On-Orbit Super-Resolution Techniques

Updated 23 April 2026

On-orbit super-resolution is a set of computational methods that reconstruct high-resolution images from low-resolution orbital data while contending with SWaP and bandwidth constraints.
It integrates classical multi-frame approaches, deep learning models, and physics-informed techniques to overcome limitations in optics, detectors, and motion-induced blur.
Recent advances, including quantum estimation and efficient onboard processing strategies, significantly improve image resolution for applications in Earth observation and astronomy.

On-orbit super-resolution is the class of computational methods that enable satellites and astronomical observatories to reconstruct high-resolution (HR) images from one or more low-resolution (LR) measurements acquired in orbit, under constraints of size, weight, power (SWaP), data bandwidth, and onboard compute. These algorithms address the physical limitations of optical, detector, and mission architectures, allowing satellites to deliver imagery with enhanced spatial fidelity beyond detector native sampling. The field encompasses classical multi-frame or single-frame reconstruction, deep-learning paradigms, physics-based inverse problems, and quantum measurement strategies, with demonstrated applications across planetary science, Earth observation, astronomy, and inter-satellite ranging.

1. Imaging Models and On-Orbit Constraints

On-orbit imaging systems introduce a cascade of physical degradation operators, including optical and atmospheric point-spread functions (PSFs), detector footprint and motion blur, geometric warping from attitude dynamics, and photon-limited or thermal noise. The measured LR frames can be described generically by a forward model:

$y_k = D\,F_k\,K\,B_k\,x + n_k$

where $x$ is the unknown HR scene, $y_k$ the $k$ th observed LR image, $B_k$ the atmospheric PSF, $K$ the optical PSF, $F_k$ the frame- and attitude-dependent motion operator, $D$ the downsampling matrix, and $n_k$ additive noise (Garg et al., 2024).

Operational platforms introduce unique constraints:

Memory/compute: On-board DRAM or SRAM is limited (<1 GB); real-time operation typically precludes high-end GPUs (Tewari et al., 2021).
Power: Missions are solar-limited, driving SWaP-aware model and hardware choices.
Bandwidth: LR acquisition for wide swath coverage, with super-resolution (SR) reconstruction either on-board for selective transmission or as post-downlink processing.
Image content: Aliased high-frequency features, subpixel scene motions (orbital, attitude, payload), and instrument variability dictate SR potential and method robustness (Garg et al., 2024).

2. Classical and Deep-Learning-Based SR Reconstruction

Classical multi-frame SR leverages subpixel misregistration among multiple LR captures to synthesize a higher-fidelity scene estimate. Algorithms typically solve a variational or Bayesian inverse problem:

$\hat{x} = \arg\min_x \sum_{k=1}^K \| y_k - D F_k K B_k x \|_2^2 + \lambda R(x)$

where $x$ 0 enforces prior knowledge (e.g., total variation, Tikhonov, bilateral TV) (Garg et al., 2024, Garg et al., 2024). Solvers are usually iterative steepest descent or conjugate-gradient, with regularization weights and blur models tuned to the platform’s modulation transfer function (MTF), jitter, and noise levels (Garg et al., 2024).

Deep learning approaches replace manual regularizers with end-to-end-trained architectures, often using convolutional neural networks (CNNs), attention modules, or hybrid physics-informed networks. Examples include:

HighRes-net: Recursive fusion of multiple co-registered views, joint learning of fusion and up-sampling, and explicit registration loss via auxiliary networks (ShiftNet) (Deudon et al., 2020).
SOCM-3: Encoder-decoder with motion-estimation sub-networks for the EOS-06 OCM-3 sensor, trained under an $x$ 1 loss over real HR/LR patches (Garg et al., 2024).
RDNLA: Residual dense networks with non-local attention for patch-based on-orbit super-resolution under SWaP constraints, using overlap-reconstruction to mitigate artifact and context loss (Tewari et al., 2021).

When implemented for in-orbit processing, these models are pruned, quantized, and designed to fit within on-chip memory and power limits, enabling real-time operation on FPGAs or low-power accelerators (Tewari et al., 2021, Deudon et al., 2020).

3. Physics- and Instrument-Informed Super-Resolution

High-fidelity applications (solar, astrophysical, planetary) require that reconstructed images are consistent with underlying physical laws and instrument idiosyncrasies. Techniques include:

Physics-informed loss functions: Gradient penalties to preserve edge structure, divergence-free or energy-conservation constraints for field images (e.g., Maxwell’s equations in magnetograms), histogram-based distribution matching (Jungbluth et al., 2019).
Instrument translation models: Learning mappings between LR/HR products from distinct instruments (e.g., SOHO/MDI $x$ 2 SDO/HMI) to harmonize archives spanning decades and generations (Jungbluth et al., 2019).
Forward-adjoint modeling: Encoding knowledge of MTF, PSF, detector sampling, and high-frequency phase transfer for optimal inversion (Garg et al., 2024).

These enhancements suppress hallucinated detail, maintain physical plausibility, and ensure spectral/radiometric consistency needed for scientific or operational use.

