acceptodds
Under review as a conference paper at ICLR 2027

CALSRN: Clustering Attention-Based Low-Dimensional Spectral Reconstruction Network for Hyperspectral Image Super-Resolution

Abstract

Hyperspectral image single-image super-resolution (HSI-SISR) aims to reconstruct a high-resolution hyperspectral image (HR-HSI) with high spatial resolution and high spectral fidelity from only a low-resolution observation. Existing methods still face two key challenges when processing hyperspectral images(HSIs). On the one hand, HSIs contain a large number of continuous spectral bands with severe spectral redundancy, and directly modeling features in the original highdimensional spectral space introduces heavy computational burden. On the other hand, pixels belonging to the same land-cover category may be spatially scattered while sharing similar spectral responses. Existing attention mechanisms based on fixed and rectangle windows, spectral-dimensional attention, or Top-k selection are still limited in directly modeling such pixel-level non-local self-similarity under relatively low computational complexity. To address these issues, this paper proposes a clustering attention-based low-dimensional spectral reconstruction network (CALSRN). First, a parameter-free spectral compression module is designed to compress the original high-dimensional HSI into a compact spectral representation, thereby reducing spectral redundancy and the computational cost of subsequent attention modeling. Second, a clustering attention block (CAB) is proposed. Specifically, clustering is performed on the compressed spectral HSI at the ×1 scale level to pre-generate a pixel membership matrix, which is then used to guide attention computation. In this way, spatially distant pixels with similar spectral responses can effectively exchange information. To address the difficulty of parallel computation caused by unequal cluster sizes, CAB further adopts a category-sorting-based fixed-size grouping strategy and an adjacent-group expansion strategy. Meanwhile, pixel-to-category-center inter-group cross-attention is introduced to incorporate global category-level contextual information. Third, hierarchical reconstruction loss and dynamic weight balancing are introduced to supervise intermediate reconstruction results and stabilize progressive multi-scale training. Experimental results demonstrate that the proposed CALSRN effectively improves spatial detail recovery and spectral fidelity for hyperspectral image super-resolution.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hyperspectral image, single-image super resolution, clustering attention, spectral compression, progressive reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_PsaPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CALSRN for hyperspectral image single-image super-resolution. It combines parameter-free spectral compression, clustering-guided attention, and progressive reconstruction to exploit non-local spectral similarities while reducing computational complexity. Experiments on three datasets demonstrate competitive reconstruction performance and model efficiency.

Critical strengths
  1. The paper addresses spectral redundancy and non-local similarities in HSIs through clustering-guided feature interactions.
  2. The proposed CAB combines intra-group and inter-group attention to capture complementary pixel-level and category-level information.
  3. CALSRN achieves competitive performance across three datasets and multiple scaling factors, with relatively few parameters and low reported FLOPs. Extensive ablation studies further support its effectiveness.
Critical weaknesses
  1. CALSRN shares similarities with SCPSN in spectral compression, non-local modeling, and progressive reconstruction. Clustering-based attention has also been explored previously. The authors should better distinguish their contributions from existing approaches.
  2. Hard K-means clustering is performed on only three compressed spectral channels, with labels propagated across scales. However, if cluster labels are not re-estimated, this may result in incorrect cross-scale feature correlations. Besides, the authors acknowledge that hard clustering may be unreliable for mixed pixels and unknown sensor degradation; however, mixed pixels are inherent in HSI SR tasks, and the authors do not effectively analyze whether K-means is truly applicable or how much it affects the proposed method.
Questions for authors
  1. What are the key methodological differences between CALSRN and SCPSN or existing clustering-based attention methods?
  2. How robust is CAB to inaccurate clustering assignments, particularly under large scaling factors?
  3. Unclear attention grouping mechanism. Fixed-size grouping may separate tokens from the same cluster, while adjacent-group expansion may introduce unrelated tokens. Further clarification of cross-cluster interactions and boundary handling is needed.
  4. The author uses averaging across adjacent bands rather than PCA, SVD, or learnable low-rank decomposition. Does this approach offer any advantages?
  5. Some figure captions in the appendix are inconsistent with the displayed content, such as Figures 7 and 8.
1 comment

Dear Reviewer,

We have posted a point-by-point response to your comments (Q1–Q5) in the thread above. Could you kindly let us know whether our revisions and responses resolve your concerns? If any question remains, we would be glad to discuss further and provide additional analysis or clarification.

Thank you for your time.

Reviewer_XpqtOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes CALSRN, a clustering attention-based low-dimensional spectral reconstruction network for hyperspectral image single-image super-resolution. The method first compresses the original hyperspectral bands into a low-dimensional representation through parameter-free spectral averaging. It then introduces a clustering attention block to capture non-local relationships among spectrally similar pixels, combining intra-group attention with pixel-to-cluster-center cross-attention. A progressive reconstruction architecture with hierarchical supervision and dynamic loss balancing is further adopted. Experiments on Pavia Center, Botswana, and Washington DC Mall under ×2, ×4, and ×8 scaling factors demonstrate competitive reconstruction accuracy and relatively low computational complexity.

Critical strengths

Relevant motivation and efficient architecture. The paper addresses spectral redundancy and non-local spatial dependencies in hyperspectral image super-resolution. Performing attention in a compressed spectral space is a reasonable approach to reducing computational cost.

Competitive quantitative performance. The proposed method achieves strong results across three benchmark datasets and multiple scaling factors. The relatively small parameter count and computational cost also suggest potential advantages for efficient hyperspectral reconstruction.

