acceptodds
Under review as a conference paper at ICLR 2027

Super-Resolving Unseen Hyperspectral Sensors at Any Scale via Spatial Operators

66 views

Abstract

Achieving cross-sensor generalization and arbitrary-scale reconstruction with a single model remains challenging in hyperspectral super-resolution (HSR). Although recent methods support arbitrary-scale reconstruction, applying them to new sensors or scales beyond the training range often requires additional data and computation to maintain reconstruction quality. To address these challenges, we propose OmniHSR, which predicts band-shared spatial operators rather than spectral values. Cross-Spectral Mapping (CSM) resamples inputs with any number of bands to fixed reference positions and predicts local operators with Gaussian supports. Continuous Operator-Field Reconstruction (COFR) composes these operators into a continuous field and applies them to all original bands for arbitrary-scale reconstruction. Experiments demonstrate that operator prediction outperforms direct spectral-value prediction on all seven datasets. Trained solely on ARAD with only 0.538M parameters, OmniHSR outperforms all directly transferred baselines on six unseen datasets without target-domain training data or adaptation. Across twelve upsampling factors from to , it improves average PSNR on Pavia U and Chikusei by 0.55 dB over the strongest baseline. It also surpasses baselines trained from scratch or adapted on the target sensor and achieves up to faster inference. Our code will be publicly released soon.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
hyperspectral image super-resolution, arbitrary-scale super-resolution, cross-sensor generalization, zero-shot transfer

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_phMfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OmniHSR, a framework for cross-sensor arbitrary-scale hyperspectral image super-resolution. Instead of directly predicting spectral values, the method learns band-shared spatial operators to reconstruct high-resolution images with different numbers of spectral bands. The framework includes two modules: Cross-Spectral Mapping (CSM), which converts the input into a fixed number of spectral bands and predicts spatial operators, and Continuous Operator-Field Reconstruction (COFR), which applies these operators to the original spectral bands at arbitrary scales. The main contribution is a unified and lightweight model that can handle different sensors without additional training or adaptation. Experiments on seven datasets show that the proposed method achieves better reconstruction accuracy and computational efficiency compared with existing methods, including on unseen sensors.

Critical strengths

The paper explore an interesting band-shared operator design. The idea of predicting band-shared spatial operators instead of spectral values is interesting. This design allows the model to handle hyperspectral images with different numbers of spectral bands without changing the network architecture. It provides a useful approach to cross-sensor generalization.

The explored framework has good computational efficiency and generalization ability. The proposed method uses a lightweight network with only 0.538M parameters and can be directly applied to unseen sensors without additional training or adaptation. The experimental results show improvements in reconstruction accuracy and inference speed compared with existing methods. These advantages make the proposed approach useful for practical applications.

Critical weaknesses
  1. Insufficient justification of the band-shared operator prior and CSM design.

Although the band-shared operator prior simplifies the model and improves computational efficiency, applying the same spatial operator to all spectral bands may limit the reconstruction performance. Hyperspectral images contain rich spectral correlations, which are important for recovering spatial details and preserving spectral information. Although CSM uses spectral information to predict spatial operators, the reconstruction stage applies the same operator to all bands. This may limit the ability to recover band-specific spatial details. In addition, CSM resamples the input into a fixed number of 31 bands, regardless of the original band number. It is unclear whether this strategy is necessary or optimal. For sensors with many spectral bands, using multiple spectral chunks and combining their features may be more effective. Although the authors provide some ablation experiments, comparisons with spectral-aware reconstruction methods and alternative spectral feature extraction strategies are still insufficient. These limitations make the advantages of the proposed design less convincing.

  1. Limited improvement over simple Bicubic interpolation at large scaling factors.

Although the proposed method outperforms several existing learning-based methods, its improvement over simple Bicubic interpolation becomes very small at large scaling factors. For example, at ×48, the PSNR improvements are only 0.03 dB on Pavia U and 0.01 dB on Chikusei. The SSIM on Pavia U is also the same as Bicubic interpolation (0.483). While the improvements at smaller scaling factors are more noticeable, the results at large factors raise concerns about the actual reconstruction benefit of the proposed method. Therefore, the experimental results do not sufficiently demonstrate the effectiveness of the proposed method at large upsampling factors, especially outside the training range.

Questions for authors

Regarding the band-shared operator prior: Could the authors further explain how the proposed method effectively exploits spectral correlations for high-quality HSI reconstruction? Since the same spatial operator is applied to all spectral bands, would incorporating spectral-adaptive operators improve the reconstruction performance? A comparison with lightweight spectral-aware methods would help justify this design choice.

