Pith. sign in

REVIEW 2 major objections 4 minor 60 references

Unregistered Spectral Image Fusion: Unmixing, Adversarial Learning, and Recoverability

T0 review · 2 major / 4 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read Unregistered hyperspectral and multispectral images covering roughly the same area can be jointly super-resolved without training data or co-registration, with the first almost-sure recoverability guarantees for both resulting high-resoluti

desk verdict First recoverability theorems for unregistered HMF that jointly super-resolve both modalities; the HSR guarantees rest on a bijective generative model that does not match classical irreversible degradation. read the letter →

arxiv 2603.21510 v3 pith:XBEK4KHD submitted 2026-03-23 eess.IV cs.CV

Jiahui Song , Sagar Shrestha , Xiao Fu This is my paper · ORCID
classification eess.IVcs.CV
keywords unregisteredimagefusionhyperspectralimagingmultispectraltensorfactorizationadversariallearningrecoverabilityspectralunmixing
verification ladder T0 review T1 audit T2 compute T3 formal

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 shows that a pair of spatially unregistered hyperspectral and multispectral images can be fused unsupervised to raise the spectral resolution of the MSI and the spatial resolution of the HSI at once. The method first unmixed both images under a shared linear mixture model via coupled block-term tensor decomposition, recovering the high-spectral-resolution image over the MSI region. It then learns a single super-resolution map by adversarially matching the distributions of randomly rotated abundance patches so that low-resolution HSI patches are translated into the spatial resolution of the MSI. Under low-rank abundance maps, known spectral degradation, continuous bijective multimodal generators, and a diversity condition on material abundances, both super-resolved images are recoverable almost surely; a robust version bounds the error when diversity holds only approximately. A sympathetic reader cares because real acquisitions are rarely perfectly aligned, prior theory addressed only the co-registered case, and most existing unregistered methods either leave the HSI unimproved or require large labeled training sets that are often unavailable.

What carries the argument

FRESCO: coupled LL1 block-term tensor unmixing that recovers shared endmembers and MSI abundances, followed by a single shared translator trained by multi-discriminator adversarial matching of randomly rotated abundance patches under the Sufficiently Diverse Abundances condition that makes the translator unique.

What would settle it

Build a controlled pair in which the HSI is produced from a known high-resolution image by an irreversible blur-and-downsample operator that cannot be inverted, run the full pipeline, and measure whether the recovered HSI-region image still matches the ground-truth high-resolution image within the linear error bound of the robust theorem; a large systematic gap falsifies the generative-model premise.

Watch

Extended reading notes

Core claim

Under the linear mixture model with shared endmembers, low-rank abundance maps, known full-row-rank spectral degradation, and a multimodal patch generative model with continuous bijective maps plus sufficiently diverse abundances, coupled block-term tensor decomposition recovers the super-resolution image over the MSI region almost surely, and multi-material adversarial distribution matching recovers the exact super-resolution map for the HSI region almost surely, with reconstruction error scaling linearly in the diameter of diversity violations when the diversity condition holds only approximately.

Load-bearing premise

Every matching pair of HSI and MSI abundance patches must be produced by continuous invertible maps from a shared low-dimensional latent code; if the true spatial degradation permanently erases information, no exact super-resolution function exists and the recovery theorems fail.

