ELMZip: Onboard Satellite Image Compression via Extreme Learning Machines for Efficient Downlink

# ELMZIP: ONBOARD SATELLITE IMAGE COMPRESSION VIA EXTREME LEARNING MACHINES FOR EFFICIENT DOWNLINK
Source: [https://arxiv.org/html/2608.06942](https://arxiv.org/html/2608.06942)
Woojin Cho Junghwan Park Sangcheol Sim Steve Andreas Immanuel Junhyuk Heo Darongsae Kwon

###### Abstract

The acquisition of multispectral imagery via small satellites \(e\.g\., CubeSats\) presents significant data downlink challenges due to high data volumes and restricted communication windows\. While onboard image compression is critical to address this bottleneck, traditional methods often struggle to adapt to the nonlinear statistics of multi\-band, multi\-resolution data\. To overcome these limitations, we propose ELMZip, a novel framework based on Extreme Learning Machines \(ELM\) and domain decomposition strategies for efficient, resolution\-free onboard neural representation\. ELMZip formulates the fitting process as a convex least\-squares problem using random\-feature single\-layer networks, thereby eliminating the need for computationally expensive backpropagation\. By adopting an asymmetric transmission protocol that sends only the compact output weights, the proposed method significantly reduces the downlink payload\. Unlike previous neural representation approaches that rely on iterative optimization and require transmitting full network parameters, ELMZip achieves significant compression efficiency while maintaining high reconstruction fidelity\. This capability enables immediate image reconstruction for analysis, allowing resource\-constrained platforms to maximize data return and advancing real\-time AI\-powered Earth observation\.

## IIntroduction

Small satellites such as CubeSats are transforming Earth observation, yet their onboard computing and communication capabilities remain tightly constrained by limited power budgets, passive thermal management, radiation effects, and short downlink windows\. As sensor performance improves, the resulting growth in high\-resolution multispectral data increasingly exceeds available satellite\-to\-ground bandwidth, forcing operators to discard data, delay transmission, or rely on conventional compression standards such as CCSDS 123\.0\[[9](https://arxiv.org/html/2608.06942#bib.bib62)\]and JPEG\[[26](https://arxiv.org/html/2608.06942#bib.bib12)\]\. While these codecs are effective, their fixed transforms may not fully adapt to the highly nonlinear, scene\-dependent statistics of diverse land covers\.

Motivated by these operational constraints, onboard edge computing has gained momentum with the availability of low\-power AI hardware \(e\.g\., the NVIDIA Jetson series\)\[[17](https://arxiv.org/html/2608.06942#bib.bib44),[13](https://arxiv.org/html/2608.06942#bib.bib55),[3](https://arxiv.org/html/2608.06942#bib.bib56),[8](https://arxiv.org/html/2608.06942#bib.bib60)\], enabling more flexible and task\-aware data reduction directly on spacecraft\. In parallel, implicit neural representations \(INRs\) have been explored as an alternative data representation for compression by modeling imagery as a continuous function over spatial coordinates rather than a discrete pixel grid\[[6](https://arxiv.org/html/2608.06942#bib.bib43),[4](https://arxiv.org/html/2608.06942#bib.bib39)\]\. Despite strong reconstruction quality, widely used INR architectures such as SIREN\[[22](https://arxiv.org/html/2608.06942#bib.bib16)\]or WIRE\[[21](https://arxiv.org/html/2608.06942#bib.bib45)\]typically require iterative backpropagation, leading to substantial latency and energy consumption that are difficult to accommodate during time\-limited orbital passes\. Moreover, transmitting full network parameters can diminish the practical compression gains, limiting their utility for bandwidth\-constrained missions\.

![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/elmzip_concept_new.png)Figure 1:Traditional satellite communication transmits raw image data, whereas the proposed method transmits compact neural representation parameters for faster ground reconstruction\.In this work, we propose Extreme Learning Machine–based neural compression \(ELMZip\), a novel framework designed specifically for onboard satellite image compression\. Drawing inspiration from domain decomposition\-based ELM architectures\[[1](https://arxiv.org/html/2608.06942#bib.bib46),[25](https://arxiv.org/html/2608.06942#bib.bib47)\]used for solving partial differential equations, we adapt the principles of domain decomposition and Extreme Learning Machines \(ELM\)\[[10](https://arxiv.org/html/2608.06942#bib.bib1)\]to the problem of multispectral image regression\. Unlike conventional deep learning methods, ELMZip reformulates per\-image training as a convex linear least\-squares problem, enabling an analytic solution in seconds without iterative backpropagation\.

