Pith. sign in

REVIEW 4 major objections 8 minor 95 references

A single model trained once can pansharpen images from any satellite by mapping every band count into one reversible latent space, and it reports beating task-specific fusers on a new global benchmark.

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 18:58 UTC pith:PRP3JDVJ

load-bearing objection Valuable dataset and a promising universal architecture, but the reversible latent projection is under-specified and the key guidance mechanism is not shown to help. the 4 major comments →

arxiv 2603.03831 v2 pith:PRP3JDVJ submitted 2026-03-04 cs.CV

Universal Pansharpening Model

Hebaixu Wang , Jing Zhang , Haonan Guo , Di Wang , Jiayi Ma , Bo Du , Liangpei Zhang This is my paper
classification cs.CV
keywords pansharpeningfoundation modelmulti-spectral image fusionlatent diffusion bridgemixture-of-expertsband-agnostic representationremote sensing benchmarkPSBench
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 paper tries to establish that pansharpening—fusing a high-resolution panchromatic image with a low-resolution multispectral image—can be solved by one universal model rather than a separate network per satellite. Its central move is to learn, per spectral band, a set of affine basis vectors that are concatenated into a mapping tensor T; multiplying the multispectral image by T sends any number of bands (4, 7, 8, or 10) into a fixed-size latent space, and the inverse T* brings it back. A latent diffusion bridge then refines the latent representation while a bridge-posterior-sampling term pulls it toward consistency with the observed pixel data. If the claims hold, a single trained model generalizes across heterogeneous sensors and scenes without fine-tuning, which would remove a major practical bottleneck in satellite imaging.

Core claim

FoundPS is the claim that a band-agnostic latent representation makes pansharpening universal. The modality-interleaved transformer (MiT) selects a top-B set of expert networks, each emitting a spectral affine basis; concatenating these bases forms a B×C mapping tensor T whose leading B×B block is set to the identity. Arbitrary-band MS images are projected to this latent space via tensor multiplication, refined by a latent diffusion bridge model with bridge posterior sampling that uses the approximate inverse T* to enforce consistency with the low-resolution MS observation, and finally projected back through T*. The paper reports that this pipeline outperforms all compared task-specific and

What carries the argument

The load-bearing object is the reversible mapping tensor T ∈ R^{B×C}, formed by concatenating band-wise spectral affine bases from a mixture-of-experts router, with its leading B×B block replaced by the identity to make projection well-posed. This tensor deterministically maps any B-band MS image into a fixed C-dimensional latent space (Z = x↑ ⊗ T) and, via its claimed inverse T*, maps back (ŷ = T* ⊗ ẑ0). The diffusion bridge (LDBM) and bridge posterior sampling (BPS) then evolve the latent toward a high-quality representation while using T and T* to couple the latent to pixel observations; the infinite-dimensional interaction block models all orders of cross-modal feature products through g

Load-bearing premise

The whole framework stands or falls on the claim that the learned mapping tensor T is invertible with a known left inverse T*; the paper asserts this by setting the leading B×B block to the identity, but never constructs or proves a true inverse for arbitrary B and fixed latent dimension C.

What would settle it

Take a trained FoundPS and feed it MS images with a band count it has seen (say 8 bands). Compute latent z = x↑ ⊗ T, reconstruct x̃ = T* ⊗ z, and measure the residual ‖x̃ − x↑‖. If the residual is not at the level of the training loss (or if spectral angles between bands change materially), the claimed reversible projection fails. A second, sharper test: for a fixed image, compare the round-trip error for B=4 vs B=10; if the error grows with B, the 'reversible' tensor is not maintaining spectral information for higher band counts.

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

If this is right

  • Training a single FoundPS model replaces the current practice of training and maintaining separate pansharpening networks for each satellite's spectral configuration.
  • The band-agnostic latent space gives a natural interface for downstream tasks (segmentation, NDVI/NDWI/NDRE/NDBI indices) that currently require band-specific pipelines.
  • Because BPS provides training-free guidance at inference, the same weights can adapt to unseen sensors and scenes by tuning one scalar weight η.
  • The PSBench benchmark (≈450K pairs, 17 land-cover classes, 5 satellite families) offers a common testbed that could standardize evaluation of universal pansharpening.
  • If the invertibility of T holds across band counts, the approach extends to other multi-modal fusion problems where input channels vary in number and order.

Where Pith is reading between the lines

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

  • The paper's own limitations section states three scope restrictions: a 1024×1024 patch ceiling, gradient-preserving BPS inference overhead, and a parameter scale that is moderate by foundation-model standards; these bound the reported universality claims.
  • The invertibility of T is asserted, not proven; a constructive check would be to verify T*T ≈ I for randomly sampled band counts and to measure spectral leakage in the round trip. If T is only approximately invertible, the 'band-agnostic latent' is lossy and the diffusion bridge operates on a corrupted encoding.
  • The same three-stage design (band-agnostic projection → latent diffusion → inverse projection) could be applied to hyperspectral pansharpening or to unifying pan-sharpening with super-resolution, where the number of input channels also varies.

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

4 major / 8 minor

Summary. The paper proposes FoundPS, a universal pansharpening foundation model intended to handle heterogeneous satellites and spectral configurations (4-, 7-, 8-, and 10-band MS) with a single trained model. Inputs are mapped into a fixed-dimension latent space by a modality-interleaved transformer that constructs a band-to-latent tensor T; a latent diffusion bridge model (LDBM) refines the representation under PAN guidance through an 'infinite-dimensional' interaction block, and bridge posterior sampling (BPS) is used for training-free adaptation. The authors also introduce PSBench, a large-scale multi-satellite benchmark with about 450K image pairs, and report extensive reduced- and full-scale experiments plus generalization and downstream-task evaluations, claiming consistent superiority over existing task-specific methods.