Editorial extensions

If this is right

  • Both the MSI-covered and HSI-covered regions can be super-resolved from one unregistered pair without external training data or explicit co-registration.
  • Recoverability of the MSI-region super-resolution image continues to hold when the images are not co-registered, extending earlier co-registered tensor results.
  • Exact recovery of the HSI super-resolution map is guaranteed when material abundance patches are sufficiently diverse, not merely when a GAN loss converges.
  • Approximate diversity still yields a reconstruction error that scales linearly with the diameter of the violating sets.
  • The pipeline remains usable when the spectral degradation matrix is estimated by simple mean-and-variance matching rather than known a priori.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The generative model replaces classical irreversible blur-and-downsample with a content-preserving map; when real sensors lose high frequencies permanently, the HSR guarantees become approximate, which the robust theorem only partially covers.
  • The same diversified multi-material distribution-matching idea could transfer to other unregistered multi-modal fusion settings (RGB-depth, multi-sensor medical imaging) whenever a shared low-dimensional spatial content is plausible.
  • Strong empirical performance under large angular misalignment and partial overlap suggests that separate co-registration stages may be unnecessary when abundance diversity is high.
  • Sliding-window inference with overlap averaging is required in practice yet lies outside the patch-wise theory; boundary consistency under non-bijective real degradations remains an open empirical question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes FRESCO, an unsupervised framework for unregistered hyperspectral-multispectral fusion that jointly recovers super-resolution images over both the MSI region (via coupled block-term tensor decomposition spectral unmixing under the linear mixture model) and the HSI region (via multi-abundance latent-space adversarial distribution matching). Under low-rank abundance maps, known full-row-rank spectral degradation, and a multimodal patch generative model with continuous bijective generators plus a sufficiently diverse abundances condition, Theorem 1 establishes almost-sure recoverability of the MSI-region SRI, while Theorems 2–3 establish exact (or robustly bounded) recoverability of the super-resolution map and thus the HSI-region SRI. The method is validated on semi-real cubes under three downsampling operators and increasing misalignment, plus a real Hyperion–Sentinel-2 pair, with ablations and comparisons to registration-plus-fusion and supervised HSR baselines.

Significance. If the claims hold, this supplies the first recoverability guarantees for unregistered HMF that target both SRIs simultaneously, without co-registration or external training data. That is a genuine advance over co-registered tensor/matrix analyses and over purely empirical unregistered pipelines that recover only the MSI-region SRI. The proofs (LL1 uniqueness plus diversified distribution matching) are machine-checkable in the supplement, the experimental suite is broad (three cubes, three degradations, misalignment sweeps, real data, ablations, estimated vs. ground-truth P^(M)), and code is promised; these are concrete strengths that raise the work above typical method-only remote-sensing papers.

major comments (2)
  1. [IV-B, Remark 3, Theorems 2–3, VI] Section IV-B (Assumption 4, Definitions 2–3, Lemma 1), Remark 3, and Theorems 2–3: the HSR recoverability statements rest on continuous bijective generators g^(H), g^(M) from a shared latent z_r, so that a continuous bijection f* exists. Classical spatial degradation (blur + downsample) is irreversible and therefore admits no such f*. The semi-real experiments in Section VI (Gaussian, nearest-neighbor, and uniform operators in Tables I–II and Figs. 7–12) employ precisely those irreversible operators. Consequently Theorems 2–3 do not apply to the data-generating process used for validation; Theorem 3 only robustifies SDA violations, not failures of bijectivity. The manuscript already notes the modeling difference in Remark 3, but the theory–practice gap on the paper’s strongest claim (first recoverability guarantees for the HSI-region SRI) still needs explicit discussion, either by additi
  2. [V-C, Tables III–IV] Section V-C and Tables III–IV: the heuristic (29) for estimating P^(M) when sensor specifications are incomplete is presented as a practical remedy, yet the recoverability statements of Theorems 1–3 assume either known full-row-rank P^(M) or perfect material-wise correspondence after unmixing. The performance drop under estimated P^(M) is small but non-zero; a short analysis of how residual permutation or scaling errors propagate into the subsequent adversarial stage would strengthen the claim that the pipeline remains recoverable when P^(M) is only approximately known.
minor comments (4)
  1. [Fig. 10] Fig. 10 caption incorrectly labels the recovered HSI-region SRI as bY^(M)_SRI; the same slip appears in the surrounding text. Correct to bY^(H)_SRI for consistency with the notation of (6) and (21).
  2. [VI-A4] In the hyper-parameter selection paragraph of VI-A4 the phrase “computes the PSNR using the estimated and observed HSI and MSI data” is slightly ambiguous; a one-sentence clarification that no ground-truth SRI is used would help readers reproduce the grid search.
  3. [II-B, IV-B] Notation for the continuous global coordinate w and the directional patch boundaries (Definitions 1–3) is introduced late; a forward reference in Section II-B would improve readability.
  4. [throughout] A few typographical inconsistencies remain (e.g., “discreminators”, “invertiability”, occasional missing spaces around math mode). A careful proof-reading pass will remove them.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor load-bearing self-citation of diversified-matching uniqueness from overlapping authors; core recoverability claims are not definitional tautologies or fitted predictions.