Regarding the CSM design: Why is it necessary to resample all spectral bands into a fixed number of 31 bands? For sensors with many spectral bands, would processing multiple spectral chunks and combining their features provide better spatial information for operator prediction? Although different values of M are evaluated in the supplementary material, could the authors clarify the advantages of the proposed CSM compared with alternative spectral feature extraction strategies?

Regarding the improvement over Bicubic interpolation: Why does the performance improvement over Bicubic interpolation become almost negligible at large scaling factors, such as ×36 and ×48? Could the authors provide further analysis or visual comparisons to demonstrate what additional spatial details are recovered by the proposed method at these factors?

Reviewer_J44COpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OmniHSR, a framework for cross-sensor and arbitrary-scale hyperspectral image super-resolution. Instead of directly predicting spectral values, the method learns band-shared spatial reconstruction operators. It introduces Cross-Spectral Mapping (CSM) to handle varying numbers of spectral bands and Continuous Operator-Field Reconstruction (COFR) to support arbitrary upsampling factors. Trained only on ARAD, OmniHSR is evaluated on six unseen datasets. The experiments demonstrate improvements over existing methods in reconstruction quality and computational efficiency, supported by ablation studies and additional robustness analyses.

Critical strengths
  1. The paper addresses the important challenge of handling different hyperspectral sensors and upsampling factors with a single model. Predicting band-shared spatial operators instead of spectral values is an interesting design that reduces dependence on the number of spectral bands and improves computational efficiency.

  2. The model is trained on ARAD and achieves strong zero-shot results across six unseen datasets. The evaluation covers multiple upsampling factors and competitive baselines, including target-sensor adaptation experiments. The ablation studies also provide useful evidence for the effectiveness of the key components and the shared-operator assumption.

Critical weaknesses
  1. The main experiments use synthetically generated low-resolution inputs with shared bicubic degradation across spectral bands. Although Appendices C.6 and C.7 provide additional experiments involving band-dependent blur, noise, and alternative degradations, the evaluation remains based on synthetic observations. Moreover, Table 18 shows that the performance advantage can diminish or reverse under stronger noise. Since real hyperspectral sensors may involve more complex degradation and spatial misalignment, it remains unclear how well the method generalizes to real low-resolution sensor measurements.

  2. Although OmniHSR supports upsampling factors up to ×48, its improvement over bicubic interpolation becomes very small at extreme scales. For example, Table 1 shows gains of only 0.03 dB on Pavia U and 0.01 dB on Chikusei at ×48. These results demonstrate that the method can operate beyond its training scale range, but the practical reconstruction benefits at such large factors remain limited. Further discussion of this limitation would help clarify the value of extreme-scale reconstruction.

Questions for authors
  1. Could the authors further clarify how OmniHSR would generalize to real sensor measurements involving noise, band-dependent degradation, and spatial misalignment, beyond the synthetic degradation settings already evaluated?

  2. Given the small improvements over bicubic interpolation at extreme upsampling factors such as ×48, could the authors discuss the practical benefits and limitations of reconstruction at these scales?

  3. CSM uses normalized band positions rather than actual wavelengths. How does this design affect reconstruction performance for sensors with irregular spectral sampling or substantially different wavelength ranges?

  4. Since OmniHSR is trained exclusively on ARAD, could the authors clarify how sensitive its cross-sensor generalization is to the choice of training dataset? Would similar performance be expected when training on a different source sensor?

Reviewer_27bkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes OmniHSR, a unified framework for cross-sensor and arbitrary-scale hyperspectral image super-resolution. By predicting band-shared spatial operators instead of spectral values, it fundamentally decouples the network architecture from the sensor's band count. Utilizing Cross-Spectral Mapping (CSM) and Continuous Operator-Field Reconstruction (COFR), the model achieves superior zero-shot performance and up to 36x faster inference on six unseen datasets.

Critical strengths
  1. Predicting spatial operators rather than spectral values decouples the computational graph from the input band count, enabling true zero-shot cross-sensor generalization.
  2. With only 0.538M parameters, it outperforms target-domain fine-tuned baselines on unseen datasets and achieves up to 36x inference acceleration.
Critical weaknesses
  1. The model relies heavily on a shared spatial operator, which may fail on real-world uncalibrated sensors with wavelength-dependent blur.
  2. CSM maps bands to [0,1] based on index order, ignoring actual physical wavelength intervals. This may lose spectral continuity priors for unevenly distributed bands (e.g., the Botswana dataset).
  3. All six baselines are general RGB arbitrary-scale methods. Recent HSI-specific arbitrary-scale or cross-band methods cited in Section 2 (e.g., SQformer, SPG-ASSR) are missing from the quantitative comparisons.
Questions for authors
  1. How does the "band-shared operator prior" perform on real-world hyperspectral images with significant chromatic aberration?
  2. Could incorporating physical wavelengths instead of uniform index mapping in CSM improve performance on sensors with discontinuous bands?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.