Reframing the aerial–satellite domain gap
Identifies why conventional synthetic downsampling and cross-sensor satellite SR assumptions can fail when the high-resolution reference comes from a distinct imaging platform.
Cross-platform transfer01 · Research Gap
Why aerial-to-satellite super-resolution is different
High-resolution aerial images provide spatial detail that is difficult to obtain from KOMPSAT-3 alone, but they are not direct pixel-level ground truth for satellite observations. Differences in sensor response, acquisition time, viewing geometry, illumination, and spatial alignment make a simple aerial-to-satellite training transfer unreliable.
The research question
How can aerial imagery contribute to ×3 KOMPSAT-3 super-resolution without assuming that independently acquired aerial and satellite images are perfectly interchangeable?
Central challengeDomain gap
Realistic sensor degradation and imperfect cross-platform correspondence must be addressed separately.
02 · Key Contributions
Research contributions
SABridge investigates a degradation-aware approach to leveraging high-resolution aerial imagery for real KOMPSAT-3 super-resolution. Its contribution is not simply a stronger SR backbone, but the treatment of the cross-platform learning problem itself.
Target-aware degradation learning
Investigates how to create satellite-like low-resolution observations from aerial imagery while better representing target-sensor image characteristics than fixed downsampling assumptions.
Realistic degradationMisalignment-conscious SR and validation
Separates reliable synthetic correspondence from imperfect real-world alignment and evaluates degradation realism independently of the learning process, alongside preliminary real-image SR comparisons.
Reliable evaluation03 · Approach at a Glance
From aerial detail to real satellite restoration
The research combines target-aware degradation modeling with super-resolution learning and validation on real KOMPSAT-3 observations. Detailed architectures, parameterizations, optimization objectives, and training protocols are available to collaborators upon request.
Understand the target
Characterize how real satellite observations differ from high-resolution aerial images.
Bridge the image domains
Study realistic aerial-to-satellite degradation rather than relying only on idealized bicubic reduction.
Validate real-world utility
Evaluate degradation characteristics and preliminary reconstruction quality on KOMPSAT-3 imagery.
04 · Evaluation
Independent simulations and real KOMPSAT-3 validation
Evaluation is separated from the learned degradation process. A controlled benchmark uses physically motivated degradation mechanisms with exact LR–HR correspondence, while real-world validation assesses practical behavior on KOMPSAT-3 imagery.
Simulation domains
Controlled degradations
Anisotropic response · photon noise · detector non-uniformity · combined degradation
Evaluation metrics
PSNR · SSIM · LPIPS · DISTS
Full-reference evaluation of reconstruction fidelity and perceptual similarity
Real-world validation
KOMPSAT-3
Qualitative assessment with supplementary registered-pair evaluation
Super-resolution setting
×3 SR
64 × 64 LR → 192 × 192 HR
05 · Results
Degradation realism validation
The currently completed experiments evaluate whether SABridge-Deg reproduces target-like spectral characteristics across real aerial–KOMPSAT-3 observations and independent simulation domains.
Closer target-domain spectral behavior
The complete SABridge-Deg was observed to reduce radially averaged log spectral distance (RALSD) in the reported experiments relative to Bicubic, EGASR, and the intermediate DDM output. RefineCNN further reduces the residual discrepancy after explicit differentiable degradation.
Real aerial–KOMPSAT-3 RALSD6.06 dB
Final SABridge-Deg after RefineCNN; lower indicates closer radial spectral agreement.
Qualitative degradation comparison. Aerial image, Bicubic, EGASR, SABridge after DDM, SABridge after RefineCNN, and the corresponding real KOMPSAT-3 observation.
Degradation realism. RALSD and relative HF-Ratio deviation are evaluated for the real aerial–KOMPSAT-3 setting and four independent simulation domains. Lower RALSD and HF-Ratio deviations closer to zero indicate greater agreement with target LR imagery.
06 · Preliminary Results
Early qualitative results on real KOMPSAT-3 imagery
Super-resolution evaluation is still in progress. The example below provides an early qualitative view of how different degradation assumptions affect reconstruction under the aerial-to-satellite domain gap.
Preliminary qualitative comparison on real KOMPSAT-3 imagery.
The comparison highlights the influence of degradation modeling on structural recovery and high-frequency detail reconstruction.
01
Bicubic training leaves real degradations unresolved
DRCT (Bicubic) is trained without an explicit model of real satellite degradation. Consequently, degradations present in KOMPSAT-3 observations are not sufficiently compensated for, limiting the recovery of fine spatial structures.
02
Satellite cross-sensor models do not fully capture the aerial–satellite gap
KANet and EGASR are designed around cross-sensor satellite super-resolution. Their degradation assumptions do not fully represent the larger cross-platform discrepancy between aerial and satellite imagery, where differences in spatial response, resolution, acquisition geometry, and image formation are more pronounced.
03
Sharper reconstruction with stable pixel fidelity
SABridge-SR appears to recover sharper structures in the preliminary visual comparison; quantitative pixel-level fidelity remains under evaluation, suggesting the potential to recover more high-frequency structure.
Quantitative evaluation and additional real-world experiments are currently in progress. Final results will be updated after the remaining experiments are completed.
Additional Technical Details · Restricted
Detailed methodology (restricted access)
This work is currently in preparation. Enter the access password to view the protected methodology, experiments, figures, and results.