  1. uniqueness imported from authors [Theorem 2 proof (Appendix C, Steps 2–3); also Theorem 3 / Appendix D citing Lemma A.3 of [44]]
    "The key idea of the proof is to link the patch generative model to distribution transfer analysis in [44]… we follow the steps in [44, Theorem 1], with slight modifications… Since the proof follows exactly the same steps as those in Lemma A.3 of [44] under the Lipschitz continuity condition (63) and the relaxed diversity condition (61)…"

    HSR recoverability (bf = f* almost surely, and the robust error bound) is forced by importing the diversified-matching uniqueness/robustness theorems of Shrestha & Fu [44], whose author set overlaps the present paper. The paper treats that uniqueness as an external mathematical fact that rules out non-identity continuous maps h preserving all R push-forwards. This is load-bearing for the central HSR claim, but only mildly circular: [44] is a published general result applied to a new generative model, not an unverified self-lemma, and the patch model / angle randomization / MSR stage remain independent content.

full rationale

The paper’s two recoverability pillars are not circular in the strong sense. Theorem 1 applies classical LL1 uniqueness (De Lathauwer) to coupled unregistered BTD; the extension beyond the co-registered case in Ding–Fu et al. [5] is real (different abundance ranks, shared permutation via P^(M)). Lemma 1 constructs f* = g^(M)∘(g^(H))^{-1} under the explicit bijectivity assumption—honestly definitional given Assumption 4, not a hidden tautology sold as free prediction. Theorems 2–3 import the diversified distribution-matching uniqueness and robustness lemmas from Shrestha–Fu [44] (same two co-authors). That is a load-bearing self-citation of a uniqueness theorem, but [44] is a peer-reviewed ICLR result with its own stated assumptions, not an unverified private lemma, and the present paper’s contribution is the patch generative model, angle randomization, and application to unregistered HMF. Hyper-parameters are tuned on observed (not ground-truth SRI) reconstruction and do not enter the theorems. No fitted parameter is renamed a prediction; no equation reduces to its own input by construction. Score 2 reflects the single non-critical uniqueness import from overlapping authors; the theory–practice gap on bijectivity vs. irreversible blur (Assumption 4) is a correctness/applicability issue, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The central recoverability claims rest on classical LMM physics, low-rank abundance smoothness, a known or moment-estimated spectral degradation matrix, and a newly postulated continuous bijective multimodal patch generative model plus a diversity condition on the latent abundances. Free parameters are the usual regularization weights, ranks, and network hyper-parameters; none of them enter the statements of Theorems 1–3.

free parameters (4)
  • λ_LR, λ_sto, λ_TV
    Regularization weights for low-rank, sum-to-one and total-variation terms in the MSR stage; selected by grid search maximizing PSNR on the observed (not ground-truth) HSI/MSI.
  • λ_inv, λ_scale
    Weights for invertibility and intensity-scale regularizers in the adversarial HSR stage; fixed at 10 and 15 after informal tuning.
  • R (number of materials) and L^(H), L^(M) (rank bounds)
    Chosen by the user; experiments use R=3–5 and implicit low ranks induced by the Schatten-p penalty.
  • Network architectures and learning-rate schedule for f, g, d_r
    U-Net + residual down-sampler + multi-task discriminator; learning rate 1e-4 with linear decay after 20 k iterations; not derived from theory.
assumptions (5)
  • domain assumption Linear mixture model with shared non-negative endmembers and sum-to-one abundances across the roughly overlapping HSI and MSI regions (Assumptions 1–2).
    Standard in hyperspectral unmixing; invoked from Section II-D onward and essential for both stages.
  • domain assumption Abundance maps of each material are low-rank matrices (Assumption 3).
    Used for LL1 uniqueness in Theorem 1; justified by spatial smoothness but not always true for sharp edges.
  • ad hoc to paper Multimodal patch generative model: patches are continuous bijective images of a shared latent z_r under random centers and rotations (Assumption 4).
    Introduced in Section IV-B to guarantee existence of f*; replaces classical blur-and-downsample degradation.
  • ad hoc to paper Sufficiently Diverse Abundances (SDA) or its η-relaxed version (Assumptions 5–6).
    Required for uniqueness of the common translator f in Theorems 2–3; may fail for nearly identical materials.
  • standard math LL1 block-term decomposition uniqueness under continuous factor distributions and the stated dimension inequalities (Lemma 2, from De Lathauwer).
    External uniqueness theorem used in the proof of Theorem 1.
invented entities (2)
  • Multimodal patch generative model with angle randomization (Assumption 4 + Definitions 2–3)
    purpose: Supplies a content-preserving, bijective relationship between HSI and MSI abundance patches that classical kernel degradation lacks, enabling recoverability of f*.
    Postulated for the paper; no independent empirical validation outside the fusion experiments.
  • FRESCO pipeline (coupled BTD + multi-abundance adversarial translator)
    purpose: Concrete unsupervised algorithm realizing the two-stage recoverability theory.
    New algorithmic combination; evaluated only on the paper’s own semi-real and real pairs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unregistered Spectral Image Fusion: Unmixing, Adversarial Learning, and Recoverability." pith.science (2026). https://pith.science/paper/XBEK4KHD

