Pith. sign in

REVIEW 3 major objections 3 minor 49 references

GSPan models pansharpening residuals as continuous 2D Gaussian primitives to enable rendering at any scale from one trained model.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

GSPan introduces continuous 2D Gaussian primitives for band-wise residuals in pansharpening to support arbitrary-scale fusion without retraining.

T0 review reviewed 2026-06-27 challenge →

load-bearing objection GSPan swaps fixed-grid prediction for 2D Gaussian primitives on pansharpening residuals to support arbitrary scales, but the stress-test worry about missing high-frequency content looks like the load-bearing assumption. the 3 major comments →

arxiv 2606.17722 v1 pith:4YKOSV4U submitted 2026-06-16 cs.CV

GSPan: A Continuous Gaussian Primitive Representation for Arbitrary-Scale Pansharpening

Fangyi Li , Xiaoyuan Yang , Yixiao Li , Zongyang Sui , Kangqing Shen , Gemine Vivone This is my paper
classification cs.CV
keywords pansharpeninggaussian splattingarbitrary scaleimage fusionmultispectral imagingremote sensingcontinuous representationresidual details
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes GSPan to replace fixed-grid pixel prediction in pansharpening with a continuous representation. Band-wise residual details are encoded as a set of learnable 2D Gaussian primitives whose parameters are inferred from paired panchromatic and multispectral inputs through a dual-stream network. These primitives are splatted into a residual field and added to an upsampled multispectral image. Because the representation is continuous, the same primitives can be rendered onto any target sampling grid without retraining or scale-specific models. The approach also supports a decoupled inference mode that estimates primitives at lower resolution and renders at the desired output resolution for faster processing of large scenes. Results on QuickBird, GaoFen-2, WorldView-3, and WorldView-3-4K datasets are reported as state-of-the-art.

Core claim

GSPan estimates a collection of 2D Gaussian primitives that encode band-wise residual details from PAN and MS inputs via a Dual-Stream Hierarchical Interaction architecture containing a Spatial-Spectral Interactive Attention module. The primitives are rendered by Gaussian splatting to produce a continuous residual detail field that is injected into the upsampled MS image. This continuous representation permits direct rendering of the fused multispectral image on arbitrary target grids and supports Scale-Decoupled Asymmetric Inference that reduces computation for large scenes while preserving fusion quality.

What carries the argument

2D Gaussian primitives that represent band-wise residual details, estimated by the DSHI network with SSIA module and rendered via Gaussian splatting to form the residual field.

Load-bearing premise

A finite collection of 2D Gaussian primitives estimated from the input observations can faithfully encode all necessary high-frequency residual information without introducing artifacts or losing detail when rendered at arbitrary scales.

What would settle it

Visible blurring, ringing, or loss of spatial detail when the same trained model renders outputs at scales differing substantially from the training resolution on the WorldView-3-4K dataset would falsify the arbitrary-scale claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single trained model can produce fused images at any desired output resolution without retraining.
  • Primitives can be estimated at reduced resolution and rendered at target resolution to accelerate inference on large scenes.
  • The method reports state-of-the-art quantitative and qualitative results on QuickBird, GaoFen-2, WorldView-3, and WorldView-3-4K benchmarks.
  • No separate scale-specific models or retraining steps are required for different output resolutions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The continuous primitive representation could extend to other remote-sensing fusion tasks that require flexible output resolutions.
  • Decoupling estimation from rendering may reduce memory demands when processing very large satellite scenes.
  • The same primitives might support on-demand multi-resolution outputs in interactive analysis tools without additional computation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes GSPan, which represents band-wise residual details for pansharpening as a set of continuous, learnable 2D Gaussian primitives (position, covariance, opacity, color per band) estimated from PAN and LRMS inputs by a Dual-Stream Hierarchical Interaction (DSHI) network containing a Spatial-Spectral Interactive Attention (SSIA) module. The primitives are rendered by splatting to produce a residual detail field that is added to the bicubically upsampled MS image; this continuous representation is claimed to support arbitrary-scale rendering without retraining and to enable Scale-Decoupled Asymmetric Inference (SDAI) that estimates primitives at reduced resolution before high-resolution rendering. Experiments on QuickBird, GaoFen-2, WorldView-3 and WorldView-3-4K report state-of-the-art quantitative results together with improved inference speed under SDAI.