Critical weaknesses

Limited methodological novelty. The proposed framework mainly integrates established techniques, including spectral averaging, K-means clustering, grouped self-attention, cluster-center cross-attention, and progressive reconstruction. The clustering attention mechanism is conceptually related to existing content-adaptive token grouping and prototype-based attention methods. The paper does not sufficiently establish a fundamentally new attention mechanism or learning principle beyond combining these existing operations.

Insufficient justification of the clustering prior and spectral compression strategy. The proposed method relies on hard K-means clustering performed on a highly compressed spectral representation, where the original hyperspectral bands are reduced to only three components. Such aggressive compression may discard discriminative spectral information, particularly for spectrally similar materials and mixed pixels. Moreover, the cluster assignments are generated at the lowest resolution and propagated to higher resolutions through nearest-neighbor interpolation. This strategy cannot explicitly account for newly reconstructed spatial details or changes in spectral characteristics across scales. The robustness and theoretical justification of these design choices remain insufficient.

Limited experimental evidence supporting the individual methodological contributions. Although the paper provides component ablation studies, the improvements attributable to several proposed mechanisms are relatively modest. For example, Table 5 shows that removing the category-sorting-based grouping strategy reduces PSNR by only approximately 0.03 dB, while removing either attention branch results in improvements of less than 0.17 dB when restored. More convincing comparisons against alternative clustering, token grouping, and attention mechanisms are needed to demonstrate the necessity of the proposed design.

Insufficient validation under realistic imaging conditions. The experiments mainly rely on synthetic Gaussian blur and downsampling under the Wald protocol. The effectiveness of CALSRN under unknown degradation, sensor noise, spectral distortion, and real-world hyperspectral acquisition conditions remains unclear. Furthermore, the computational analysis primarily emphasizes parameter counts and FLOPs, while the additional overhead associated with clustering, sorting, and grouping deserves more comprehensive evaluation.

Questions for authors

What is the fundamental novelty of the proposed clustering attention mechanism compared with existing content-adaptive attention, token clustering, and prototype-based attention methods?

Why is simple band-wise averaging adopted for spectral compression instead of more advanced dimensionality reduction methods, such as PCA or learnable spectral projections? How can the authors ensure that discriminative spectral information is preserved when compressing hundreds of spectral bands into only three components?

The clustering membership matrix is generated only at the lowest spatial resolution and propagated to higher resolutions using nearest-neighbor interpolation. How does this strategy handle mixed pixels, inaccurate cluster assignments, and newly recovered spatial structures?

Why is hard K-means clustering preferred over soft clustering or differentiable feature grouping? How sensitive is the reconstruction performance to clustering initialization and incorrect cluster assignments?

Table 5 indicates relatively limited improvements from several internal CAB components. Can the authors provide more comprehensive comparisons with alternative grouping strategies and attention mechanisms under matched computational budgets?

The experiments are mainly conducted using synthetic Gaussian degradation. How does CALSRN perform under realistic sensor noise, unknown blur kernels, and different hyperspectral imaging conditions?

Although the proposed method achieves competitive FLOPs and parameter counts, how much additional computational overhead is introduced by K-means clustering, token sorting, and grouping? More comprehensive end-to-end runtime and memory comparisons would help establish the practical efficiency of the method.

How stable are the reported results across different random seeds, clustering initializations, and training configurations? Can the authors provide statistical evaluations to demonstrate the significance of the improvements?

1 comment

Dear Reviewer,

We have posted a point-by-point response to your comments (Q1–Q8) in the thread above. Could you kindly let us know whether our revisions and responses resolve your concerns? If any question remains, we would be glad to discuss further and provide additional analysis or clarification.

Thank you for your time.

Reviewer_Mff6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The manuscript proposed to solve the intrinsic problem possessed by hyperspectral data in remote sensing, which is spectral redundancy and spatially scattered while spectrally similar pixels. The author designed a parameter-free spectral compression module to compress the HSI into a compact representation, and developed a clustering attention mechanism based on the observation that pixels belonging to the same land-cover class often exhibit similar spectral responses.

Critical strengths
  1. The manuscript is clearly presented and easy to follow.
  2. The idea of compressing spectral bands in a parameter-free manner is interesting, and the strategy of grouping spectrally similar yet spatially distant pixels provides an insightful way to exploit non-local self-similariry.
Critical weaknesses
  1. The proposed method is more of an incremental method built on the existing paradigm instead of an independent research, which somewhat lacks of novelty;
  2. The claim 'HSI possess two crucial intrinsic priors' is not accurate since HSI includes remote sensing HSI, natural scene HSI and microscope HSI, it is suggested to limit the discussion in 'HSI in remote sensing' as pixels in remote sensing are spatially distant while spectrally similar;
  3. The label of y-axis in Fig.2 should be inverted. Since in (wang et al.) it can be cited as 'for a band, it has stronger correlation with the adjacent bands within a certain range and low correlation with farther bands', the closer the bands are, the more similar they should be, the authors might make a mistake in marking;
  4. The visualisation demonstration is quite limited, only one scene in one dataset is shown;
  5. The proposed method is evaluated on datasets released before 2009, which leaves the proposed method less convincing. In the cited literature such as Xiao et al., the method was evaluated on datasets released after 2017. The authors are encouraged to evaluate their method on more recent datasets to further support their conclusions.
Questions for authors

See weakness.

1 comment

Dear Reviewer,

We have posted a point-by-point response to your comments (W1–W5) in the thread above. Could you kindly let us know whether our revisions and responses resolve your concerns? If any question remains, we would be glad to discuss further and provide additional analysis or clarification.

Thank you for your time.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.