@misc{pith2026260321510,
  author       = {Pith},
  title        = {Pith review of: Unregistered Spectral Image Fusion: Unmixing, Adversarial Learning, and Recoverability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XBEK4KHD}},
  note         = {Machine review of arXiv:2603.21510}
}
read the original abstract

This paper addresses the fusion of a pair of spatially unregistered hyperspectral image (HSI) and multispectral image (MSI) covering roughly overlapping regions. HSIs offer high spectral but low spatial resolution, while MSIs provide the opposite. The goal is to integrate their complementary information to enhance both HSI spatial resolution and MSI spectral resolution. While hyperspectral-multispectral fusion (HMF) has been widely studied, the unregistered setting remains challenging. Many existing methods focus solely on MSI super-resolution, leaving HSI unchanged. Supervised deep learning approaches were proposed for HSI super-resolution, but rely on accurate training data, which is often unavailable. Moreover, theoretical analyses largely address the co-registered case, leaving unregistered HMF poorly understood. In this work, an unsupervised framework is proposed to simultaneously super-resolve both MSI and HSI. The method integrates coupled spectral unmixing for MSI super-resolution with latent-space adversarial learning for HSI super-resolution. Theoretical guarantees on the recoverability of the super-resolution MSI and HSI are established under reasonable generative models -- providing, to our best knowledge, the first such insights for unregistered HMF. The approach is validated on semi-real and real HSI-MSI pairs across diverse conditions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 1 linked inside Pith

  1. [1]

    Hyperspectral and Mul- tispectral Data Fusion: A comparative review of the recent literature,

    N. Yokoya, C. Grohnfeldt, and J. Chanussot, “Hyperspectral and Mul- tispectral Data Fusion: A comparative review of the recent literature,” IEEE Geosci. Remote Sens. Mag., vol. 5, no. 2, pp. 29–56, 2017

    2017

  2. [2]

    The use of intensity-hue- saturation transformations for merging SPOT panchromatic and mul- tispectral image data,

    W. Carper, T. Lillesand, and R. Kiefer, “The use of intensity-hue- saturation transformations for merging SPOT panchromatic and mul- tispectral image data,”Photogramm. Eng. Remote Sens., vol. 56, no. 4, pp. 459–467, 1990

    1990

  3. [3]

    An MTF-based spectral distortion minimizing model for pan-sharpening of very high resolution multispectral images of urban areas,

    B. Aiazzi, L. Alparone, S. Baronti, A. Garzelli, and M. Selva, “An MTF-based spectral distortion minimizing model for pan-sharpening of very high resolution multispectral images of urban areas,” inProc. 2nd IEEE/ISPRS Joint Workshop Remote Sens. Data Fusion Urban Areas, pp. 90–94, 2003

    2003

  4. [4]

    Hyperspec- tral super-resolution: A coupled tensor factorization approach,

    C. I. Kanatsoulis, X. Fu, N. D. Sidiropoulos, and W.-K. Ma, “Hyperspec- tral super-resolution: A coupled tensor factorization approach,”IEEE Trans. Signal Process., vol. 66, no. 24, pp. 6503–6517, 2018

    2018

  5. [5]

    Hyperspectral super-resolution via interpretable block-term tensor modeling,

    M. Ding, X. Fu, T.-Z. Huang, J. Wang, and X.-L. Zhao, “Hyperspectral super-resolution via interpretable block-term tensor modeling,”IEEE J. Sel. Topics Signal Process., vol. 15, no. 3, pp. 641–656, 2020

    2020

  6. [6]

    Coupled nonnegative matrix factorization unmixing for hyperspectral and multispectral data fusion,