Significance. A working continuous Gaussian-primitive representation would constitute a genuine departure from fixed-grid regression in pansharpening and could enable genuinely scale-agnostic fusion pipelines plus efficient large-scene processing. The SDAI decoupling of estimation and rendering resolution is a practical engineering contribution. Realization of these benefits, however, hinges on whether the finite set of learned Gaussians can faithfully encode high-frequency residuals across unseen target grids; the significance is therefore conditional on stronger empirical and analytical support for that claim.

major comments (3)
  1. [§3] §3 (Gaussian primitive estimation and rendering): the central claim that a finite collection of 2D Gaussians estimated at training resolution can encode all high-frequency residual content without loss or artifacts on arbitrary target grids is load-bearing yet unsupported by any frequency-domain analysis, anti-aliasing mechanism, or explicit covariance scaling rule; the rendering equation therefore risks introducing blurring or ringing precisely where the arbitrary-scale and SDAI claims are asserted.
  2. [Experiments] Experiments section (quantitative tables and ablations): no error bars, no explicit train/validation/test splits, and no ablation on the number of primitives or on scale factors outside the training distribution are reported; without these it is impossible to determine whether reported gains over baselines are attributable to the Gaussian representation or to other architectural choices.
  3. [§4] §4 (SDAI strategy): the claim that primitives estimated at reduced resolution can be rendered at full target resolution without quality degradation is asserted but not accompanied by a controlled study that isolates the effect of the resolution decoupling from the choice of primitive count or network capacity.
minor comments (3)
  1. [Abstract] Abstract: quantitative SOTA numbers and the exact baselines used are omitted, making the performance claim difficult to contextualize.
  2. [§3] Notation: the precise parameterization of each Gaussian primitive (mean, covariance matrix, opacity, per-band color) and the exact form of the splatting renderer should be stated with equations at the first appearance in §3.
  3. Figure captions: several figures lack scale bars or explicit indication of the target rendering resolution, complicating visual assessment of arbitrary-scale behavior.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We are grateful for the referee's insightful comments, which have helped us identify areas for improvement. We address each major comment below, indicating the revisions we plan to make to the manuscript.

read point-by-point responses

  1. Referee: [§3] §3 (Gaussian primitive estimation and rendering): the central claim that a finite collection of 2D Gaussians estimated at training resolution can encode all high-frequency residual content without loss or artifacts on arbitrary target grids is load-bearing yet unsupported by any frequency-domain analysis, anti-aliasing mechanism, or explicit covariance scaling rule; the rendering equation therefore risks introducing blurring or ringing precisely where the arbitrary-scale and SDAI claims are asserted.

    Authors: We thank the referee for highlighting this important aspect. While the paper emphasizes empirical results, we agree that additional analysis would strengthen the claims. In the revised version, we will add a section discussing the continuous nature of the Gaussian primitives and how the splatting process inherently supports arbitrary scales through the covariance parameters. We will also provide a brief analysis of potential artifacts and why they are mitigated in practice based on the learned opacities and positions. However, a full frequency-domain study may require further theoretical work beyond the scope of this revision. revision: partial

  2. Referee: [Experiments] Experiments section (quantitative tables and ablations): no error bars, no explicit train/validation/test splits, and no ablation on the number of primitives or on scale factors outside the training distribution are reported; without these it is impossible to determine whether reported gains over baselines are attributable to the Gaussian representation or to other architectural choices.

    Authors: We appreciate this feedback on experimental rigor. In the revised manuscript, we will include error bars computed over multiple training runs and explicitly state the train/validation/test splits for each dataset. We will also add an ablation study varying the number of Gaussian primitives to show its impact on performance. For scale factors outside the training distribution, this is a good point; our experiments include various scales but we will attempt to include additional results on unseen scales to demonstrate generalization. revision: partial

  3. Referee: [§4] §4 (SDAI strategy): the claim that primitives estimated at reduced resolution can be rendered at full target resolution without quality degradation is asserted but not accompanied by a controlled study that isolates the effect of the resolution decoupling from the choice of primitive count or network capacity.