Significance. If the claims hold, the paper addresses a real bottleneck in pansharpening: existing learned methods are largely fixed to one spectral configuration and generalize poorly across sensors. The proposed band-agnostic projection plus a single universal model is a plausible route, and the reported margins over 24 baselines across four band configurations are large and internally consistent. The manuscript also has concrete strengths: a new large-scale benchmark, multiple model variants showing clean scaling behavior, generalization tests on unseen sensors/scenes, and downstream segmentation and spectral-index evaluations. However, the central methodology is currently under-specified at its load-bearing points: the invertibility of the mapping tensor T is asserted rather than proved, the inverse T* is never defined, and the key equations for diffusion guidance contain a dimensional mismatch. Because the universality claim rests on this latent round-trip, the contributions cannot be fully assessed without correcting these issues.

major comments (4)
  1. [§III-D, §III-F, Algorithms 1–2] The invertibility of T is asserted but not established. T ∈ R^{B×C}; setting the leading B×B block to the identity only guarantees a one-sided inverse if C ≥ B, a condition that is never stated. If C = B, T is the identity and the projection is vacuous; if C < B, no left inverse exists. The text never defines T* beyond calling it 'the inverse of T', yet Eqs. (19)–(20), Eq. (28), and both algorithms rely on exact round-tripping z0 = T y then T* z0 ≈ y. Please specify C, define T* explicitly (e.g., the Moore–Penrose right inverse), and state whether it is fixed or learned. Without this, the band-agnostic representation and the BPS guidance are not well-defined.
  2. [§III-F, Eq. (19)–(20)] Eq. (19) is dimensionally inconsistent as printed. zT is defined in Eq. (6) and Algorithm 1 as T⊗x↑, with shape H×W×C. The right-hand side T(T* z0 ↓ + n) has shape h×w×C because ↓ in Eq. (18) is the spatial degradation from H×W to h×w. Unless an upsampling operator is silently inserted, the BPS loss in Eq. (20) compares tensors of different spatial sizes. This affects the validity of the guidance gradient and of Algorithm 2. Please correct the equation (likely zT = T((T* z0 ↓)↑ + n)), or define zT in the low-resolution space and upsample in the latent domain.
  3. [§III-E, §III-F, Apps. A–B] The paper repeatedly refers to appendices for the derivation of the bridge SDE (Eq. (7)), the score/likelihood approximation (Eq. (17)), and the Jensen-gap bound (Eq. (21)), but no appendices are present in this version. Because these equations are the theoretical foundation of LDBM and BPS, and because Eq. (7) is taken from the authors' own preprint [68], the manuscript in its current form does not provide sufficient support for the central methodology. Please include the appendices or move the derivations into the main text.
  4. [§III-G, Eqs. (26)–(27)] The 'infinite-dimensional' interaction block uses the element-wise geometric series K_geo(W^T V) = 1/(1−W^T V) and the exponential series. For the geometric kernel, any entry of W^T V with |x| ≥ 1 makes the series diverge, so the operation is not well defined without a boundedness/normalization argument. The paper does not state how W^T V is bounded. If the kernels are intended as formal power series or are clipped in practice, this should be stated. As written, the infinitude of the interaction space is not rigorously established.
minor comments (8)
  1. [§III-F] 'Bayer's theorem' should be 'Bayes' theorem'.
  2. [§I] The phrase 'thereby enhancing generalization capability' is duplicated in the introduction.
  3. [Abstract / Title] The abstract refers to 'UniPS' while the rest of the paper uses 'FoundPS'; please make the naming consistent.
  4. [Algorithm 2, §III-F] Algorithm 2 uses ∇zt ||zT − T(T* z0 ↓)||² with z0, whereas Eq. (22) and the surrounding text use ẑ0. This is likely a typo, but it makes the algorithm ambiguous.
  5. [Table I] The GaoFen-7 row lists MS spatial resolution as 0.65 m and PAN as 2.6 m, which appears to be swapped relative to the usual specifications; please verify.
  6. [§III-D] Minor typo: 'arbitary' should be 'arbitrary'.
  7. [Fig. 4] The label 'Inifite-dimentional' should be 'Infinite-dimensional'.
  8. [§IV-B] It would be helpful to report standard deviations across multiple training runs for the main quantitative tables, since the reported margins are otherwise hard to judge for significance.

Circularity Check

1 steps flagged

Diffusion-bridge core is imported from same-author preprint; universal pansharpening empirical claim remains independently tested.

specific steps
  1. self citation load bearing [Sec. III-E (Eqs. 7-14) and Sec. IV-A.1 (hyperparameters)]
    "let zt be a finite random variable governed by the generalized diffusion bridge process in [68], with terminal condition zT . The evolution of its marginal distribution p(zt|zT ) satisfies the following stochastic differential equation under a fixed stationary coefficient ratio λ with noise schedule θt (detailed in Apps. A) ... We adopt the default noise schedule and stationary variance λ = 0.001 in [68] with total diffusion steps as 1000."