    N. Yokoya, T. Yairi, and A. Iwasaki, “Coupled nonnegative matrix factorization unmixing for hyperspectral and multispectral data fusion,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 2, pp. 528–537, 2012

    2012

  7. [7]

    Hyperspec- tral and multispectral image fusion based on a sparse representation,

    Q. Wei, J. Bioucas-Dias, N. Dobigeon, and J.-Y . Tourneret, “Hyperspec- tral and multispectral image fusion based on a sparse representation,” IEEE Trans. Geosci. Remote Sens., vol. 53, no. 7, pp. 3658–3668, 2015

    2015

  8. [8]

    A convex formulation for hyperspectral image superresolution via subspace-based regularization,

    M. Sim ˜oes, J. Bioucas-Dias, L. B. Almeida, and J. Chanussot, “A convex formulation for hyperspectral image superresolution via subspace-based regularization,”IEEE Trans. Geosci. Remote Sens., vol. 53, no. 6, pp. 3373–3388, 2015

    2015

Show all 60 references
  1. [9]

    Deep blind hyperspectral image super-resolution,

    L. Zhang, J. Nie, W. Wei, Y . Li, and Y . Zhang, “Deep blind hyperspectral image super-resolution,”IEEE Trans. Neural Netw. Learn. Syst., vol. 32, no. 6, pp. 2388–2400, 2020

    2020

  2. [10]

    Unsupervised deep hyperspectral super-resolution with unregistered images,

    J. Nie, L. Zhang, W. Wei, C. Ding, and Y . Zhang, “Unsupervised deep hyperspectral super-resolution with unregistered images,” inProc. IEEE Int. Conf. Multimedia Expo (ICME), pp. 1–6, IEEE, 2020

    2020

  3. [11]

    NonRegSR- Net: A nonrigid registration hyperspectral super-resolution network,

    K. Zheng, L. Gao, D. Hong, B. Zhang, and J. Chanussot, “NonRegSR- Net: A nonrigid registration hyperspectral super-resolution network,” IEEE Trans. Geosci. and Remote Sens., vol. 60, pp. 1–16, 2021

    2021

  4. [12]

    Unsupervised and unregistered hyperspectral image super-resolution with mutual Dirichlet- Net,

    Y . Qu, H. Qi, C. Kwan, N. Yokoya, and J. Chanussot, “Unsupervised and unregistered hyperspectral image super-resolution with mutual Dirichlet- Net,”IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–18, 2021

    2021

  5. [13]

    Is there any recovery guarantee with coupled structured matrix factorization for hyperspectral super- resolution?,

    H. Liu, R. Wu, and W.-K. Ma, “Is there any recovery guarantee with coupled structured matrix factorization for hyperspectral super- resolution?,” inProc. IEEE Int. Workshop Comput. Adv. Multi-Sensor Adapt. Process. (CAMSAP), pp. 480–484, IEEE, 2019

    2019

  6. [14]

    Hyperspectral super- resolution with coupled Tucker approximation: Recoverability and SVD- based algorithms,

    C. Pr ´evost, K. Usevich, P. Comon, and D. Brie, “Hyperspectral super- resolution with coupled Tucker approximation: Recoverability and SVD- based algorithms,”IEEE Trans. Signal Process., vol. 68, pp. 931–946, 2020

    2020

  7. [15]

    Hyperspectral super-resolution: Exact recovery in polynomial time,

    Q. Li, W.-K. Ma, and Q. Wu, “Hyperspectral super-resolution: Exact recovery in polynomial time,” inProc. IEEE Stat. Signal Process. Workshop (SSP), pp. 378–382, IEEE, 2018

    2018

  8. [16]

    Advances and challenges in multimodal remote sensing image registration,

    B. Zhu, L. Zhou, S. Pu, J. Fan, and Y . Ye, “Advances and challenges in multimodal remote sensing image registration,”IEEE J. Miniatur. Air Space Syst., vol. 4, no. 2, pp. 165–174, 2023

    2023

  9. [17]

    An integrated approach to registration and fusion of hyperspectral and multispectral images,

    Y . Zhou, A. Rangarajan, and P. D. Gader, “An integrated approach to registration and fusion of hyperspectral and multispectral images,”IEEE Trans. Geosci. Remote Sens., vol. 58, no. 5, pp. 3020–3033, 2019

    2019

  10. [18]

    Unaligned hyperspectral image fusion via registration and interpolation modeling,