    Authors: We agree that a more controlled study would better isolate the SDAI benefits. In the revision, we will add a controlled experiment where we fix the number of primitives and network capacity while varying the estimation resolution, comparing to full-resolution estimation, to demonstrate that the quality is maintained due to the continuous representation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper introduces a neural architecture (DSHI+SSIA) that learns parameters of 2D Gaussian primitives from PAN+MS inputs; these primitives are then rendered by standard splatting to produce a residual field added to an upsampled MS image. This is a standard learned representation whose parameters are optimized against reconstruction loss on training data, not defined in terms of the output metric or target scale. Arbitrary-scale rendering follows directly from the continuous splatting formulation rather than being presupposed. No self-citation chains, fitted-input-as-prediction, or ansatz smuggling appear in the provided description; empirical results on held-out datasets constitute independent evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities beyond the high-level claim that Gaussian primitives can represent residuals; ledger is therefore empty.

reviewed 2026-06-27 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GSPan: A Continuous Gaussian Primitive Representation for Arbitrary-Scale Pansharpening." pith.science (2026). https://pith.science/paper/4YKOSV4U

@misc{pith2026260617722,
  author       = {Pith},
  title        = {Pith review of: GSPan: A Continuous Gaussian Primitive Representation for Arbitrary-Scale Pansharpening},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4YKOSV4U}},
  note         = {Machine review of arXiv:2606.17722}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Pansharpening aims to generate high-resolution multispectral (HRMS) images by fusing low-resolution multispectral (LRMS) and panchromatic (PAN) observations. Most existing deep learning methods treat pansharpening as fixed-grid prediction, which limits scale adaptation. To address this, we propose GSPan, a framework that introduces 2D Gaussian Splatting (GS) into pansharpening. Instead of directly predicting pixels, GSPan represents band-wise residual details as continuous and learnable 2D Gaussian primitives. We design a Dual-Stream Hierarchical Interaction (DSHI) architecture with a Spatial-Spectral Interactive Attention (SSIA) module to estimate these primitives from complementary PAN and MS observations. The predicted primitives are rendered as a residual detail field and injected into the upsampled MS image. This continuous representation allows GSPan to render fused images on arbitrary target sampling grids without scale-specific retraining. It further enables a Scale-Decoupled Asymmetric Inference (SDAI) strategy, which estimates primitives at a reduced resolution and renders the fused image at the target resolution for efficient large-scene pansharpening. Experiments on QuickBird, GaoFen-2, WorldView-3, and WorldView-3-4K datasets show that GSPan delivers state-of-the-art fusion performance. Moreover, SDAI markedly accelerates inference, achieving a favorable trade-off between computational efficiency and fusion quality. Our results demonstrate the potential of continuous Gaussian residual representations as a flexible and scale-decoupled alternative to fixed-grid prediction.

Figures

Figures reproduced from arXiv: 2606.17722 by Fangyi Li, Gemine Vivone, Kangqing Shen, Xiaoyuan Yang, Yixiao Li, Zongyang Sui.