    The latent diffusion bridge model is presented as a central contribution, but its governing SDE (Eq. 7), analytical marginals (Eqs. 8-12), reverse sampling formulas (Eqs. 13-14), and the actual sampling hyperparameters are all imported from [68], the authors' own arXiv preprint. The manuscript defers the underlying derivations to 'Apps. A/B', which do not appear in the paper, so the generative dynamics that produce the fused output are not derived or independently supported here. This is load-bearing: without [68] the LDBM equations and hyperparameters have no in-paper grounding. The pansharpening comparisons are external, so the universality claim is not forced by this import, but the diffusion-bridge component itself rests on a same-author citation.

full rationale

The paper's central 'universal pansharpening' claim is validated on held-out test sets from PSBench and on unseen sensors (SegGF/GaoFen-2, Quickbird), so the performance claim is not a fitted constant renamed as a prediction. The MiT mapping, diffusion training, and BPS guidance are trained/executed end-to-end on paired data; no parameter is fit to the evaluation metrics and then reported as a prediction. The under-specification of the latent dimension C and the one-sided inverse T* in Eqs. (19)-(20) is a real correctness/reproducibility concern, but it is not circularity in the strict sense: no quantity in the derivation reduces by construction to an input it was supposed to predict. The only substantive circularity flag is the self-citation of [68] for the entire diffusion-bridge engine and its default schedule; this is a load-bearing same-author citation, but since the universal fusion result is independently benchmarked, the appropriate score is 4 rather than higher.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The paper introduces no physical entities; the main load-bearing assumptions are mathematical: exact invertibility of the learned band-mapping tensor, the correctness of the diffusion bridge equations drawn from unpublished self-cited work, and convergence of the 'infinite-dimensional' interaction kernels. The free parameters are mostly standard training hyperparameters, but the latent dimension and expert count are central design choices left unspecified.

free parameters (5)
  • Latent dimension C
    The fixed dimensionality of the unified latent space is a design choice, never specified numerically; all subsequent stages depend on it.
  • Number of experts L
    Mixture-of-experts count is not listed; the routing selects top-B experts, but total L is unspecified.
  • Diffusion steps = T=1000 training, 3 inference
    Chosen as operational default after ablation (Sec. IV-C2); performance varies with step count.
  • Auxiliary loss weight gamma = 0.001
    Set in Sec. IV-A to balance fusion quality and expert utilization.
  • Stationary variance lambda = 0.001
    Adopted from the authors' prior diffusion bridge work [68] without independent justification.
axioms (6)
  • ad hoc to paper T is invertible with a known left inverse T* after replacing the leading B×B block with identity
    Sec. III-D: 'we replace the leading B×B block of T with the identity matrix, ensuring well-posed projection.' No construction or proof of T* is given.
  • ad hoc to paper Approximation ∇zt log p(z_T|z_t) ≈ ∇zt log p(z_T|ẑ_t0)
    Eq. (17) in Sec. III-F; the derivation is relegated to the missing Appendix B.
  • standard math Diffusion bridge SDE (Eq. 7) and its analytical solution are correct as stated
    Taken directly from the authors' prior preprint [68]; no derivation is provided in the main text.
  • domain assumption Geometric kernel K_geo=1/(1-W^TV) converges as an infinite series
    Sec. III-G: for the geometric series to converge, all entries of W^TV must have absolute value <1; no normalization or proof is given.
  • domain assumption Wald's protocol provides valid reduced-scale supervision
    Used to create training/test pairs (Sec. III-B); standard in pansharpening but known to be an approximation to real degradations.
  • domain assumption Image data lie on a low-dimensional manifold (manifold assumption)
    Invoked in Sec. III-F to justify choosing z0-prediction mode over epsilon-prediction.

pith-pipeline@v1.3.0-alltime-deepseek · 37914 in / 11000 out tokens · 102689 ms · 2026-08-02T18:58:29.738767+00:00 · methodology

0 comments
read the original abstract

Pansharpening generates the high-resolution multi-spectral (MS) image by integrating spatial details from a texture-rich panchromatic (PAN) image and spectral attributes from a low-resolution MS image. Existing methods are predominantly satellite-specific and scene-dependent, which severely limits their generalization across heterogeneous sensors and varied scenes, thereby reducing their real-world practicality. To address these challenges, we present UniPS, a universal pansharpening model for satellite-agnostic and scene-robust fusion. Specifically, we introduce a modality-interleaved transformer that learns band-wise modal specializations to form reversible spectral affine bases, mapping arbitrary-band MS into a unified latent space via tensor multiplication. Building upon this, we construct a latent diffusion bridge model to progressively evolve latent representations, and incorporate bridge posterior sampling to couple latent diffusion with pixel-space observations, enabling stable and controllable fusion. Furthermore, we devise infinite-dimensional pixel-to-latent interaction mechanisms to comprehensively capture the cross-domain dependencies between PAN observations and MS representations, thereby facilitating complementary information fusion. In addition, to support large-scale training and evaluation, we construct a comprehensive pansharpening benchmark, termed PSBench, consisting of worldwide MS and PAN image pairs from multiple satellites across diverse scenes. Extensive experiments verify that UniPS consistently outperforms state-of-the-art methods, exhibiting superior generalization and robustness across a wide range of tasks.

Figures

Figures reproduced from arXiv: 2603.03831 by Bo Du, Di Wang, Haonan Guo, Hebaixu Wang, Jiayi Ma, Jing Zhang, Liangpei Zhang.