![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/elmzip_new.png)Figure 2:Overview of ELMZip: the satellite fits only the output weightsμk\\mu^\{k\}using fixed random featuresθfix\\theta\_\{\\text\{fix\}\}pre\-shared between the satellite and the ground station, and transmits only theμk\\mu^\{k\}, enabling the ground station to reconstructV^​\(𝒳;Θk\)\\hat\{V\}\(\\mathcal\{X\};\\Theta^\{k\}\)\.To maximize downlink efficiency, ELMZip adopts an asymmetric transmission protocol\. By synchronizing the random feature initialization between the satellite and the ground segment \(via a shared deterministic initialization\), the satellite needs to transmit only a compact set of output parameters \(i\.e\., the final\-layer weights\), rather than the full network parameters as in standard INR pipelines\[[5](https://arxiv.org/html/2608.06942#bib.bib57)\]\. As a result, the ground station can reconstruct a high\-fidelity preview rapidly, enabling practical operational decisions—such as prioritizing scenes for immediate download—and supporting downstream remote\-sensing workflows, e\.g\., vegetation monitoring\[[7](https://arxiv.org/html/2608.06942#bib.bib26),[15](https://arxiv.org/html/2608.06942#bib.bib27)\], urban change detection\[[2](https://arxiv.org/html/2608.06942#bib.bib58)\], and few\-shot segmentation\[[11](https://arxiv.org/html/2608.06942#bib.bib61)\]without waiting for full data transfer\.

Overall, ELMZip provides an effective trade\-off between reconstruction quality and transmission cost, offering a deployment\-oriented neural compression strategy for resource\-constrained Earth observation platforms\. The main contributions of this work are summarized as:

- •We introduce ELMZip, an ELM\-based neural compression method that enables fast onboard fitting via a convex least\-squares formulation without backpropagation\.
- •We present an asymmetric downlink protocol that transmits only compact output parameters for efficient reconstruction and preview generation at the ground station\.
- •To assess general performance, we evaluate our proposed method on Level\-0 and Level\-1C datasets covering diverse natural environments\.

## IIBackground and Related Work

### II\-AImplicit Neural Representations for Images

INRs fundamentally alter how data is stored and processed\[[23](https://arxiv.org/html/2608.06942#bib.bib17),[12](https://arxiv.org/html/2608.06942#bib.bib53),[6](https://arxiv.org/html/2608.06942#bib.bib43)\]\. Instead of storing a discrete grid of pixel intensitiesV∈ℝH×W×CV\\in\\mathbb\{R\}^\{H\\times W\\times C\}, an INR learns a continuous mapping functionΦ:ℝ2→ℝC\\Phi:\\mathbb\{R\}^\{2\}\\rightarrow\\mathbb\{R\}^\{C\}that maps 2D spatial coordinates𝒳=\(x,y\)\\mathcal\{X\}=\(x,y\)to their corresponding spectral intensity valuesV​\(𝒳\)V\(\\mathcal\{X\}\)\. A standard Multi\-layer Perceptron \(MLP\) acts as the function approximator, defined recursively as follows\.

