pith. sign in

arxiv: 2601.16602 · v1 · submitted 2026-01-23 · 📡 eess.IV · cs.GR· eess.SP

Unsupervised Super-Resolution of Hyperspectral Remote Sensing Images Using Fully Synthetic Training

Xinxin Xu (LTCI , IDS , IP Paris , IMAGES) , Yann Gousseau (LTCI , Christophe Kervazo (IDS , IMAGES , LTCI)
show 1 more author
Sa\"id Ladjal (IMAGES
This is my paper

Pith reviewed 2026-05-16 12:08 UTC · model grok-4.3

classification 📡 eess.IV cs.GReess.SP
keywords hyperspectral super-resolutionunsupervised learningsynthetic training datadead leaves modelspectral unmixingabundance mapsremote sensing
0
0 comments X

The pith

A neural network trained only on synthetic abundance maps can super-resolve real hyperspectral images without ground-truth data.

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

The paper introduces an unsupervised super-resolution method for hyperspectral remote sensing images. It first decomposes the input image into endmembers and abundances through unmixing. A neural network learns to increase the spatial resolution of abundances after training exclusively on synthetic abundance maps produced by the dead leaves model to match real statistical properties. The upsampled abundances are then recombined with the original endmembers to produce the high-resolution hyperspectral output. This approach matters because collecting matched low- and high-resolution hyperspectral pairs for supervised training is rarely feasible in remote sensing.

Core claim

The central claim is that synthetic abundance maps generated by the dead leaves model faithfully reproduce the statistics of real hyperspectral abundance data, allowing an abundance super-resolution network trained solely on these maps to generalize to real inputs and deliver effective unsupervised super-resolution when recombined with endmembers.

What carries the argument

The dead leaves model for generating synthetic abundance maps that mimic real hyperspectral abundance statistics to train the super-resolution network.

If this is right

  • The method removes any need for paired ground-truth high-resolution hyperspectral data during training.
  • It can be applied directly to any real low-resolution hyperspectral image through unmixing, network upsampling, and recombination.
  • Experiments demonstrate that the synthetic training data supports effective generalization and overall method performance.
  • Spectral fidelity is preserved because endmembers remain unchanged while only spatial resolution of abundances is enhanced.

Where Pith is reading between the lines

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

  • The same synthetic-generation strategy could extend to other remote-sensing enhancement tasks where real paired data is scarce.
  • Success hinges on how accurately the dead leaves model captures abundance correlations, pointing to tests with alternative generative models for abundances.
  • The modular split between unmixing and abundance super-resolution suggests the pipeline could incorporate improved unmixing algorithms without retraining the spatial network.

Load-bearing premise

The dead leaves model generates synthetic abundance maps that faithfully mimic the statistics of real abundance data extracted from hyperspectral images.

What would settle it

A network trained on dead-leaves synthetic abundances performing no better than one trained on mismatched or random maps when tested on real hyperspectral super-resolution tasks would falsify the claim.

read the original abstract

Considerable work has been dedicated to hyperspectral single image super-resolution to improve the spatial resolution of hyperspectral images and fully exploit their potential. However, most of these methods are supervised and require some data with ground truth for training, which is often non-available. To overcome this problem, we propose a new unsupervised training strategy for the super-resolution of hyperspectral remote sensing images, based on the use of synthetic abundance data. Its first step decomposes the hyperspectral image into abundances and endmembers by unmixing. Then, an abundance super-resolution neural network is trained using synthetic abundances, which are generated using the dead leaves model in such a way as to faithfully mimic real abundance statistics. Next, the spatial resolution of the considered hyperspectral image abundances is increased using this trained network, and the high resolution hyperspectral image is finally obtained by recombination with the endmembers. Experimental results show the training potential of the synthetic images, and demonstrate the method effectiveness.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes an unsupervised super-resolution method for hyperspectral remote sensing images. It first applies unmixing to decompose a low-resolution hyperspectral image into abundance maps and endmembers. A neural network for abundance super-resolution is then trained exclusively on synthetic abundance maps generated via the dead leaves model, chosen to mimic the spatial statistics of real abundances. The trained network is applied to the real low-resolution abundances, and the resulting high-resolution abundances are recombined with the endmembers to yield the super-resolved hyperspectral image. The abstract states that experiments demonstrate the training potential of the synthetic data and the overall effectiveness of the method.