Figure 1
Figure 1. Figure 1: Pansharpening performance comparisons on PSBench across [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Mainstream pansharpening paradigms: (a) Satellite-specific methods employ independent encoder–decoder architectures per spectral [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Satellite observation locations and distribution of PSBench. It contains four representative spectral configurations for pansharpening [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The overall framework of the proposed FoundPS. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The network architecture of Infinite-UNet in latent diffusion bridge model. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: 8-band visual comparisons of reduced scale on PSBench. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 7
Figure 7. Figure 7: 7-band visual comparisons of reduced scale on PSBench. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 12
Figure 12. Figure 12: 8-band visual comparisons of full scale on PSBench. [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: 10-band visual comparisons of full scale on PSBench. [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Distribution of data and parameters by t-SNE visualization. (a) Expert parameters and their activation frequency. Larger circles [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Visual comparisons of reduced scale on SegGF. [PITH_FULL_IMAGE:figures/full_fig_p013_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visual comparisons of reduced scale on Quickbird. [PITH_FULL_IMAGE:figures/full_fig_p013_16.png] view at source ↗
Figure 19
Figure 19. Figure 19: Reduced-scale segmentation comparisons on SegGF. [PITH_FULL_IMAGE:figures/full_fig_p014_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Full-scale segmentation comparisons on SegGF . [PITH_FULL_IMAGE:figures/full_fig_p014_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Remote sensing application of NDVI. PAN MS GT PNN PanNet GPPNN U2Net PanFlowNet P2Sharpen DISPNet CANConv PAPS UniPAN WFANet CSLP DAISP FoundPS BAGDC MS SFIM IHS MTF_GLP PRACS AWLP GS C_BDSD MF BT-H BDSDPC 1 0 [PITH_FULL_IMAGE:figures/full_fig_p015_21.png] view at source ↗
Figure 24
Figure 24. Figure 24: Remote sensing application of NDBI. In accordance with these definitions, the normalized difference indices are confined to the range of [−1, 1], which can serve as the spectral accuracy indicators of the pansharpening process to a certain extent. The quantitative results over these four indices are presented in Tab. XII. Clearly, our method achieves the highest values across all these objective metrics, … 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

95 extracted references · 9 linked inside Pith

  1. [1]

    A semantic-enhanced multi-modal remote sensing foundation model for earth observation,

    K. Wu, Y . Zhang, L. Ru, B. Dang, J. Lao, L. Yu, J. Luo, Z. Zhu, Y . Sun, J. Zhanget al., “A semantic-enhanced multi-modal remote sensing foundation model for earth observation,”Nat. Mach. Intell., vol. 7, no. 8, pp. 1235–1249, 2025

    2025

  2. [2]

    Hipandas: Hyperspectral image joint denoising and super-resolution by image fusion with the panchromatic image,

    S. Xu, Z. Zhao, H. Bai, C. Yu, J. Peng, X. Cao, and D. Meng, “Hipandas: Hyperspectral image joint denoising and super-resolution by image fusion with the panchromatic image,” inProc. IEEE Int. Conf. Comput. Vis., 2025, pp. 12 002–12 011

    2025

  3. [3]

    Hyperspectral pansharpening: Critical review, tools, and future perspectives,

    M. Ciotola, G. Guarino, G. Vivone, G. Poggi, J. Chanussot, A. Plaza, and G. Scarpa, “Hyperspectral pansharpening: Critical review, tools, and future perspectives,”IEEE Geosci. Remote Sens. Mag., 2024

    2024

  4. [4]

    Toward resolution mismatching: Modality-aware feature-aligned network for pan-sharpening,

    M. Zhou, X. He, and D. Hong, “Toward resolution mismatching: Modality-aware feature-aligned network for pan-sharpening,”IEEE Trans. Pattern Anal. Mach. Intell., 2025

    2025

  5. [5]

    Deep adaptive unfolded network via spatial morphology stripping and spectral filtration for pan-sharpening,

    H. Wang and J. Ma, “Deep adaptive unfolded network via spatial morphology stripping and spectral filtration for pan-sharpening,” inProc. IEEE Int. Conf. Comput. Vision, 2025, pp. 10 730–10 740

    2025

  6. [6]

    Hyperspectral image super- resolution meets deep learning: A survey and perspective,

    X. Wang, Q. Hu, Y . Cheng, and J. Ma, “Hyperspectral image super- resolution meets deep learning: A survey and perspective,”Proc. Int. Joint Conf. Artif. Intell., vol. 10, no. 8, pp. 1664–1687, 2023

    2023

  7. [7]

    From classical image fusion to deep representation learning: A survey of the advances in hyperspectral image pan-sharpening,

    S. K. Roy, G. Vivone, S. Das, B. Raman, P. Singhet al., “From classical image fusion to deep representation learning: A survey of the advances in hyperspectral image pan-sharpening,”Inf. Fusion, p. 103834, 2025

    2025

  8. [8]

    Robust band-dependent spatial-detail approaches for panchromatic sharpening,

    G. Vivone, “Robust band-dependent spatial-detail approaches for panchromatic sharpening,”IEEE Trans. Geosci. Remote. Sens., vol. 57, no. 9, pp. 6421–6433, 2019

    2019

  9. [9]

    Spatial methods for multispectral pansharpening: Multiresolution analysis demystified,

    L. Alparone, S. Baronti, B. Aiazzi, and A. Garzelli, “Spatial methods for multispectral pansharpening: Multiresolution analysis demystified,” IEEE Trans. Geosci. Remote. Sens., vol. 54, no. 5, pp. 2563–2576, 2016

    2016

  10. [10]

    Better image filter for pansharp- ening,

    A. Guo, R. Dian, N. Wang, and S. Li, “Better image filter for pansharp- ening,”IEEE Trans. Image Process, vol. 34, pp. 8171–8184, 2025

    2025

  11. [11]

    Progress and challenges in intelligent remote sens. satellite systems,

    B. Zhang, Y . Wu, B. Zhao, J. Chanussot, D. Hong, J. Yao, and L. Gao, “Progress and challenges in intelligent remote sens. satellite systems,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 15, pp. 1814– 1822, 2022

    2022

  12. [12]

    A variational pan-sharpening with local gradient constraints,

    X. Fu, Z. Lin, Y . Huang, and X. Ding, “A variational pan-sharpening with local gradient constraints,” inProc. IEEE Conf. Comput. Vis. Pattern Recogn., 2019, pp. 10 265–10 274

    2019

  13. [13]

    An overview of gradient descent optimization algorithms,

    S. Ruder, “An overview of gradient descent optimization algorithms,” arXiv preprint arXiv:1609.04747, 2016

    Pith/arXiv arXiv 2016

  14. [14]