Φ​\(𝒳;Θ\)=WL​\(σL−1​\(…​σ1​\(W1⋅𝒳\+b1\)​…\)\)\+bL,\\Phi\(\\mathcal\{X\};\\Theta\)=W\_\{L\}\(\\sigma\_\{L\-1\}\(\\dots\\sigma\_\{1\}\(W\_\{1\}\\cdot\\mathcal\{X\}\+b\_\{1\}\)\\dots\)\)\+\{b\}\_\{L\},\(1\)whereWlW\_\{l\}andblb\_\{l\}represent the weights and biases of thell\-th layer,LLis the number of layers, andσ\\sigmais a nonlinear activation function\. While standard ReLU networks fail to capture high\-frequency image details\[[24](https://arxiv.org/html/2608.06942#bib.bib51)\], recent advancements have introduced periodic activations to overcome this spectral bias\[[22](https://arxiv.org/html/2608.06942#bib.bib16),[14](https://arxiv.org/html/2608.06942#bib.bib49),[27](https://arxiv.org/html/2608.06942#bib.bib50)\]\. For instance, SIREN\[[22](https://arxiv.org/html/2608.06942#bib.bib16)\]employs a sine activation function, allowing the network to model fine details and their derivatives effectively\. Similarly, WIRE\[[21](https://arxiv.org/html/2608.06942#bib.bib45)\]utilizes a complex Gabor wavelet activation,ei​ω0​𝒳⋅e−\|𝒳\|2e^\{i\\omega\_\{0\}\\mathcal\{X\}\}\\cdot e^\{\-\|\\mathcal\{X\}\|^\{2\}\}, to provide optimally concentrated space\-frequency localization\. Despite their reconstruction quality, these architectures require finding the optimal parametersΘ\\Thetaby minimizing a non\-convex loss functionℒ=∑‖Φ​\(𝒳;Θ\)−V​\(𝒳\)‖2\\mathcal\{L\}=\\sum\|\|\\Phi\(\\mathcal\{X\};\\Theta\)\-V\(\\mathcal\{X\}\)\|\|^\{2\}via gradient descent\. This iterative optimization is prohibitively slow for onboard processing, often taking minutes to hours to converge, which is incompatible with the real\-time constraints of Low Earth Orbit \(LEO\) operations\.

### II\-BExtreme Learning Machines

To circumvent the latency of gradient\-based training, we look to ELM\. The ELM theory posits that for a feedforward network, the input weights and biasesθfix=\{Wfix,bfix\}\\theta\_\{\\text\{fix\}\}=\\\{W\_\{\\text\{fix\}\},b\_\{\\text\{fix\}\}\\\}need not be tuned iteratively\. Instead, they can be randomly assigned from a continuous probability distribution and fixed\. The training process is thus reduced to learning only the output weightsμ\\mu\. The ELM output is expressed as follows\.

Φ​\(𝒳;Θ\)=𝐇⋅μ,𝐇=σ​\(Wfix⋅𝒳\+bfix\)\\Phi\(\\mathcal\{X\};\\Theta\)=\\mathbf\{H\}\\cdot\\mu,\\quad\\mathbf\{H\}=\\sigma\(W\_\{\\text\{fix\}\}\\cdot\\mathcal\{X\}\+b\_\{\\text\{fix\}\}\)\(2\)This formulation allows the optimal output weightsμ\\muto be computed analytically using the output matrix𝐇\\mathbf\{H\}\.

## IIIProposed Method: ELMZip

We propose the ELMZip framework, which adapts the ELM with domain decomposition frameworks to the problem of high\-speed multispectral image compression\.

The core objective is to approximate the multispectral image functionV​\(𝒳\)V\(\\mathcal\{X\}\)using a collection of localized Extreme Learning Machines, enabling highly efficient data transmission\. Our ELMZip framework can be summarized as follows:

- •Onboard Encoding:The satellite generates the random input layer using a fixed initialization, computes the analytical solution for the output weightsμ\\mu\.
- •Transmission:The downlink payload consists only of the quantized output weightsμ\\mu\. The input weights and biasesθfix=\{Wfix,bfix\}\\theta\_\{\\text\{fix\}\}=\\\{W\_\{\\text\{fix\}\},b\_\{\\text\{fix\}\}\\\}are never transmitted\.
- •Ground Decoding:The ground station receivesμ\\mufrom onboard satellite and combines it with the shared fixed initializationθfix\\theta\_\{\\text\{fix\}\}to reconstruct the high\-dimensional multispectral satellite image\.