Figure 1
Figure 1. Figure 1: Conceptual comparison of different representation paradigms. (a) CNN Feature Map: Discrete grid-based repre￾sentation, illustrated by a CNN feature map as a representative example: features are organized on a regular pixel lattice, mak￾ing reconstruction dependent on predefined output grids. (b) INR Feature Map: representation is achieved via coordinate￾based queries on a latent grid. (c) GS Feature Map (O… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of the proposed GSPan framework. The pipeline consists of three primary stages: (A) Feature Encoder and Primitive Seed Embedding, which anchors spatial and spectral features into raw embeddings via window cross￾attention. (B) Dual-Stream Hierarchical Interaction (DSHI) Block, where Swin Transformer and Spatial-Spectral Interactive Attention (SSIA) modules facilitate intra-stream evolut… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison between the proposed GSPan and other methods on the QB dataset. Rows 1 and 3 show pansharpened fused images, and Rows 2 and 4 show corresponding absolute error maps against the ground truth (GT). AWLP TV BDSD PNN FusionNet LAGConv PSGAN MSAN PanMamba MARNet GSPan GT [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison between the proposed GSPan and other methods on the GF2 dataset. Rows 1 and 3 show pansharpened fused images, and Rows 2 and 4 show corresponding absolute error maps against the ground truth (GT). F. Li: Preprint submitted to Elsevier Page 8 of 16 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison between the proposed GSPan and other methods on the WV3 dataset. Rows 1 and 3 show pansharpened fused images, and Rows 2 and 4 show corresponding absolute error maps against the ground truth (GT) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison on the QB dataset under full-resolution testing. The top and third rows show the fused images, while the second and fourth rows display the HQNR maps. 0 0 0 0 0 0 𝑀𝑆 ̃/PAN AWLP TV BDSD PNN FusionNet LAGConv PSGAN MSAN PanMamba MARNet GSPan [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on the GF2 dataset under full-resolution testing. The top and third rows show the fused images, while the second and fourth rows display the HQNR maps. F. Li: Preprint submitted to Elsevier Page 10 of 16 [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on the WV3 dataset under full-resolution testing. The top and third rows show the fused images, while the second and fourth rows display the HQNR maps [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of Scale-Decoupled Asymmetric Inference (SDAI). Unlike standard full-resolution inference, which estimates Gaussian primitive attributes and renders the output on the same full-resolution grid, SDAI estimates primitive attributes at a reduced resolution and renders the fused image on the target high-resolution grid. This design reduces large-scene inference cost by decoupling primitive attribu… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison on the WV3-4K dataset. The first and third rows show fused images and the second and fourth rows show HQNR maps. Scale × 1.5 Scale × 2 Scale × 3 Scale × 4 FR fused result [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Arbitrary-scale pansharpening results on WV3-4K. Images from × 1.5 to × 4 are rendered from RR inputs by changing the target sampling grid, and are compared with the FR fused result obtained from original-scale inputs. Zoomed-in patches show that our method provides stable reconstruction quality and high-fidelity textures across different spatial resolutions. where 𝑠 (⋅) denotes the upsampling operation … view at source ↗
Figure 12
Figure 12. Figure 12: Visual analysis of the learned Gaussian primitives on the GaoFen-2 dataset. (a) Full-resolution fused image; (b) residual map representing the residual detail field; (c) Gaussian density map illustrating the spatial distribution of primitives, where each Gaussian is rendered as a uniform circular spot at its predicted position 𝜇 to reveal content-adaptive clustering; (d) Gaussian highlight map where red d… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Vivone, M

    G. Vivone, M. Dalla Mura, A. Garzelli, R. Restaino, G. Scarpa, M. O. Ulfarsson, L. Alparone, J. Chanussot, A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods, IEEEGeoscienceandRemoteSensingMagazine9(1)(2020)53–81

    2020

  2. [2]

    J. Li, D. Hong, L. Gao, J. Yao, K. Zheng, B. Zhang, J. Chanussot, Deep learning in multimodal remote sensing data fusion: A compre- hensive review, International Journal of Applied Earth Observation and Geoinformation 112 (2022) 102926

    2022

  3. [3]

    Zhang, H

    H. Zhang, H. Xu, X. Tian, J. Jiang, J. Ma, Image fusion meets deep learning: A survey and perspective, Information Fusion 76 (2021) 323–336

    2021

  4. [4]

    Vivone, L

    G. Vivone, L. Alparone, J. Chanussot, M. Dalla Mura, A. Garzelli, G. A. Licciardi, R. Restaino, L. Wald, A critical comparison among pansharpening algorithms, IEEE Transactions on Geoscience and Remote Sensing 53 (5) (2015) 2565–2586

    2015

  5. [5]