4. Quantum and Information-Theoretic Approaches

Recent advances harness quantum estimation theory to break classical Rayleigh limits and achieve sub-diffraction super-resolution on orbit. There are two principal strategies:

Quantum super-resolution with balanced homodyne detection (BHD): By shaping local oscillators to spatial (e.g., Hermite–Gauss HG10) modes, BHD attains the quantum Fisher information (QFI) limit, yielding resolvable separations ( $x$ 3) orders of magnitude below classical PSF constraints, practical with current small-satellite hardware (Gosalia et al., 2023).
Multiparameter quantum estimation for passive imaging: For microwave-band remote sensing (e.g., SMOS), an optimal unitary mixing of antenna modes followed by photon counting, and maximum-likelihood estimation, enables HR reconstructions with pixel sizes %%%%14 $x$ 015%%%% smaller than the interferometric Rayleigh limit (Köse et al., 2023). This approach is robust to losses and feasible for current or near-term quantum hardware.

These paradigms have demonstrated, in simulation, resolution gains from $x$ 675 m (classical) to $x$ 71 m (quantum) (Gosalia et al., 2023) and 35 km (classical) to 3 km (quantum) (Köse et al., 2023).

5. In-Orbit System Architectures and Performance

On-orbit SR pipelines are often embedded within wider system architectures:

Selective transmission: On-board SR supports context-aware inference (e.g. crater detection, mineral mapping), flagging only pixels of scientific interest for downlink, massively reducing bandwidth (Tewari et al., 2021).
Patch-wise and overlap reconstruction: To address limited memory, SR networks operate on overlapping patches, with context-focused loss functions (mask-PSNR) to avoid artifacts (Tewari et al., 2021).
Subpixel shift calibration: Platforms like Resourcesat-3S/3SA leverage mechanical pixel staggering, TDI subarrays, and stereo payloads for deterministic, modelable LR-frame offsets, maximizing attainable SR gain (Garg et al., 2024).

Quantitative performance is usually reported in terms of PSNR, block-sensitive PSNR, BRISQUE scores, full-width-at-half-maximum (FWHM), super-resolution ratio (SR-ratio), and spectral fidelity metrics. Representative results:

Mission/Method	Resolution Gain	Notable Metrics/Notes
OCM-3/SOCM-3 (Garg et al., 2024)	2 $x$ 8 (360 $x$ 9180 m; 1080 $y_k$ 0540 m)	BRISQUE $y_k$ 1 by $y_k$ 240, FWHM ratio $y_k$ 31.7–1.8
Resourcesat-3S (Garg et al., 2024)	1.45 $y_k$ 4 (2.5 $y_k$ 51.7 m)	MTF and SNR sensitivity, 25,000-run Monte Carlo
Lunar SR/RDNLA (Tewari et al., 2021)	PSNR +0.1 dB, block-PSNR +0.19 dB	48% memory and 67% power savings vs. baseline RDN
Quantum BHD (Gosalia et al., 2023)	10–100 $y_k$ 6 beyond Rayleigh	Minimum $y_k$ 7 order 1/10 classical limit
Quantum Passive (Köse et al., 2023)	10 $y_k$ 8 beyond standard interferometry	3 km pix. at 758 km vs. 35 km classical limit

6. Applications and Representative Missions

On-orbit super-resolution is deployed across a spectrum of domains:

Earth Observation: EOS-06/OCM-3 (ocean color, vegetation, cryosphere), Resourcesat-3S/3SA (stereoscopic PAN imaging), lunar mapping (Kaguya DOMs), PROBA-V archive harmonization (Garg et al., 2024, Garg et al., 2024, Tewari et al., 2021, Deudon et al., 2020).
Astrophysics and Solar Physics: Up-sampling solar magnetograms for solar dynamo and space weather studies (Jungbluth et al., 2019).
Quantum Sensing and Navigation: Sub-meter inter-satellite ranging, quantum-enhanced interferometric Earth remote sensing (Gosalia et al., 2023, Köse et al., 2023).
Astronomical Imaging: Quantum measurement protocols for resolving stellar sources below the atmospheric or telescope-limited PSF, optimal transport-based SR for imaging with sparse/irregular arrays (Gosalia et al., 2023, Rawson et al., 2022).
Spectral and Environmental Monitoring: NDVI, chlorophyll, TSM, AOD retrievals, maintaining accuracy under SR-enhanced sampling (Garg et al., 2024).

7. Evaluation, Limitations, and Future Directions

Evaluation methodologies incorporate perceptual metrics (BRISQUE), edge/line sharpness (LSF, FWHM), spectral signature preservation, and operational product integrity (e.g., NDVI, ocean-color indices) (Garg et al., 2024). Robustness is assessed with power spectral density, entropy, and cross-instrument validation (Jungbluth et al., 2019).

Limitations include:

Sensitivity to registration: Subpixel shift accuracy, MTF, SNR, jitter, and PSF estimation errors are dominant limiting factors (Garg et al., 2024).
Failure cases: High noise, poorly characterized blur, or under-sampled strong features degrade SR fidelity (Garg et al., 2024, Jungbluth et al., 2019).
Physical priors/correctness: Inadequate or absent physical constraints may induce artifacts or hallucinated structure (Jungbluth et al., 2019, Garg et al., 2024).
Quantum hardware readiness: While lab and simulation results are promising, on-orbit quantum SR remains an active area for engineering translation (Köse et al., 2023, Gosalia et al., 2023).

A plausible implication is continued convergence of physics-aware deep learning, model-based inversion, and quantum information theory, enabling multi-modal, trustworthy SR for diverse orbital imaging platforms. Robust, lightweight architectures and hybrid cloud–on-orbit processing are expected to further enhance the scientific value of future missions.