TABLE I:Geographic Coordinates of Benchmark DatasetDataLevelLatitudeLongitudeEnvironmentsAntucoL037∘35’05”S71∘13’37”WVolcanoPusztaL047∘28’18”N19∘53’27”EGrasslandAndamanL012∘26’54”N93∘56’27”EMarineCairoL1C30∘01’29”N31∘55’18”EDesertMerapiL1C07∘32’29”S110∘26’46”EVolcanoSeoulL1C37∘31’30”N126∘55’36”EUrban

![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/L0_antuco.png)\(a\)Antuco \(L0\)
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/L0_puszta.png)\(b\)Puszta \(L0\)
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/L0_andaman.png)\(c\)Andaman \(L0\)
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/L1_cairo_band03.png)\(d\)Cairo \(L1C\)
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/L1_Merapi_band08.png)\(e\)Merapi \(L1C\)
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/L1_Seoul_band08.png)\(f\)Seoul \(L1C\)

Figure 3:Sentinel\-2 benchmark dataset\. The dataset consists of six distinct regions covering diverse environments across both Level\-0 \(\(a\)\-\(c\)\) and Level\-1C \(\(d\)\-\(f\)\) data products\.### III\-AExtreme Learning Machine with Domain Decomposition

A single global ELM struggles to capture the variations of a high\-resolution satellite scene due to global spectral bias\[[19](https://arxiv.org/html/2608.06942#bib.bib48)\]\. To address this, we employ the domain decomposition strategy found in\[[18](https://arxiv.org/html/2608.06942#bib.bib52),[1](https://arxiv.org/html/2608.06942#bib.bib46)\]\.

Given a high\-resolution multispectral image, we first decompose the spatial domainΩ\\OmegaintoKKoverlapping subdomains\{Ωk\}k=1K\\\{\\Omega\_\{k\}\\\}\_\{k=1\}^\{K\}\. For each subdomain, we assign a local ELM approximatorϕk​\(𝒳;Θk\)\\phi^\{k\}\(\\mathcal\{X\};\\Theta^\{k\}\)\. We utilize sine function for the activation functionσ\\sigma\. The global image reconstruction is defined as the weighted sum of these local models:

V^​\(𝒳;Θ\)=∑k=1Kωk​\(𝒳\)⋅ϕk​\(𝒳;Θk\),Θ=\{θfixk,μk\}k=1K\\hat\{V\}\(\\mathcal\{X\};\{\\Theta\}\)=\\sum\_\{k=1\}^\{K\}\\omega^\{k\}\(\\mathcal\{X\}\)\\cdot\\phi^\{k\}\(\\mathcal\{X\};\{\{\\Theta\}^\{k\}\}\),\\quad\\Theta=\\\{\\theta^\{k\}\_\{\\text\{fix\}\},\\mu^\{k\}\\\}\_\{k=1\}^\{K\}\(3\)whereωk​\(𝒳\)\\omega^\{k\}\(\\mathcal\{X\}\)are smooth window functions forming a partition of unity such that∑k=1Kωk​\(𝒳\)=1\\sum\_\{k=1\}^\{K\}\\omega^\{k\}\(\\mathcal\{X\}\)=1\. Concretely, the input parametersθfixk=\{Wnk,bnk\}\\theta^\{k\}\_\{\\text\{fix\}\}=\\\{\{W\}\_\{n\}^\{k\},b\_\{n\}^\{k\}\\\}are initialized using a built\-in function in PyTorch\. By sharing the fixed parametersθfix=\{θfixk\}k=1K\\theta\_\{\\text\{fix\}\}=\\\{\\theta\_\{\\text\{fix\}\}^\{k\}\\\}\_\{k=1\}^\{K\}between the satellite and the ground station, these parameters become deterministic and do not need to be transmitted\.

### III\-BOptimization for ELMZip

Unlike standard INRs that minimize loss via backpropagation, ELMZip formulates the training as a linear least\-squares problem\. As shown in Fig\.[2](https://arxiv.org/html/2608.06942#S1.F2), for a specific subdomainkk, we aim to minimize the reconstruction error between the model output and the ground truth pixel valuesVgtkV^\{k\}\_\{\\text\{gt\}\}\. Since the input projection parameters are fixed, the optimization objective for the output weightsμk\\mu^\{k\}is strictly convex:

minμk⁡12​‖𝐇k​μk−Vgtk‖22\\min\_\{\\mu^\{k\}\}\\frac\{1\}\{2\}\\\|\\mathbf\{H\}^\{k\}\\mu^\{k\}\-\{V\}^\{k\}\_\{\\text\{gt\}\}\\\|\_\{2\}^\{2\}\(4\)In this equation,𝐇k\\mathbf\{H\}^\{k\}is the hidden layer activation matrix computed as𝐇k=ωk​\(𝒳\)⋅σ​\(Wfixk⋅𝒳\+bfixk\)\\mathbf\{H\}^\{k\}=\\omega^\{k\}\(\\mathcal\{X\}\)\\cdot\\sigma\(W\_\{\\text\{fix\}\}^\{k\}\\cdot\\mathcal\{X\}\+b\_\{\\text\{fix\}\}^\{k\}\)\. The optimal solutionμ^k\\hat\{\\mu\}^\{k\}is obtained in a single computational step by solving the normal equations:

μ^k=\(𝐇k⊤​𝐇k\)−1​𝐇k⊤​Vg​tk\\hat\{\\mu\}^\{k\}=\\left\(\\mathbf\{H\}^\{k\\top\}\\mathbf\{H\}^\{k\}\\right\)^\{\-1\}\\mathbf\{H\}^\{k\\top\}V^\{k\}\_\{gt\}\(5\)This closed\-form solution reduces training to matrix multiplications and a linear solve, which can be efficiently accelerated by CUDA\-optimized linear algebra libraries on embedded GPUs such as the NVIDIA Jetson Nano\. As a result, the model can fit each image patch rapidly\.

TABLE II:PSNR\(↑\\uparrow\) and SSIM\(↑\\uparrow\) comparison for the ELMZip \(ours\) against baseline methodsModelAntuco \(L0\)Puszta \(L0\)Andaman \(L0\)Cairo \(L1C\)Merapi \(L1C\)Seoul \(L1C\)PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMMLP14\.5210\.47517\.9790\.46621\.1130\.62015\.4330\.77213\.6890\.41012\.8770\.211SIREN22\.3020\.94023\.9700\.90624\.3880\.85417\.8420\.88315\.8520\.71516\.0260\.735FFN22\.8180\.94822\.1770\.82524\.7040\.85617\.9920\.88415\.7690\.67116\.0970\.720GaussNet16\.7410\.74421\.0020\.79322\.5350\.76416\.8430\.84615\.0160\.63114\.9550\.624WIRE16\.4690\.72519\.5580\.72121\.7590\.71615\.4740\.78714\.7050\.60814\.4610\.582ELMZip26\.9430\.98132\.4260\.98830\.0390\.96520\.9610\.94620\.8470\.92519\.5850\.899

![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/L0_antuco.png)\(a\)Ground Truth
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L0_antuco_mlp_baseline.png)\(b\)MLP
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L0_antuco_siren_baseline.png)\(c\)SIREN
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L0_antuco_ffn_baseline.png)\(d\)FFN
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L0_antuco_gaussnet_baseline.png)\(e\)GaussNet
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L0_antuco_wire_baseline.png)\(f\)WIRE
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/ELM_L0_antuco.png)\(g\)ELMZip
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/L1_cairo_band03.png)\(h\)Ground Truth
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L1_cairo_band03_mlp_baseline.png)\(i\)MLP
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L1_cairo_band03_siren_baseline.png)\(j\)SIREN
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L1_cairo_band03_ffn_baseline.png)\(k\)FFN
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L1_cairo_band03_gaussnet_baseline.png)\(l\)GaussNet
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/L1_cairo_band03_wire_baseline.png)\(m\)WIRE
![Refer to caption](https://arxiv.org/html/2608.06942v1/Figures/exp_result/ELM_L1_cairo_band03.png)\(n\)ELMZip

Figure 4:Qualitative reconstruction comparisons on Sentinel\-2 MSI under unequal energy budgets\.The top row shows Antuco \(L0, B8\) and the bottom row shows Cairo \(L1C, B3\)\. INR baselines are optimized with approximately 10×\\timeshigher electrical energy than ELMZip, yet ELMZip produces reconstructions that are visually closest to the ground truth\.

## IVExperiments

To validate the proposed method under realistic onboard constraints, we utilized Sentinel\-2 MSI Level\-0 and Level\-1C data collected from six distinct geographic regions representing diverse land covers\[[16](https://arxiv.org/html/2608.06942#bib.bib54)\], as summarized in Table[I](https://arxiv.org/html/2608.06942#S3.T1)and visualized in Fig\.[3](https://arxiv.org/html/2608.06942#S3.F3)\. In particular, to approximate an onboard satellite scenario as closely as possible, our benchmark suite explicitly includes Level\-0 \(L0\) products, which are the most similar to raw sensor measurements and therefore capture the practical characteristics of minimally processed imagery available in early downlink pipelines\. The onboard processing environment was simulated using an NVIDIA Jetson Nano, a representative low\-power edge computing platform for nanosatellites\.