Significance. If the synthetic abundances are shown to match real statistics sufficiently well, the approach would offer a practical unsupervised route to hyperspectral super-resolution in domains where paired high-resolution ground truth is unavailable, potentially reducing reliance on supervised training and enabling broader application in remote sensing.

major comments (2)
  1. [Abstract] Abstract: the claim that 'experimental results show the training potential of the synthetic images, and demonstrate the method effectiveness' is unsupported by any reported datasets, quantitative metrics (PSNR, SSIM, SAM, etc.), baselines, or error analysis. Without these, the central empirical claim cannot be evaluated.
  2. [Method (synthetic generation)] Synthetic abundance generation (described in the method): the assertion that the dead leaves model 'faithfully mimic[s] real abundance statistics' is load-bearing for generalization, yet no quantitative validation is supplied (e.g., power spectra, variograms, autocorrelation functions, or distributional distances between synthetic and unmixing-derived real abundance maps). Visual examples alone are insufficient to rule out overfitting to synthetic artifacts.
minor comments (2)
  1. [Method (unmixing step)] Clarify the precise unmixing algorithm and any regularization used to extract the real abundances that serve as the reference for synthetic mimicry.
  2. [Method (network training)] Specify the architecture and loss function of the abundance super-resolution network, including how the sum-to-one and non-negativity constraints are enforced during training and inference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript arXiv:2601.16602. We address each major comment point by point below, agreeing where revisions are needed to strengthen the empirical claims and validation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'experimental results show the training potential of the synthetic images, and demonstrate the method effectiveness' is unsupported by any reported datasets, quantitative metrics (PSNR, SSIM, SAM, etc.), baselines, or error analysis. Without these, the central empirical claim cannot be evaluated.

    Authors: We acknowledge the referee's point that the abstract's phrasing requires clearer empirical grounding. The manuscript's Section 4 presents results on real hyperspectral remote sensing datasets via visual comparisons of super-resolved images. To directly address this, we will revise the abstract to explicitly reference the datasets used and summarize key quantitative metrics (such as PSNR and SAM) along with baseline comparisons from the experiments. This revision will ensure the central claim is properly supported without altering the method description. revision: yes

  2. Referee: Synthetic abundance generation (described in the method): the assertion that the dead leaves model 'faithfully mimic[s] real abundance statistics' is load-bearing for generalization, yet no quantitative validation is supplied (e.g., power spectra, variograms, autocorrelation functions, or distributional distances between synthetic and unmixing-derived real abundance maps). Visual examples alone are insufficient to rule out overfitting to synthetic artifacts.

    Authors: We agree that quantitative validation of the dead leaves model's fidelity to real abundance statistics is important for supporting generalization claims. The current manuscript relies on visual examples in the method section to illustrate the similarity. In the revised version, we will incorporate additional quantitative analysis, including comparisons of power spectra, variograms, and autocorrelation functions between the synthetic abundances and those derived from real images via unmixing. This will provide stronger evidence against overfitting to synthetic artifacts and better justify the model's use. revision: yes

Circularity Check

0 steps flagged

No circularity; unsupervised pipeline uses independent dead-leaves synthesis

full rationale

The paper's chain is a forward unsupervised pipeline: unmix the input HS image into abundances/endmembers, generate synthetic abundance maps via the dead-leaves model (tuned to match real statistics), train an SR network exclusively on the synthetics, then apply the network to the unmixed abundances and recombine. No step reduces a claimed prediction or first-principles result to the input data by construction, no fitted parameter is renamed as a prediction, and no self-citation chain is invoked to justify a uniqueness or ansatz that would collapse the argument. The mimicry assumption is external to the derivation and does not create self-reference.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on one key domain assumption about the fidelity of the dead leaves model to real abundance statistics; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption The dead leaves model can generate synthetic abundance maps that faithfully mimic the statistics of real abundance data from hyperspectral images.
    This assumption underpins the entire training strategy and generalization to real data as stated in the abstract.

pith-pipeline@v0.9.0 · 5507 in / 1181 out tokens · 34638 ms · 2026-05-16T12:08:43.274724+00:00 · methodology

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

30 extracted references · 30 canonical work pages · 2 internal anchors

  1. [1]

    Consequently, HSIs are 3-dimensional images having two spatial dimensions and one spectral dimen- sion

    INTRODUCTION In contrast to usual multispectral images, hyperspectral images (HSI) acquire a scene in many contiguous and nar- row spectral bands. Consequently, HSIs are 3-dimensional images having two spatial dimensions and one spectral dimen- sion. Due to their rich spectral information, HSIs have many applications in remote sensing, such as land cover ...

    work page

  2. [2]

    Unsupervised Super-Resolution of Hyperspectral Remote Sensing Images Using Fully Synthetic Training

    to only cite a few. However, due to optical constraints, hyperspectral sensors generally have larger pixels than their multispectral counterparts, leading to images having lower spatial resolution. The work was partially supported by Agence de l’Innovation de D´efense – AID - via Centre Interdisciplinaire d’Etudes pour la D´efense et la S´ecurit´e – CIEDS...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