    Panchromatic and multispectral image fusion for remote sens. and earth observation: Concepts, taxonomy, literature review, evaluation methodologies and challenges ahead,

    K. Zhang, F. Zhang, W. Wan, H. Yu, J. Sun, J. Del Ser, E. Elyan, and A. Hussain, “Panchromatic and multispectral image fusion for remote sens. and earth observation: Concepts, taxonomy, literature review, evaluation methodologies and challenges ahead,”Inf. Fusion, vol. 93, pp. 227–242, 2023

    2023

  15. [15]

    P2sharpen: A progressive pansharpening network with deep spectral transformation,

    H. Zhang, H. Wang, X. Tian, and J. Ma, “P2sharpen: A progressive pansharpening network with deep spectral transformation,”Inf. Fusion, vol. 91, pp. 103–122, 2023

    2023

  16. [16]

    From forgotten to pan-sharpening,

    J. Wang, Y . Lin, C. Chen, X. Huang, R. Zhang, Y . Wang, and T. Lu, “From forgotten to pan-sharpening,”Pattern Recog., p. 112653, 2025

    2025

  17. [17]

    Upangan: Unsupervised pan- sharpening based on the spectral and spatial loss constrained generative adversarial network,

    Q. Xu, Y . Li, J. Nie, Q. Liu, and M. Guo, “Upangan: Unsupervised pan- sharpening based on the spectral and spatial loss constrained generative adversarial network,”Inf. Fusion, vol. 91, pp. 31–46, 2023

    2023

  18. [18]

    Mft-gan: A multiscale feature- guided transformer network for unsupervised hyperspectral pansharpen- ing,

    Y . Shang, J. Liu, J. Zhang, and Z. Wu, “Mft-gan: A multiscale feature- guided transformer network for unsupervised hyperspectral pansharpen- ing,”IEEE Trans. Geosci. Remote. Sens., vol. 62, pp. 1–16, 2024

    2024

  19. [19]

    Linearly-evolved transformer for pan-sharpening,

    J. Hou, Z. Cao, N. Zheng, X. Li, X. Chen, X. Liu, X. Cong, D. Hong, and M. Zhou, “Linearly-evolved transformer for pan-sharpening,” inProc. ACM Int. Conf. Multimed., 2024, pp. 1486–1494

    2024

  20. [20]

    Transformer-ensemble based implicit spectral-spatial functions for arbitrary-resolution hyperspectral pansharpening,

    L. He, H. Ye, J. Li, and A. Plaza, “Transformer-ensemble based implicit spectral-spatial functions for arbitrary-resolution hyperspectral pansharpening,”IEEE Trans. Geosci. Remote. Sens., 2025

    2025

  21. [21]

    Pan-mamba: Effective pan-sharpening with state space model,

    X. He, K. Cao, J. Zhang, K. Yan, Y . Wang, R. Li, C. Xie, D. Hong, and M. Zhou, “Pan-mamba: Effective pan-sharpening with state space model,”Inf. Fusion, vol. 115, p. 102779, 2025

    2025

  22. [22]

    Mambamtl: Progressive mutual- guided mamba multi-task learning for hyperspectral image pansharpen- ing and classification,

    S. Hou, S. Xiao, J. Qu, and W. Dong, “Mambamtl: Progressive mutual- guided mamba multi-task learning for hyperspectral image pansharpen- ing and classification,”IEEE Trans. Geosci. Remote. Sens., 2025

    2025

  23. [23]

    Zero-sharpen: A universal pansharpening method across satellites for reducing scale-variance gap via zero-shot variation,

    H. Wang, H. Zhang, X. Tian, and J. Ma, “Zero-sharpen: A universal pansharpening method across satellites for reducing scale-variance gap via zero-shot variation,”Inf. Fusion, vol. 101, p. 102003, 2024

    2024

  24. [24]

    Hyperspectral pansharpening via diffusion models with iteratively zero- shot guidance,

    J.-L. Xiao, T.-Z. Huang, L.-J. Deng, G. Lin, Z. Cao, C. Li, and Q. Zhao, “Hyperspectral pansharpening via diffusion models with iteratively zero- shot guidance,” inProc. IEEE Conf. Comput. Vis. Pattern Recogn., 2025, pp. 12 669–12 678

    2025

  25. [25]

    Zero-shot semi- supervised learning for pansharpening,

    Q. Cao, L.-J. Deng, W. Wang, J. Hou, and G. Vivone, “Zero-shot semi- supervised learning for pansharpening,”Inf. Fusion, vol. 101, p. 102001, 2024

    2024

  26. [26]

    Enpowering your pansharpening models with generalizability: Unified distribution is all you need,

    Y . Cui, P. Liu, and H. Zhang, “Enpowering your pansharpening models with generalizability: Unified distribution is all you need,” inProc. IEEE Int. Conf. Comput. Vision, 2025, pp. 11 850–11 860

    2025

  27. [27]

    Rethinking pan-sharpening: A new training process for full-resolution generalization,

    R. Zhang, X. He, L. Xueheng, K. Cao, L. Liu, W. Xu, F. Jiabin, Y . Qize, and J. Zhang, “Rethinking pan-sharpening: A new training process for full-resolution generalization,”arXiv preprint arXiv:2507.15059, 2025

    arXiv 2025

  28. [28]

    A survey on mixture of experts in large language models,

    W. Cai, J. Jiang, F. Wang, J. Tang, S. Kim, and J. Huang, “A survey on mixture of experts in large language models,”IEEE Trans. Knowl. Data Eng., 2025

    2025

  29. [29]