### IV\-AExperimental Setup

We compared ELMZip against representative INR baselines, including an MLP with ReLU, SIREN\[[22](https://arxiv.org/html/2608.06942#bib.bib16)\], Fourier Feature Networks \(FFN\)\[[24](https://arxiv.org/html/2608.06942#bib.bib51)\], GaussNet\[[20](https://arxiv.org/html/2608.06942#bib.bib59)\], and WIRE\[[21](https://arxiv.org/html/2608.06942#bib.bib45)\]\. All methods were implemented in PyTorch\. For a controlled and fair comparison, all benchmark images were standardized to a fixed spatial dimension of10241024so that the evaluation is not confounded by varying scene sizes or different coordinate sampling densities\. We also designed the evaluation to reflect practical downlink constraints by enforcing neural compression to achieve at least an approximately10×10\\timesreduction in payload relative to the uncompressed image\. This was implemented by constraining the transmitted representation size across methods, for example through a common parameter budget and quantized payload settings\. We measure the electrical energy consumed during fitting\. Due to iterative backpropagation, INR baselines require approximately10×10\\timesmore energy than ELMZip to reach their reported performance\. We evaluated performance using two key metrics: Peak Signal\-to\-Noise Ratio \(PSNR\) to measure reconstruction fidelity, and Structural Similarity Index Measure \(SSIM\) to assess perceptual quality\.

### IV\-BExperimental Results

Table[II](https://arxiv.org/html/2608.06942#S3.T2)summarizes reconstruction fidelity across six Sentinel\-2 scenes for both Level\-0 and Level\-1C products\. Importantly, the compared INR baselines are evaluated using substantially larger power/energy budgets than ELMZip\. Specifically, each INR consumes approximately 10×\\timesmore electrical energy than ELMZip to obtain the reported results, reflecting the higher computational cost of iterative backpropagation\-based optimization\.

Despite this disadvantage in resource budget, ELMZip achieves the best reconstruction quality across all scenes, consistently outperforming all baselines in PSNR and SSIM\. This trend is especially pronounced in structurally complex regions\. For Cairo and Seoul, which contain dense man\-made structures and heterogeneous textures, ELMZip improves PSNR by 16\.5% and 21\.7%, respectively, over the strongest INR baseline, even though the INR baseline uses 10×\\timeshigher energy\. We provide detailed visualization of the experimental results in Fig\.[4](https://arxiv.org/html/2608.06942#S3.F4)\. While INR baselines trained with much higher energy budgets still exhibit over\-smoothing and residual artifacts in fine structures, ELMZip reconstructs sharper edges and preserves local contrast more faithfully, producing outputs that most closely match the ground truth\.

## VDiscussion

ELMZip demonstrates a practical pathway to rapid onboard compression and fast preview reconstruction under tight downlink and power constraints\. By formulating the image fitting process as a convex least\-squares problem and transmitting only the output parameters, the method supports timely screening of acquired data at the ground station, which is beneficial for operational workflows such as preview\-based quality assessment, vegetation monitoring, and urban change detection\. In addition, its low\-energy fitting aligns well with the strict power budgets of small satellite platforms\. Future work will extend the evaluation to larger\-scale benchmarks and explore numerical solvers to improve robustness and efficiency under constrained onboard memory and compute\.