    PROPOSED METHOD 2.1. Overview To perform super-resolution, the objective is to obtain a high spatial resolution hyperspectral image HSI HR ∈ RL×H×W from a low resolution one, HSI LR ∈ RL×h×w, where h, w, H and W (h < H and w < W ) are the spatial dimensions and L is the spectral dimension. Due to its sim- plicity and despite its limitations [14], we assum...

    work page 1968

  4. [4]

    Dataset & Setup The Urban dataset, a widely-used dataset where there is an exploitable ground-truth of the abundance maps and end- members, is considered here

    EXPERIMENTAL RESULTS 3.1. Dataset & Setup The Urban dataset, a widely-used dataset where there is an exploitable ground-truth of the abundance maps and end- members, is considered here. Urban is a dataset captured by the Hyperspectral Digital Image Collection Experiment (HY- DICE) sensor. It is 307 × 307 pixels in size for 210 bands in the 400 − 2500 nm r...

    work page

  5. [5]

    The main idea of our method is to train a supervised super- resolution network with synthetic dead leaves images only

    CONCLUSION In this paper, we have proposed an unsupervised method for Super Resolution of remote sensing hyperspectral images. The main idea of our method is to train a supervised super- resolution network with synthetic dead leaves images only. This overcomes the problem of lack of datasets encountered by most super-resolution methods. Our method compare...

    work page

  6. [6]

    Prediction of land use/land cover change meth- ods: A study,

    Oumayma Bounouh, Houcine Essid, and Imed Riadh Farah, “Prediction of land use/land cover change meth- ods: A study,” in 2017, ATSIP. IEEE, 2017, pp. 1–7

    work page 2017

  7. [7]

    Evaluation of aerial remote sensing techniques for vegetation management in power-line cor- ridors,

    Steven J Mills, Marcos P Gerardo Castro, Zhengrong Li, Jinhai Cai, Ross Hayward, Luis Mejias, and Rod- ney A Walker, “Evaluation of aerial remote sensing techniques for vegetation management in power-line cor- ridors,” IEEE TGRS, vol. 48, no. 9, pp. 3379–3390, 2010

    work page 2010

  8. [8]

    Unrolling palm for sparse semi- blind source separation,

    Mohammad Fahes, Christophe Kervazo, J´erˆome Bobin, and Florence Tupin, “Unrolling palm for sparse semi- blind source separation,” in ICLR, 2022

    work page 2022

  9. [9]

    Cou- pled convolutional neural network with adaptive response function learning for unsupervised hyperspectral super resolution,

    Ke Zheng, Lianru Gao, Wenzhi Liao, Danfeng Hong, Bing Zhang, Ximin Cui, and Jocelyn Chanussot, “Cou- pled convolutional neural network with adaptive response function learning for unsupervised hyperspectral super resolution,” IEEE TGRS, vol. 59, no. 3, pp. 2487–2502, 2020

    work page 2020

  10. [10]

    Decoupled- and-coupled networks: Self-supervised hyperspectral im- age super-resolution with subpixel fusion,

    Danfeng Hong, Jing Yao, Chenyu Li, Deyu Meng, Naoto Yokoya, and Jocelyn Chanussot, “Decoupled- and-coupled networks: Self-supervised hyperspectral im- age super-resolution with subpixel fusion,” IEEE TGRS, 2023

    work page 2023

  11. [11]

    Super-resolution reconstruction of hyper- spectral images,

    Toygar Akgun, Yucel Altunbasak, and Russell M Mersereau, “Super-resolution reconstruction of hyper- spectral images,” IEEE TIP, vol. 14, no. 11, pp. 1860– 1875, 2005

    work page 2005

  12. [12]

    On projec- tion algorithms for solving convex feasibility problems,

    Heinz H Bauschke and Jonathan M Borwein, “On projec- tion algorithms for solving convex feasibility problems,” SIAM review, vol. 38, no. 3, pp. 367–426, 1996

    work page 1996

  13. [13]

    Hy- perspectral image superresolution by transfer learning,

    Yuan Yuan, Xiangtao Zheng, and Xiaoqiang Lu, “Hy- perspectral image superresolution by transfer learning,” IEEE JSTARS, vol. 10, no. 5, pp. 1963–1974, 2017

    work page 1963

  14. [14]

    Hyperspectral image spatial super- resolution via 3d full convolutional neural network,

    Shaohui Mei, Xin Yuan, Jingyu Ji, Yifan Zhang, Shuai Wan, and Qian Du, “Hyperspectral image spatial super- resolution via 3d full convolutional neural network,” Re- mote Sensing, vol. 9, no. 11, pp. 1139, 2017

    work page 2017

  15. [15]