    G. Masi, D. Cozzolino, L. Verdoliva, G. Scarpa, Pansharpening by convolutional neural networks, Remote Sensing 8 (7) (2016) 594

    2016

  6. [6]

    J. Yang, X. Fu, Y. Hu, Y. Huang, X. Ding, J. Paisley, Pannet: A deep networkarchitectureforpan-sharpening,in:ProceedingsoftheIEEE international conference on computer vision, 2017, pp. 5449–5457

    2017

  7. [7]

    L.-J. Deng, G. Vivone, C. Jin, J. Chanussot, Detail injection-based deep convolutional neural networks for pansharpening, IEEE Trans- actionsonGeoscienceandRemoteSensing59(8)(2020)6995–7010

    2020

  8. [8]

    Q. Liu, H. Zhou, Q. Xu, X. Liu, Y. Wang, Psgan: A generative adversarial network for remote sensing image pan-sharpening, IEEE Transactions on Geoscience and Remote Sensing 59 (12) (2020) 10227–10242

    2020

  9. [9]

    J. Ma, W. Yu, C. Chen, P. Liang, X. Guo, J. Jiang, Pan-gan: An un- supervised pan-sharpening method for remote sensing image fusion, Information Fusion 62 (2020) 110–120

    2020

  10. [10]

    H.Zhou,Q.Liu,Y.Wang,Panformer:Atransformerbasedmodelfor pan-sharpening, in: 2022 IEEE international conference on multime- dia and expo (ICME), IEEE, 2022, pp. 1–6

    2022

  11. [11]

    H. Lu, Y. Yang, S. Huang, R. Liu, H. Guo, Msan: Multiscale self- attentionnetworkforpansharpening,PatternRecognition162(2025) 111441

    2025

  12. [12]

    Q. Meng, W. Shi, S. Li, L. Zhang, Pandiff: A novel pansharpen- ing method based on denoising diffusion probabilistic model, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–17

    2023

  13. [13]

    X. He, K. Cao, J. Zhang, K. Yan, Y. Wang, R. Li, C. Xie, D. Hong, M. Zhou, Pan-mamba: Effective pan-sharpening with state space model, Information Fusion 115 (2025) 102779

    2025

  14. [14]

    L. He, J. Zhu, J. Li, A. Plaza, J. Chanussot, Z. Yu, Cnn-based hyper- spectral pansharpening with arbitrary resolution, IEEE Transactions on Geoscience and Remote Sensing 60 (2022) 1–21

    2022

  15. [15]

    L. He, Z. Fang, J. Li, H. Ye, A. Plaza, Arbitrary-resolution hy- perspectral pansharpening neural operators, IEEE Transactions on Geoscience and Remote Sensing 63 (2025) 1–19

    2025

  16. [16]

    T. Wang, Z. Yan, J. Li, X. Zhao, C. Wang, M. Ng, Hyperspectral andmultispectralimagefusionwitharbitraryresolutionthroughself- supervisedrepresentations,InternationalJournalofComputerVision 133 (11) (2025) 7515–7535

    2025

  17. [17]

    Wang,Learningcontinuous imagerepresentation with local implicit image function, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp

    Y.Chen, S.Liu,X. Wang,Learningcontinuous imagerepresentation with local implicit image function, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 8628–8638

    2021

  18. [18]

    J. Cao, Q. Wang, Y. Xian, Y. Li, B. Ni, Z. Pi, K. Zhang, Y. Zhang, R. Timofte, L. Van Gool, Ciaosr: Continuous implicit attention- in-attention network for arbitrary-scale image super-resolution, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 1796–1807

    2023

  19. [19]

    J. Lee, K. H. Jin, Local texture estimator for implicit representation function, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 1929–1938

    2022

  20. [20]

    Liang, Z

    Y.-J. Liang, Z. Cao, S. Deng, H.-X. Dou, L.-J. Deng, Fourier- enhanced implicit neural fusion network for multispectral and hyper- spectral image fusion, Advances in Neural Information Processing Systems 37 (2024) 63441–63465