## References

- \[1\]\(2024\)ELM\-fbpinn: efficient finite\-basis physics\-informed neural networks\.arXiv preprint arXiv:2409\.01949\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p3.1),[§III\-A](https://arxiv.org/html/2608.06942#S3.SS1.p1.1)\.
- \[2\]K\. Basavaraju, N\. Sravya,et al\.\(2024\)Recent advances in urban expansion monitoring through deep learning\-based semantic change detection techniques from satellite imagery\.In2024 IEEE Space, Aerospace and Defence Conference \(SPACE\),pp\. 169–173\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p4.1)\.
- \[3\]N\. Buonaiuto, M\. Louie, J\. Aarestad, R\. Mital, D\. Mateik, R\. Sivilli, A\. Bhopale, C\. Kief, and B\. Zufelt\(2017\)Satellite identification imaging for small satellites using nvidia\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p2.1)\.
- \[4\]Z\. Chen, Y\. Chen, J\. Liu, X\. Xu, V\. Goel, Z\. Wang, H\. Shi, and X\. Wang\(2022\)Videoinr: learning video implicit neural representation for continuous space\-time super\-resolution\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 2047–2057\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p2.1)\.
- \[5\]W\. Cho, S\. A\. Immanuel, J\. Heo, and D\. Kwon\(2025\)Fourier\-modulated implicit neural representation for multispectral satellite image compression\.arXiv preprint arXiv:2506\.01234\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p4.1)\.
- \[6\]W\. Cho, M\. Jo, K\. Lee, and N\. Park\(2025\)Neural functions for learning periodic signal\.arXiv preprint arXiv:2506\.09526\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p2.1),[§II\-A](https://arxiv.org/html/2608.06942#S2.SS1.p1.4)\.
- \[7\]E\. Dwyer, S\. Pinnock, J\. Grégoire, and J\. Pereira\(2000\)Global spatial and temporal distribution of vegetation fire as determined from satellite observations\.International Journal of Remote Sensing21\(6\-7\),pp\. 1289–1302\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p4.1)\.
- \[8\]G\. Giuffrida, L\. Fanucci, G\. Meoni, M\. Batič, L\. Buckley, A\. Dunne, C\. Van Dijk, M\. Esposito, J\. Hefele, N\. Vercruyssen,et al\.\(2021\)TheΦ\\Phi\-sat\-1 mission: the first on\-board deep neural network demonstrator for satellite earth observation\.IEEE Transactions on Geoscience and Remote Sensing60,pp\. 1–14\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p2.1)\.
- \[9\]M\. Hernández\-Cabronero, A\. B\. Kiely, M\. Klimesh, I\. Blanes, J\. Ligo, E\. Magli, and J\. Serra\-Sagrista\(2021\)The ccsds 123\.0\-b\-2 “low\-complexity lossless and near\-lossless multispectral and hyperspectral image compression” standard: a comprehensive review\.IEEE Geoscience and Remote Sensing Magazine9\(4\),pp\. 102–119\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p1.1)\.
- \[10\]G\. Huang, Q\. Zhu, and C\. Siew\(2006\)Extreme learning machine: theory and applications\.Neurocomputing70\(1\-3\),pp\. 489–501\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p3.1)\.
- \[11\]S\. A\. Immanuel, W\. Cho, J\. Heo, and D\. Kwon\(2025\)Tackling few\-shot segmentation in remote sensing via inpainting diffusion model\.arXiv preprint arXiv:2503\.03785\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p4.1)\.
- \[12\]M\. Jo, W\. Cho, U\. B\. Mudiyanselage, S\. Lee, N\. Park, and K\. Lee\(2025\)PDEfuncta: spectrally\-aware neural representation for pde solution modeling\.arXiv preprint arXiv:2506\.12790\.Cited by:[§II\-A](https://arxiv.org/html/2608.06942#S2.SS1.p1.4)\.
- \[13\]L\. S\. Karumbunathan\(2022\)Nvidia jetson agx orin series\.A Giant Leap Forward for Robotics and Edge AI Applications\. Technical Brief\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p2.1)\.
- \[14\]Z\. Liu, H\. Zhu, Q\. Zhang, J\. Fu, W\. Deng, Z\. Ma, Y\. Guo, and X\. Cao\(2024\)Finer: flexible spectral\-bias tuning in implicit neural representation by variable\-periodic activation functions\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 2713–2722\.Cited by:[§II\-A](https://arxiv.org/html/2608.06942#S2.SS1.p1.12)\.