    J. Ying, H.-L. Shen, and S.-Y . Cao, “Unaligned hyperspectral image fusion via registration and interpolation modeling,”IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–14, 2021

    2021

  11. [19]

    Simultaneous hyperspectral image super-resolution and geometric alignment with a hybrid camera system,

    Y . Fu, Y . Zheng, L. Zhang, Y . Zheng, and H. Huang, “Simultaneous hyperspectral image super-resolution and geometric alignment with a hybrid camera system,”Neurocomputing, vol. 384, pp. 282–294, 2020

    2020

  12. [20]

    Unsupervised blind hyperspectral super- resolution for unregistered images,

    B. Hu, X. Zhang, and K. Gao, “Unsupervised blind hyperspectral super- resolution for unregistered images,” inProc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), pp. 9349–9352, 2024

    2024

  13. [21]

    Stereo cross-attention network for unregistered hyperspectral and multispectral image fusion,

    Y . Guo, X. Fu, M. Xu, and S. Jia, “Stereo cross-attention network for unregistered hyperspectral and multispectral image fusion,”IEEE Trans. Geosci. Remote Sens., 2023

    2023

  14. [22]

    Deep unfolding network enhanced by transformer priors for unregistered hyperspectral and mul- tispectral image fusion,

    J. Fang, J. Yang, A. Khader, and L. Xiao, “Deep unfolding network enhanced by transformer priors for unregistered hyperspectral and mul- tispectral image fusion,”IEEE Trans. Geosci. Remote Sens., 2024

    2024

  15. [23]

    Unregistered hyperspectral and multispectral image fusion with synchronous nonnegative matrix factorization,

    W. Chen and X. Lu, “Unregistered hyperspectral and multispectral image fusion with synchronous nonnegative matrix factorization,” inProc. Chinese Conf. Pattern Recognit. Comput. Vis. (PRCV), pp. 602–614, Springer, 2020

    2020

  16. [24]

    Mixed 2D/3D convolutional network for hyperspectral image super-resolution,

    Q. Li, Q. Wang, and X. Li, “Mixed 2D/3D convolutional network for hyperspectral image super-resolution,”Remote Sens., vol. 12, no. 10, p. 1660, 2020

    2020

  17. [25]

    Blind super-resolution of single remotely sensed hyperspectral image,

    Z. Liang, S. Wang, T. Zhang, and Y . Fu, “Blind super-resolution of single remotely sensed hyperspectral image,”IEEE Trans. Geosci. Remote Sens., vol. 61, pp. 1–14, 2023

    2023

  18. [26]

    Hyperspectral super-resolution accounting for spectral variability: Coupled tensor LL1-based recovery and blind unmixing of the unknown super-resolution image,

    C. Pr ´evost, R. A. Borsoi, K. Usevich, D. Brie, J. C. Bermudez, and C. Richard, “Hyperspectral super-resolution accounting for spectral variability: Coupled tensor LL1-based recovery and blind unmixing of the unknown super-resolution image,”SIAM J. Imaging Sci., vol. 15, no. ...

    2022

  19. [27]

    A signal processing perspective on hyperspectral unmixing: Insights from remote sensing,

    W.-K. Ma, J. M. Bioucas-Dias, T.-H. Chan, N. Gillis, P. Gader, A. J. Plaza, A. Ambikapathi, and C.-Y . Chi, “A signal processing perspective on hyperspectral unmixing: Insights from remote sensing,”IEEE Signal Process. Mag., vol. 31, no. 1, pp. 67–81, 2013

    2013

  20. [28]

    Provable unregistered hyperspectral- multispectral image fusion via spectral unmixing and adversarial learn- ing,

    J. Song, S. Shrestha, and X. Fu, “Provable unregistered hyperspectral- multispectral image fusion via spectral unmixing and adversarial learn- ing,” inProc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP),

  21. [29]

    Nonnegative ma- trix factorization for signal and data analytics: Identifiability, algorithms, and applications,

    X. Fu, K. Huang, N. D. Sidiropoulos, and W.-K. Ma, “Nonnegative ma- trix factorization for signal and data analytics: Identifiability, algorithms, and applications,”IEEE Signal Process. Mag., vol. 36, no. 2, pp. 59–80, 2019

    2019

  22. [30]

    Self-dictionary sparse regression for hyperspectral unmixing: Greedy pursuit and pure pixel search are related,