    Vitpose++: Vision transformer for generic body pose estimation,

    Y . Xu, J. Zhang, Q. Zhang, and D. Tao, “Vitpose++: Vision transformer for generic body pose estimation,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 2, pp. 1212–1230, 2023

    2023

  30. [30]

    Evomoe: An evolutional mixture-of-experts training frame- work via dense-to-sparse gate,

    X. Nie, X. Miao, S. Cao, L. Ma, Q. Liu, J. Xue, Y . Miao, Y . Liu, Z. Yang, and B. Cui, “Evomoe: An evolutional mixture-of-experts training frame- work via dense-to-sparse gate,”arXiv preprint arXiv:2112.14397, 2021

    Pith/arXiv arXiv 2021

  31. [31]

    Mixture of lora experts,

    X. Wu, S. Huang, and F. Wei, “Mixture of lora experts,” inInt. Conf. Learn. Represent

  32. [32]

    Dense training, sparse inference: Rethinking training of mixture-of-experts language models,

    B. Pan, Y . Shen, H. Liu, M. Mishra, G. Zhang, A. Oliva, C. Raffel, and R. Panda, “Dense training, sparse inference: Rethinking training of mixture-of-experts language models,”arXiv preprint arXiv:2404.05567, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Deepspeed-moe: Advancing mixture-of-experts inference and training to power next-generation ai scale,

    S. Rajbhandari, C. Li, Z. Yao, M. Zhang, R. Y . Aminabadi, A. A. Awan, J. Rasley, and Y . He, “Deepspeed-moe: Advancing mixture-of-experts inference and training to power next-generation ai scale,” inProc. Int. Conf. Mach. Learn.PMLR, 2022, pp. 18 332–18 346. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 17

    2022

  34. [34]

    Skywork-moe: A deep dive into training techniques for mixture-of-experts language models,

    T. Wei, B. Zhu, L. Zhao, C. Cheng, B. Li, W. L ¨u, P. Cheng, J. Zhang, X. Zhang, L. Zenget al., “Skywork-moe: A deep dive into training techniques for mixture-of-experts language models,”arXiv preprint arXiv:2406.06563, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Base layers: Simplifying training of large, sparse models,

    M. Lewis, S. Bhosale, T. Dettmers, N. Goyal, and L. Zettlemoyer, “Base layers: Simplifying training of large, sparse models,” inProc. Int. Conf. Mach. Learn.PMLR, 2021, pp. 6265–6274

    2021

  36. [36]

    Mixtral of experts,

    A. Q. Jiang, A. Sablayrolles, A. Roux, A. Mensch, B. Savary, C. Bam- ford, D. S. Chaplot, D. d. l. Casas, E. B. Hanna, F. Bressandet al., “Mixtral of experts,”arXiv preprint arXiv:2401.04088, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Jamba: Hybrid transformer-mamba language models,

    B. Lenz, O. Lieber, A. Arazi, A. Bergman, A. Manevich, B. Peleg, B. Aviram, C. Almagor, C. Fridman, D. Padnoset al., “Jamba: Hybrid transformer-mamba language models,” inInt. Conf. Learn. Represent., 2025

    2025

  38. [38]

    Yuan 2.0-m32: Mixture of experts with attention router,

    S. Wu, J. Luo, X. Chen, L. Li, X. Zhao, T. Yu, C. Wang, Y . Wang, F. Wang, W. Qiaoet al., “Yuan 2.0-m32: Mixture of experts with attention router,”arXiv preprint arXiv:2405.17976, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Pangu-{\Sigma}: Towards trillion parameter language model with sparse heterogeneous comput- ing,

    X. Ren, P. Zhou, X. Meng, X. Huang, Y . Wang, W. Wang, P. Li, X. Zhang, A. Podolskiy, G. Arshinovet al., “Pangu-{\Sigma}: Towards trillion parameter language model with sparse heterogeneous comput- ing,”arXiv preprint arXiv:2303.10845, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Hmoe: Heterogeneous mixture of experts for language modeling,

    A. Wang, X. Sun, R. Xie, S. Li, J. Zhu, Z. Yang, P. Zhao, W. Han, Z. Kang, D. Wanget al., “Hmoe: Heterogeneous mixture of experts for language modeling,” inProc. Conf. Empir. Methods Nat. Lang. Process, 2025, pp. 21 954–21 968

    2025

  41. [41]

    Uni-moe: Scaling unified multimodal llms with mixture of experts,

    Y . Li, S. Jiang, B. Hu, L. Wang, W. Zhong, W. Luo, L. Ma, and M. Zhang, “Uni-moe: Scaling unified multimodal llms with mixture of experts,”IEEE Trans. Pattern Anal. Mach. Intell., 2025

    2025

  42. [42]

    A new look at ihs-like image fusion methods,

    T.-M. Tu, S.-C. Su, H.-C. Shyu, and P. S. Huang, “A new look at ihs-like image fusion methods,”Inf. Fusion, vol. 2, no. 3, pp. 177–186, 2001

    2001

  43. [43]

    Haze correction for contrast-based multispectral pansharpening,

    S. Lolli, L. Alparone, A. Garzelli, and G. Vivone, “Haze correction for contrast-based multispectral pansharpening,”IEEE Geosci. Remote Sens. Lett., vol. 14, no. 12, pp. 2255–2259, 2017

    2017

  44. [44]

    Improving component substitution pansharpening through multivariate regression of ms+pan data,

    B. Aiazzi, S. Baronti, and M. Selva, “Improving component substitution pansharpening through multivariate regression of ms+pan data,”IEEE Trans. Geosci. Remote. Sens., vol. 45, no. 10, pp. 3230–3239, 2007

    2007

  45. [45]

    Introduction of sensor spectral response into image fusion methods. application to wavelet-based methods,