    2024

  21. [21]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, G. Drettakis, 3d gaussian splattingforreal-timeradiancefieldrendering,ACMTransactionson Graphics 42 (4) (2023)

    2023

  22. [22]

    G. Chen, W. Wang, A survey on 3d gaussian splatting, ACM Com- puting Surveys 58 (12) (2026)

    2026

  23. [23]

    D. Chen, L. Chen, Z. Zhang, L. Zhang, Generalized and efficient 2d gaussiansplattingforarbitrary-scalesuper-resolution,in:Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 26435–26445

    2025

  24. [24]

    Scarpa, S

    G. Scarpa, S. Vitale, D. Cozzolino, Target-adaptive cnn-based pan- sharpening, IEEE Transactions on Geoscience and Remote Sensing 56 (9) (2018) 5443–5457

    2018

  25. [25]

    X.Liu,Q.Liu,Y.Wang,Remotesensingimagefusionbasedontwo- stream fusion network, Information Fusion 55 (2020) 1–15

    2020

  26. [26]

    Jin, L.-J

    C. Jin, L.-J. Deng, T.-Z. Huang, G. Vivone, Laplacian pyramid net- works:Anewapproachformultispectralpansharpening,Information Fusion 78 (2022) 158–170. F. Li:Preprint submitted to Elsevier Page 15 of 16 GSPan for Arbitrary-Scale Pansharpening

    2022

  27. [27]

    I.Pereira-Sánchez,E.Sans,J.Navarro,J.Duran,Multi-headattention residual unfolded network for model-based pansharpening, Interna- tional Journal of Computer Vision 134 (2) (2026) 55

    2026

  28. [28]

    Y. Chen, Z. Wan, Z. Chen, M. Wei, Cslp: A novel pansharpening method based on compressed sensing and l-pnn, Information Fusion 118 (2025) 103002

    2025

  29. [29]

    Zhang, Z

    Y.Yan,Y.Wang,W.Tu,J.Wang,B.Cai,Q.Zhuang,X.Zuo,Y.Chen, H. Zhang, Z. Shao, S3mamba: Pan-sharpening via spatial–spectral synergistic state space model, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 19 (2026) 12820– 12834

    2026

  30. [30]

    X. Rui, X. Cao, L. Pang, Z. Zhu, Z. Yue, D. Meng, Unsupervised hy- perspectralpansharpeningvialow-rankdiffusionmodel,Information Fusion 107 (2024) 102325

    2024

  31. [31]

    Cao, L.-J

    Q. Cao, L.-J. Deng, W. Wang, J. Hou, G. Vivone, Zero-shot semi- supervised learning for pansharpening, Information Fusion 101 (2024) 102001

    2024

  32. [32]

    H. Wang, H. Zhang, X. Tian, J. Ma, Zero-sharpen: A universal pansharpening method across satellites for reducing scale-variance gap via zero-shot variation, Information Fusion 101 (2024) 102003

    2024

  33. [33]

    K. Shen, X. Yang, S. Lolli, G. Vivone, A continual learning-guided trainingframeworkforpansharpening,ISPRSJournalofPhotogram- metry and Remote Sensing 196 (2023) 45–57

    2023

  34. [34]

    K. Shen, X. Yang, Z. Li, J. Jiang, F. Jiang, H. Ren, Y. Li, Docsnet: a dual-outputandcross-scalestrategyforpan-sharpening,International Journal of Remote Sensing 43 (5) (2022) 1609–1629

    2022

  35. [35]

    Z. Yang, S. Yin, J. Liang, L.-J. Deng, G-zap: A generalizable zero- shot framework for arbitrary-scale pansharpening, arXiv preprint arXiv:2603.14412 (2026)

    work page internal anchor Pith review arXiv 2026

  36. [36]

    Zhang, X

    X. Zhang, X. Ge, T. Xu, D. He, Y. Wang, H. Qin, G. Lu, J. Geng, J. Zhang, Gaussianimage: 1000 fps image representation and com- pression by 2d gaussian splatting, in: European Conference on Com- puter Vision, Springer, 2024, pp. 327–345

