep learning in remote sensing: a comprehensive review and list of resources. IEEE Geosci. Remote Sens. Mag. 5 (4), p. 8–36. External Links: Document Cited by: §I. [3] W. Li, Y. Li, D. Chen, and J. C. Chan (2019) Thin cloud removal with residual symmetrical concatenation network. ISPRS J. Photogramm. Remote Sens. 153, p. 137–150. External Links: ISSN 0924-2716, Document, Link Cited by: §I, §I-A. [4] A. Meraner, P. Ebel, X. X. Zhu, and M. Schmitt (2020) Cloud removal in Sentinel-2 imagery using a deep residual neural network and SAR-optical data fusion. ISPRS J. Photogramm. Remote Sens. 166, p. 333–346. External Links: ISSN 0924-2716, Document, Link Cited by: §I. [5] X. Zhang, X. Chen, C. Gu, S. Zhu, K. Tsang, and X. Guan (2026) Memory augment is all you need for image restoration. IEEE Trans. Consum. Electron. 72 (2), p. 3764–3772. Cited by: §I, §IV-A, TABLE I. [6] H. Ding, Y. Zi, and F. Xie (2023) Uncertainty-based thin cloud removal network via conditional variational autoencoders. In Proc. Asian Conf. Comput. Vis. (ACCV), L. Wang, J. Gall, T. Chin, I. Sato, and R. Chellappa (Eds.), Cham, p. 52–68. External Links: ISBN 978-3-031-26313-2 Cited by: §I, §IV-A, TABLE I. [7] J. Sui, Y. Ma, W. Yang, X. Zhang, M. Pun, and J. Liu (2024) Diffusion enhancement for cloud removal in ultra-resolution remote sensing imagery. IEEE Trans. Geosci. Remote Sens. 62 (), p. 1–14. External Links: Document Cited by: 3rd item, §I, §IV-A, TABLE I, TABLE I. [8] N. Razmi, B. Matthiesen, A. Dekorsy, and P. Popovski (2022) Ground-assisted federated learning in leo satellite constellations. IEEE Wireless Communications Letters 11 (4), p. 717–721. External Links: Document, ISSN 2162-2345 Cited by: §I. [9] Y. Shi, L. Zeng, J. Zhu, Y. Zhou, C. Jiang, and K. B. Letaief (2024) Satellite federated edge learning: architecture design and convergence analysis. IEEE Trans. Wireless Commun. 23 (10), p. 15212–15229. Cited by: §I. [10] M. Yang, J. Zhang, L. Wang, N. Vo, and X. Wang (2026) DFedSat: communication-efficient and robust decentralized federated learning for LEO satellite constellations. IEEE Trans. Cogn. Commun. Netw. 12 (), p. 4972–4985. Cited by: §I. [11] K. Roy, A. Jaiswal, and P. Panda (2019) Towards spike-based machine intelligence with neuromorphic computing. Nature 575 (7784), p. 607–617. External Links: Document Cited by: §I. [12] S. Chen, T. Krajnik, F. Arvin, and A. Atapour-Abarghouei (2026) Exploring the potentials of spiking neural networks for image deraining. Proc. AAAI Conf. Artif. Intell. 40 (4), p. 3029–3037. External Links: Document Cited by: §I, §I-A. [13] P. Yang, T. Wang, H. Cai, Y. Shi, C. Jiang, and L. Kuang (2025) Brain-inspired decentralized satellite learning in space computing power networks. IEEE Trans. Mobile Comput. 24 (12), p. 12935–12949. External Links: Document Cited by: §I-B, §I-B. [14] I. Loshchilov and F. Hutter (2019) Decoupled weight decay regularization. In Int. Conf. Learn. Represent. (ICLR), External Links: Link Cited by: §I-B. [15] K. Yu, T. Zhang, H. Wang, and Q. Xu (2025) FSTA-SNN: frequency-based spatial-temporal attention module for spiking neural networks. In Proc. AAAI Conf. Artif. Intell., Vol. 39, p. 22227–22235. External Links: Document Cited by: 2nd item, §I-C. [16] B. Zubair, W. Hu, and J. Peng (2025) Structural-perceptual image super resolution using Charbonnier-SSIM loss in an efficient sub-pixel convolutional network. In Photonics Electromagn. Res. Symp. Fall (PIERS-Fall), Vol. , p. 1–8. Cited by: §I-D. [17] O. Ronneberger, P. Fischer, and T. Brox (2015) U-Net: convolutional networks for biomedical image segmentation. In Int. Conf. Med. Image Comput. Comput.-Assist. Interv. (MICCAI), Vol. 9351, p. 234–241. External Links: Document Cited by: 2nd item. [18] M. Horowitz (2014) 1.1 computing’s energy problem (and what we can do about it). In IEEE Int. Solid-State Circuits Conf. (ISSCC), Vol. , p. 10–14. Cited by: §IV-B. [19] N. Qiao, H. Mostafa, F. Corradi, M. Osswald, F. Stefanini, D. Sumislawska, and G. Indiveri (2015) A reconfigurable on-line learning spiking neuromorphic processor comprising 256 neurons and 128K synapses. Front. Neurosci. 9. External Links: Document, ISSN 1662-453X Cited by: §IV-B. [20] G. Indiveri, F. Corradi, and N. Qiao (2015) Neuromorphic architectures for spiking deep neural networks. In IEEE Int. Electron Devices Meeting (IEDM), Vol. , p. 4.2.1–4.2.4. External Links: Document Cited by: §IV-B.