    X. Otazu, M. Gonz ´alez-Aud´ıcana, O. Fors, and J. N ´u˜nez, “Introduction of sensor spectral response into image fusion methods. application to wavelet-based methods,”IEEE Trans. Geosci. Remote. Sens., vol. 43, no. 10, pp. 2376–2385, 2005

    2005

  46. [46]

    Implementation of image fusion algorithm using matlab (laplacian pyramid),

    M. Pradeep, “Implementation of image fusion algorithm using matlab (laplacian pyramid),” inProc. Int. Multi-Conf. Autom. Comput. Commun. Control Compress. Sens.IEEE, 2013, pp. 165–168

    2013

  47. [47]

    A critical comparison among pansharpening algorithms,

    G. Vivone, L. Alparone, J. Chanussot, M. Dalla Mura, A. Garzelli, G. A. Licciardi, R. Restaino, and L. Wald, “A critical comparison among pansharpening algorithms,”IEEE Trans. Geosci. Remote. Sens., vol. 53, no. 5, pp. 2565–2586, 2014

    2014

  48. [48]

    A variational model for p+ xs image fusion,

    C. Ballester, V . Caselles, L. Igual, J. Verdera, and B. Roug ´e, “A variational model for p+ xs image fusion,”Int. J. Comput. Vis., vol. 69, no. 1, pp. 43–58, 2006

    2006

  49. [49]

    Sirf: Simultaneous satellite image registration and fusion in a unified framework,

    C. Chen, Y . Li, W. Liu, and J. Huang, “Sirf: Simultaneous satellite image registration and fusion in a unified framework,”IEEE Trans. Image Process, vol. 24, no. 11, pp. 4213–4224, 2015

    2015

  50. [50]

    A unified pansharpening model based on band-adaptive gradient and detail correction,

    H. Lu, Y . Yang, S. Huang, W. Tu, and W. Wan, “A unified pansharpening model based on band-adaptive gradient and detail correction,”IEEE Trans. Image Process, vol. 31, pp. 918–933, 2021

    2021

  51. [51]

    Pansharpening by convolutional neural networks,

    G. Masi, D. Cozzolino, L. Verdoliva, and G. Scarpa, “Pansharpening by convolutional neural networks,”Remote Sens., vol. 8, no. 7, p. 594, 2016

    2016

  52. [52]

    Pannet: A deep network architecture for pan-sharpening,

    J. Yang, X. Fu, Y . Hu, Y . Huang, X. Ding, and J. Paisley, “Pannet: A deep network architecture for pan-sharpening,” inProc. IEEE Int. Conf. Comput. Vision, 2017, pp. 5449–5457

    2017

  53. [53]

    Panformer: A transformer based model for pan-sharpening,

    H. Zhou, Q. Liu, and Y . Wang, “Panformer: A transformer based model for pan-sharpening,” inInt. Conf. Multimedia and Expo.IEEE, 2022, pp. 1–6

    2022

  54. [54]

    Pan-mamba: Effective pan-sharpening with state space model,

    X. He, K. Cao, J. Zhang, K. Yan, Y . Wang, R. Li, C. Xie, D. Hong, and M. Zhou, “Pan-mamba: Effective pan-sharpening with state space model,”Inf. Fusion, p. 102779, 2024

    2024

  55. [55]

    Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images,

    Z. Cao, S. Cao, L.-J. Deng, X. Wu, J. Hou, and G. Vivone, “Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images,”Inf. Fusion, vol. 104, p. 102158, 2024

    2024

  56. [56]

    A general spatial-frequency learning framework for multimodal image fusion,

    M. Zhou, J. Huang, K. Yan, D. Hong, X. Jia, J. Chanussot, and C. Li, “A general spatial-frequency learning framework for multimodal image fusion,”IEEE Trans. Pattern Anal. Mach. Intell., 2024

    2024

  57. [57]

    Deep spatial–spectral fusion transformer for remote sens. pansharpening,

    M. Ma, Y . Jiang, M. Zhao, X. Ma, W. Zhang, and S. Song, “Deep spatial–spectral fusion transformer for remote sens. pansharpening,”Inf. Fusion, vol. 118, p. 102980, 2025

    2025

  58. [58]

    Cslp: A novel pansharpening method based on compressed sensing and l-pnn,

    Y . Chen, Z. Wan, Z. Chen, and M. Wei, “Cslp: A novel pansharpening method based on compressed sensing and l-pnn,”Inf. Fusion, vol. 118, p. 103002, 2025

    2025

  59. [59]

    Enhanced pansharpening via quaternion spatial-spectral interactions,

    D. Li, C. Luo, Y . Bao, G. Yang, J. Xiao, X. Fu, and Z.-J. Zha, “Enhanced pansharpening via quaternion spatial-spectral interactions,” inProc. IEEE Conf. Comput. Vis. Pattern Recogn., 2025, pp. 10 908– 10 918

    2025

  60. [60]

    Pan-sharpening via multiscale dynamic convolutional neural network,

    J. Hu, P. Hu, X. Kang, H. Zhang, and S. Fan, “Pan-sharpening via multiscale dynamic convolutional neural network,”IEEE Trans. Geosci. Remote. Sens., vol. 59, no. 3, pp. 2231–2244, 2020

    2020

  61. [61]

    Memory- augmented deep conditional unfolding network for pan-sharpening,

    G. Yang, M. Zhou, K. Yan, A. Liu, X. Fu, and F. Wang, “Memory- augmented deep conditional unfolding network for pan-sharpening,” in Proc. IEEE Conf. Comput. Vis. Pattern Recogn., 2022, pp. 1788–1797

    2022

  62. [62]