    2024

  37. [37]

    L. Peng, A. Wu, W. Li, P. Xia, X. Dai, X. Zhang, X. Di, H. Sun, R. Pei, Y. Wang, et al., Pixel to gaussian: Ultra-fast continu- ous super-resolution with 2d gaussian modeling, arXiv preprint arXiv:2503.06617 (2025)

    work page Pith review arXiv 2025

  38. [38]

    J. Hu, B. Xia, B. Chen, W. Yang, L. Zhang, Gaussiansr: High fi- delity2dgaussiansplattingforarbitrary-scaleimagesuper-resolution, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, 2025, pp. 3554–3562

    2025

  39. [39]

    10012–10022

    Z.Liu,Y.Lin,Y.Cao,H.Hu,Y.Wei,Z.Zhang,S.Lin,B.Guo,Swin transformer: Hierarchical vision transformer using shifted windows, in: Proceedings of the IEEE/CVF international conference on com- puter vision, 2021, pp. 10012–10022

    2021

  40. [40]

    L.-J. Deng, G. Vivone, M.E. Paoletti, G. Scarpa, J. He, Y. Zhang, J. Chanussot, A. Plaza, Machine learning in pansharpening: A bench- mark, from shallow to deep networks, IEEE Geoscience and Remote Sensing Magazine 10 (3) (2022) 279–315

    2022

  41. [41]

    L. Wald, T. Ranchin, M. Mangolini, Fusion of satellite images of differentspatialresolutions:Assessingthequalityofresultingimages, Photogrammetricengineeringandremotesensing63(6)(1997)691– 699

    1997

  42. [42]

    L.Alparone,S.Baronti,A.Garzelli,F.Nencini,Aglobalqualitymea- surement of pan-sharpened multispectral imagery, IEEE Geoscience and Remote Sensing Letters 1 (4) (2004) 313–317

    2004

  43. [43]

    Alparone, L

    L. Alparone, L. Wald, J. Chanussot, C. Thomas, P. Gamba, L. M. Bruce, Comparison of pansharpening algorithms: Outcome of the 2006grs-sdata-fusioncontest,IEEETransactionsonGeoscienceand Remote Sensing 45 (10) (2007) 3012–3021

    2007

  44. [44]

    Garzelli, F

    A. Garzelli, F. Nencini, Hypercomplex quality assessment of multi/hyperspectral images, IEEE Geoscience and Remote Sensing Letters 6 (4) (2009) 662–665

    2009

  45. [45]

    Arienzo, G

    A. Arienzo, G. Vivone, A. Garzelli, L. Alparone, J. Chanussot, Full-resolutionqualityassessmentofpansharpening:Theoreticaland hands-on approaches, IEEE Geoscience and Remote Sensing Maga- zine 10 (3) (2022) 168–201

    2022

  46. [46]

    Garzelli, F

    A. Garzelli, F. Nencini, L. Capobianco, Optimal mmse pan sharpen- ing of very high resolution multispectral images, IEEE Transactions on Geoscience and Remote Sensing 46 (1) (2007) 228–236

    2007

  47. [47]

    Otazu, M

    X. Otazu, M. González-Audícana, O. Fors, J. Núñez, Introduction of sensor spectral response into image fusion methods. application to wavelet-based methods, IEEE Transactions on Geoscience and Remote Sensing 43 (10) (2005) 2376–2385

    2005

  48. [48]

    Palsson, J

    F. Palsson, J. R. Sveinsson, M. O. Ulfarsson, A new pansharpening algorithm based on total variation, IEEE Geoscience and Remote Sensing Letters 11 (1) (2013) 318–322

    2013

  49. [49]

    Jin, T.-J

    Z.-R. Jin, T.-J. Zhang, T.-X. Jiang, G. Vivone, L.-J. Deng, Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pansharpening, in: Proceedings of the AAAI conference on artificial intelligence, Vol. 36, 2022, pp. 1113–1121. F. Li:Preprint submitted to Elsevier Page 16 of 16

    2022

This paper was first reviewed by grok-4.3 on June 27, 2026.