    X. Fu, W.-K. Ma, T.-H. Chan, and J. M. Bioucas-Dias, “Self-dictionary sparse regression for hyperspectral unmixing: Greedy pursuit and pure pixel search are related,”IEEE J. Sel. Topics Signal Process., vol. 9, no. 6, pp. 1128–1141, 2015

    2015

  23. [31]

    A variable splitting augmented Lagrangian ap- proach to linear spectral unmixing,

    J. M. Bioucas-Dias, “A variable splitting augmented Lagrangian ap- proach to linear spectral unmixing,” inProc. IEEE WHISPERS, pp. 1–4, 2009

    2009

  24. [32]

    Fast and structured block-term tensor decomposition for hyperspectral unmixing,

    M. Ding, X. Fu, and X.-L. Zhao, “Fast and structured block-term tensor decomposition for hyperspectral unmixing,”IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 16, pp. 1691–1709, 2023

    2023

  25. [33]

    Matrix-vector nonnegative tensor factorization for blind unmixing of hyperspectral imagery,

    Y . Qian, F. Xiong, S. Zeng, J. Zhou, and Y . Y . Tang, “Matrix-vector nonnegative tensor factorization for blind unmixing of hyperspectral imagery,”IEEE Trans. Geosci. Remote Sens., vol. 55, no. 3, pp. 1776– 1792, 2016

    2016

  26. [34]

    Hyperspectral unmixing via total variation regularized nonnegative tensor factorization,

    F. Xiong, Y . Qian, J. Zhou, and Y . Y . Tang, “Hyperspectral unmixing via total variation regularized nonnegative tensor factorization,”IEEE Trans. Geosci. Remote Sens., vol. 57, no. 4, pp. 2341–2357, 2018

    2018

  27. [35]

    Decompositions of a higher-order tensor in block terms—Part II: Definitions and uniqueness,

    L. De Lathauwer, “Decompositions of a higher-order tensor in block terms—Part II: Definitions and uniqueness,”SIAM J. Matrix Anal. Appl., vol. 30, no. 3, pp. 1033–1066, 2008

    2008

  28. [36]

    Spectrum cartography via coupled block-term tensor decomposition,

    G. Zhang, X. Fu, J. Wang, X.-L. Zhao, and M. Hong, “Spectrum cartography via coupled block-term tensor decomposition,”IEEE Trans. Signal Process., vol. 68, pp. 3660–3675, 2020

    2020

  29. [37]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” inAdv. Neural Inf. Process. Syst., 2014

    2014

  30. [38]

    Multi-domain image gen- eration and translation with identifiability guarantees,

    S. Xie, L. Kong, M. Gong, and K. Zhang, “Multi-domain image gen- eration and translation with identifiability guarantees,” inThe Eleventh international conference on learning representations, 2023

    2023

  31. [39]

    Partial disentanglement for domain adaptation,

    L. Kong, S. Xie, W. Yao, Y . Zheng, G. Chen, P. Stojanov, V . Akinwande, and K. Zhang, “Partial disentanglement for domain adaptation,” inIn- ternational conference on machine learning, pp. 11455–11472, PMLR, 2022

    2022

  32. [40]

    Understanding latent correlation- based multiview learning and self-supervision: An identifiability per- spective,

    Q. Lyu, X. Fu, W. Wang, and S. Lu, “Understanding latent correlation- based multiview learning and self-supervision: An identifiability per- spective,” inInternational Conference on Learning Representations, 2022

    2022

  33. [41]

    Self-supervised disentanglement by leveraging structure in data augmentations,

    C. Eastwood, J. von K ¨ugelgen, L. Ericsson, D. Bouchacourt, P. Vincent, M. Ibrahim, and B. Sch ¨olkopf, “Self-supervised disentanglement by leveraging structure in data augmentations,” inCausal Representation Learning Workshop at NeurIPS 2023, 2023

    2023

  34. [42]

    Identifiable causal representation learning: Unsupervised, multi-view, and multi-environment,

    J. von K ¨ugelgen, “Identifiable causal representation learning: Unsupervised, multi-view, and multi-environment,”arXiv preprint arXiv:2406.13371, 2024. 16

    2024 arXiv

  35. [43]

    Rethinking coupled tensor analysis for hy- perspectral super-resolution: Recoverable modeling under endmember variability,