    Pan-gan: An unsupervised pan-sharpening method for remote sens. image fusion,

    J. Ma, W. Yu, C. Chen, P. Liang, X. Guo, and J. Jiang, “Pan-gan: An unsupervised pan-sharpening method for remote sens. image fusion,” Inf. Fusion, vol. 62, pp. 110–120, 2020

    2020

  63. [63]

    Ssdiff: Spatial- spectral integrated diffusion model for remote sens. pansharpening,

    Y . Zhong, X. Wu, Z. Cao, H.-X. Dou, and L.-J. Deng, “Ssdiff: Spatial- spectral integrated diffusion model for remote sens. pansharpening,”Adv. Neural Inform. Process. Syst., vol. 37, pp. 77 962–77 986, 2024

    2024

  64. [64]

    Leveraging large-scale pretrained spatial- spectral priors for general zero-shot pansharpening,

    Y . Cui, P. Liu, and Y . Zeng, “Leveraging large-scale pretrained spatial- spectral priors for general zero-shot pansharpening,”arXiv preprint arXiv:2512.02643, 2025

    arXiv 2025

  65. [65]

    Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images,

    L. Wald, T. Ranchin, and M. Mangolini, “Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images,” Photogramm. Eng. Remote Sens., vol. 63, no. 6, pp. 691–699, 1997

    1997

  66. [66]

    A survey on vision transformer,

    K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, Y . Xuet al., “A survey on vision transformer,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 1, pp. 87–110, 2022

    2022

  67. [67]

    Bidirectional diffusion bridge models,

    D. Kieu, K. Do, T. Nguyen, D. Nguyen, and T. Nguyen, “Bidirectional diffusion bridge models,” inProc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., 2025, pp. 1139–1148

    2025

  68. [68]

    Residual diffusion bridge model for image restoration,

    H. Wang, J. Zhang, H. Chen, H. Guo, D. Wang, J. Ma, and B. Du, “Residual diffusion bridge model for image restoration,”arXiv preprint arXiv:2510.23116, 2025

    Pith/arXiv arXiv 2025

  69. [69]

    Tweedie’s formula and selection bias,

    B. Efron, “Tweedie’s formula and selection bias,”J. Am. Stat. Assoc., vol. 106, no. 496, pp. 1602–1614, 2011

    2011

  70. [70]

    Diffusion posterior sampling for general noisy inverse problems,

    H. Chung, J. Kim, M. T. McCann, M. L. Klasky, and J. C. Ye, “Diffusion posterior sampling for general noisy inverse problems,” inInt. Conf. Learn. Represent., 2023

    2023

  71. [71]

    Back to basics: Let denoising generative models denoise,

    T. Li and K. He, “Back to basics: Let denoising generative models denoise,”arXiv preprint arXiv:2511.13720, 2025

    Pith/arXiv arXiv 2025

  72. [72]

    Rewrite the stars,

    X. Ma, X. Dai, Y . Bai, Y . Wang, and Y . Fu, “Rewrite the stars,” inProc. IEEE Conf. Comput. Vis. Pattern Recogn., 2024, pp. 5694–5703

    2024

  73. [73]

    Smoothing filter-based intensity modulation: A spectral preserve image fusion technique for improving spatial details,

    J. Liu, “Smoothing filter-based intensity modulation: A spectral preserve image fusion technique for improving spatial details,”Int. J. Remote Sens., vol. 21, no. 18, pp. 3461–3472, 2000

    2000

  74. [74]

    Mtf- tailored multiscale fusion of high-resolution ms and pan imagery,

    B. Aiazzi, L. Alparone, S. Baronti, A. Garzelli, and M. Selva, “Mtf- tailored multiscale fusion of high-resolution ms and pan imagery,” Photogramm. Eng. Remote Sens., vol. 72, no. 5, pp. 591–596, 2006

    2006

  75. [75]

    A new adaptive component-substitution- based satellite image fusion by using partial replacement,

    J. Choi, K. Yu, and Y . Kim, “A new adaptive component-substitution- based satellite image fusion by using partial replacement,”IEEE Trans. Geosci. Remote. Sens., vol. 49, no. 1, pp. 295–309, 2010

    2010

  76. [76]

    Pansharpening of multispectral images based on nonlocal parameter optimization,

    A. Garzelli, “Pansharpening of multispectral images based on nonlocal parameter optimization,”IEEE Trans. Geosci. Remote. Sens., vol. 53, no. 4, pp. 2096–2107, 2014

    2096

  77. [77]

    Fusion of multispectral and panchromatic images based on morphological operators,

    R. Restaino, G. Vivone, M. Dalla Mura, and J. Chanussot, “Fusion of multispectral and panchromatic images based on morphological operators,”IEEE Trans. Image Process, vol. 25, no. 6, pp. 2882–2895, 2016

    2016

  78. [78]

    Deep gradient projection networks for pan-sharpening,

    S. Xu, J. Zhang, Z. Zhao, K. Sun, J. Liu, and C. Zhang, “Deep gradient projection networks for pan-sharpening,” inProc. IEEE Conf. Comput. Vis. Pattern Recogn., 2021, pp. 1366–1375

    2021

  79. [79]

    U2net: A general framework with spatial-spectral-integrated double u-net for image fusion,

    S. Peng, C. Guo, X. Wu, and L.-J. Deng, “U2net: A general framework with spatial-spectral-integrated double u-net for image fusion,” inProc. ACM Int. Conf. Multimed., 2023, pp. 3219–3227

    2023

  80. [80]

    Panflownet: A flow-based deep network for pan-sharpening,

    G. Yang, X. Cao, W. Xiao, M. Zhou, A. Liu, X. Chen, and D. Meng, “Panflownet: A flow-based deep network for pan-sharpening,” inProc. IEEE Conf. Comput. Vis. Pattern Recogn., 2023, pp. 16 857–16 867

    2023

Showing first 80 references.