- \[15\]R\. Ma, H\. Duan, X\. Gu, and S\. Zhang\(2008\)Detecting aquatic vegetation changes in taihu lake, china using multi\-temporal satellite imagery\.Sensors8\(6\),pp\. 3988–4005\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p4.1)\.
- \[16\]G\. Meoni, R\. Del Prete, F\. Serva, A\. De Beusscher, O\. Colin, and N\. Longépé\(2024\)Unlocking the use of raw multispectral earth observation imagery for onboard artificial intelligence\.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing17,pp\. 12521–12537\.Cited by:[§IV](https://arxiv.org/html/2608.06942#S4.p1.1)\.
- \[17\]S\. Mittal\(2019\)A survey on optimized implementation of deep learning models on the nvidia jetson platform\.Journal of Systems Architecture97,pp\. 428–442\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p2.1)\.
- \[18\]B\. Moseley, A\. Markham, and T\. Nissen\-Meyer\(2023\)Finite basis physics\-informed neural networks \(fbpinns\): a scalable domain decomposition approach for solving differential equations\.Advances in Computational Mathematics49\(4\),pp\. 62\.Cited by:[§III\-A](https://arxiv.org/html/2608.06942#S3.SS1.p1.1)\.
- \[19\]N\. Rahaman, A\. Baratin, D\. Arpit, F\. Draxler, M\. Lin, F\. Hamprecht, Y\. Bengio, and A\. Courville\(2019\)On the spectral bias of neural networks\.InInternational conference on machine learning,pp\. 5301–5310\.Cited by:[§III\-A](https://arxiv.org/html/2608.06942#S3.SS1.p1.1)\.
- \[20\]S\. Ramasinghe and S\. Lucey\(2022\)Beyond periodicity: towards a unifying framework for activations in coordinate\-mlps\.InEuropean Conference on Computer Vision,pp\. 142–158\.Cited by:[§IV\-A](https://arxiv.org/html/2608.06942#S4.SS1.p1.3)\.
- \[21\]V\. Saragadam, D\. LeJeune, J\. Tan, G\. Balakrishnan, A\. Veeraraghavan, and R\. G\. Baraniuk\(2023\)Wire: wavelet implicit neural representations\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 18507–18516\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p2.1),[§II\-A](https://arxiv.org/html/2608.06942#S2.SS1.p1.12),[§IV\-A](https://arxiv.org/html/2608.06942#S4.SS1.p1.3)\.
- \[22\]V\. Sitzmann, J\. Martel, A\. Bergman, D\. Lindell, and G\. Wetzstein\(2020\)Implicit neural representations with periodic activation functions\.Advances in neural information processing systems33,pp\. 7462–7473\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p2.1),[§II\-A](https://arxiv.org/html/2608.06942#S2.SS1.p1.12),[§IV\-A](https://arxiv.org/html/2608.06942#S4.SS1.p1.3)\.
- \[23\]Y\. Strümpler, J\. Postels, R\. Yang, L\. V\. Gool, and F\. Tombari\(2022\)Implicit neural representations for image compression\.InECCV,Cited by:[§II\-A](https://arxiv.org/html/2608.06942#S2.SS1.p1.4)\.
- \[24\]M\. Tancik, P\. Srinivasan, B\. Mildenhall, S\. Fridovich\-Keil, N\. Raghavan, U\. Singhal, R\. Ramamoorthi, J\. Barron, and R\. Ng\(2020\)Fourier features let networks learn high frequency functions in low dimensional domains\.Advances in neural information processing systems33,pp\. 7537–7547\.Cited by:[§II\-A](https://arxiv.org/html/2608.06942#S2.SS1.p1.12),[§IV\-A](https://arxiv.org/html/2608.06942#S4.SS1.p1.3)\.
- \[25\]J\. W\. van Beek, V\. Dolean, and B\. Moseley\(2026\)Local feature filtering for scalable and well\-conditioned domain\-decomposed random feature methods\.Computer Methods in Applied Mechanics and Engineering449,pp\. 118583\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p3.1)\.
- \[26\]G\. K\. Wallace\(1992\)The jpeg still picture compression standard\.Communications of the ACM34\(4\),pp\. 30–44\.Cited by:[§I](https://arxiv.org/html/2608.06942#S1.p1.1)\.
- \[27\]T\. Yeom, S\. Lee, and J\. Lee\(2024\)Fast training of sinusoidal neural fields via scaling initialization\.arXiv preprint arXiv:2410\.04779\.Cited by:[§II\-A](https://arxiv.org/html/2608.06942#S2.SS1.p1.12)\.