    M. Ding and X. Fu, “Rethinking coupled tensor analysis for hy- perspectral super-resolution: Recoverable modeling under endmember variability,”SIAM Journal on Imaging Sciences,accepted, 2025

    2025

  36. [44]

    Towards identifiable unsupervised domain translation: A diversified distribution matching approach,

    S. Shrestha and X. Fu, “Towards identifiable unsupervised domain translation: A diversified distribution matching approach,” inProc. 12th Int. Conf. Learn. Represent. (ICLR), 2024

    2024

  37. [45]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inProc. Int. Conf. Med. Image Comput. Comput.-Assist. Interv. (MICCAI), pp. 234–241, Springer, 2015

    2015

  38. [46]

    220-band A VIRIS hyperspectral image data set: June 12, 1992, Indian Pines test site 3

    M. Baumgardner, L. Biehl, and D. Landgrebe, “220-band A VIRIS hyperspectral image data set: June 12, 1992, Indian Pines test site 3.” Purdue University Research Repository, 2015

    1992

  39. [47]

    Normalized total gradient: A new measure for multispectral image registration,

    S.-J. Chen, H.-L. Shen, C. Li, and J. H. Xin, “Normalized total gradient: A new measure for multispectral image registration,”IEEE Trans. Image Process., vol. 27, no. 3, pp. 1297–1310, 2017

    2017

  40. [48]

    Generalized assorted pixel camera: postcapture control of resolution, dynamic range, and spectrum,

    F. Yasuma, T. Mitsunaga, D. Iso, and S. K. Nayar, “Generalized assorted pixel camera: postcapture control of resolution, dynamic range, and spectrum,”IEEE Trans. Image Process., vol. 19, no. 9, pp. 2241–2253, 2010

    2010

  41. [49]

    AID: A benchmark data set for performance evaluation of aerial scene classification,

    G.-S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y . Zhong, L. Zhang, and X. Lu, “AID: A benchmark data set for performance evaluation of aerial scene classification,”IEEE Trans. Geosci. Remote Sens., vol. 55, no. 7, pp. 3965–3981, 2017

    2017

  42. [50]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 586–595, 2018

    2018

  43. [51]

    GANs trained by a two time-scale update rule converge to a local Nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “GANs trained by a two time-scale update rule converge to a local Nash equilibrium,”Adv. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  44. [52]

    Lanczos filtering in one and two dimensions,

    C. E. Duchon, “Lanczos filtering in one and two dimensions,”J. Appl. Meteorol., pp. 1016–1022, 1979

    1979

  45. [53]

    Earth Explorer

    U.S. Geological Survey, “Earth Explorer.” https://earthexplorer.usgs. gov/. Accessed: Jun. 23, 2025. 17 Supplementary Materials APPENDIXA PROOF OFTHEOREM1 The proof is similar to that under co-registered HMF settings [5], with some critical differences. For example, the unregi...

    2025

  46. [54]

    Conv(C in →C out, K-3, S-1, ZP-1, BN, LReLU)

  47. [55]

    Conv(C out →C out, K-3, S-1, ZP-1, BN, LReLU) The decoder block Dec(C in →C out, Skip-layer-i) performs bilinear upsampling, followed by convolution and concatena- tion with the corresponding encoder feature map, and two additional convolutional layers:

  48. [56]

    Conv(C in →C out, K-1, S-1, ZP-0, BN, LReLU)

  49. [57]

    Conv(2×C out →C out, K-3, S-1, ZP-1, BN, LReLU)

  50. [58]

    Conv(C out →C out, K-3, S-1, ZP-1, BN, LReLU) Inverse Mapping Neural Networkg.we adopt a simpler architecture for inverse mappingg, consisting of a sequence of residual blocks and a bilinear downsampling in the mid- dle. The residual block ResBlock(C in →C out) operates two co...

  51. [59]

    Conv(C in →C out, K-3, S-1, RP-1, BN, LReLU)

  52. [60]

    Discriminatord r.Finally, to handle the multiple abundance maps translation task, we employ a multi-task discriminator described in Table

    Conv(C out →C out, K-3, S-1, RP-1, BN,#) The architecture of the inverse mapping neural networkgis presented in Table X. Discriminatord r.Finally, to handle the multiple abundance maps translation task, we employ a multi-task discriminator described in Table. XI, whereRdenotes...

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.