    Dual 1d-2d spatial-spectral cnn for hyperspectral image super-resolution,

    Jiaojiao Li, Ruxing Cui, Bo Li, Yunsong Li, Shaohui Mei, and Qian Du, “Dual 1d-2d spatial-spectral cnn for hyperspectral image super-resolution,” in IGARSS 2019-2019 IEEE IGRSS. IEEE, 2019, pp. 3113–3116

    work page 2019

  16. [16]

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

    Qiang Li, Qi Wang, and Xuelong Li, “Mixed 2d/3d convolutional network for hyperspectral image super- resolution,” Remote sensing, vol. 12, no. 10, pp. 1660, 2020

    work page 2020

  17. [17]

    Hyperspectral im- age superresolution using spectrum and feature context,

    Qi Wang, Qiang Li, and Xuelong Li, “Hyperspectral im- age superresolution using spectrum and feature context,” IEEE TIE, vol. 68, no. 11, pp. 11276–11285, 2020

    work page 2020

  18. [18]

    Fully synthetic training for image restoration tasks,

    Rapha¨el Achddou, Yann Gousseau, and Sa ¨ıd Ladjal, “Fully synthetic training for image restoration tasks,” CVIU, vol. 233, pp. 103723, 2023

    work page 2023

  19. [19]

    Provably robust blind source separation of linear- quadratic near-separable mixtures,

    Christophe Kervazo, Nicolas Gillis, and Nicolas Dobi- geon, “Provably robust blind source separation of linear- quadratic near-separable mixtures,” SIAM Journal on Imaging Sciences, vol. 14, no. 4, pp. 1848–1889, 2021

    work page 2021

  20. [20]

    Modele s ´equentiel de partition al´eatoire,

    Georges Matheron, “Modele s ´equentiel de partition al´eatoire,” Tech. Rep., Technical report, CMM, 1968

    work page 1968

  21. [21]

    The dead leaves model: a general tessellation modeling occlusion,

    Charles Bordenave, Yann Gousseau, and Francois Roueff, “The dead leaves model: a general tessellation modeling occlusion,” Advances in applied probability, vol. 38, no. 1, pp. 31–46, 2006

    work page 2006

  22. [22]

    The size of objects in natural and artificial images,

    Luis Alvarez, Yann Gousseau, and Jean-Michel Morel, “The size of objects in natural and artificial images,” in AIEP, vol. 111, pp. 167–242. Elsevier, 1999

    work page 1999

  23. [23]

    Oc- clusion models for natural images: A statistical study of a scale-invariant dead leaves model,

    Ann B Lee, David Mumford, and Jinggang Huang, “Oc- clusion models for natural images: A statistical study of a scale-invariant dead leaves model,”IJCV, vol. 41, pp. 35–59, 2001

    work page 2001

  24. [24]

    Modeling occlu- sion and scaling in natural images,

    Yann Gousseau and Franc ¸ois Roueff, “Modeling occlu- sion and scaling in natural images,” MMS, vol. 6, no. 1, pp. 105–134, 2007

    work page 2007

  25. [25]

    Perfect simulation in stochastic geometry,

    Wilfrid S Kendall and Elke Th¨onnes, “Perfect simulation in stochastic geometry,” Pattern recognition, vol. 32, no. 9, pp. 1569–1586, 1999

    work page 1999

  26. [26]

    Residual dense network for image super- resolution,

    Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu, “Residual dense network for image super- resolution,” in IEEE CVPR, 2018, pp. 2472–2481

    work page 2018

  27. [27]

    Hyperspectral Unmixing: Ground Truth Labeling, Datasets, Benchmark Performances and Survey

    Feiyun Zhu, “Hyperspectral unmixing: ground truth labeling, datasets, benchmark performances and survey,” arXiv preprint arXiv:1708.05125, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  28. [28]

    A review of spatial enhancement of hyperspectral remote sensing imaging techniques,

    Nour Aburaed, Mohammed Q Alkhatib, Stephen Mar- shall, Jaime Zabalza, and Hussain Al Ahmad, “A review of spatial enhancement of hyperspectral remote sensing imaging techniques,” IEEE JSTARS, vol. 16, pp. 2275– 2300, 2023

    work page 2023

  29. [29]

    Learning spatial-spectral prior for super-resolution of hyperspectral imagery,

    Junjun Jiang, He Sun, Xianming Liu, and Jiayi Ma, “Learning spatial-spectral prior for super-resolution of hyperspectral imagery,” IEEE TCI, vol. 6, pp. 1082– 1096, 2020

    work page 2020

  30. [30]

    Hyperspectral image super-resolution with rgb image super-resolution as an auxiliary task,

    Ke Li, Dengxin Dai, and Luc Van Gool, “Hyperspectral image super-resolution with rgb image super-resolution as an auxiliary task,” in IEEE/CVF WACV, 2022, pp. 3193–3202

    work page 2022