An Unsupervised Cascade Fusion Network for Radiometrically-Accurate Vis-NIR-SWIR Hyperspectral Sharpening

Round 1

Reviewer 1 Report

The manuscript describes a new unsupervised pansharpening method that is specifically targeting SWIR bands in the multispectral images. These bands currently suffer from low spatial resolution. The manuscript provides a thorough analysis of literature and offers a new solution. The proposed method seems to be superior that other more established techniques.

 

One comment:

 

It is unclear whether the authors will upload the code to the UCFNet, and other codes relevant to the study. There is no section such as Data Sharing and no Data Availability Statement

Author Response

Thanks for your valuable comments on our manuscript. We are working on publishing the code of the proposed method. Once we get approval from the government sponsor, the code will be uploaded as soon as possible.

Reviewer 2 Report

The authors present a novel image sharpening method by fusing spectral details from LR-HSI with the spatial information from the HR-MSI. The paper is interesting, well written, and in good shape. 

Line 67: How was the reference data generated to evaluate the per-pixel classification?

Section 3.4: How does the cascade up-sampling method perform better with the large up-sampling ratio? Please clarify it in the manuscript.

Line 230: The sentence "sharpening. Specifically, our method only learns to improve the spatial resolution of the LR-HSI by fusing the spectral" seems grammatically incorrect. Please revise it.

The Section 4.6 needs to be improved/made elaborate by reasoning a bit on the effects of up-sampling to initialize the HR-VNIR-HSL.

Author Response

Thank you for the valuable comments. We have carefully revised the manuscript based on the suggestions.

1) Line 67: How was the reference data generated to evaluate the per-pixel classification?

A: The reference data is generated based on the reference HR-HSI. More precisely, the MaxD and Gram matrix algorithms are used to extract endmembers, and the SAM is applied to create the ground truth class map. The details are introduced in section 4.2 Evaluation Criterion.

We also clarified this in the introduction part. Please check it in the revised manuscript.

2) Section 3.4: How does the cascade up-sampling method perform better with the large up-sampling ratio? Please clarify it in the manuscript.

A: The advantage of the cascade up-sampling method is demonstrated by our previous work^[1] (in the ablation study section). Intuitively speaking, the cascade sharpening method allows the model to iteratively learn to expand the spatial resolution of the HSI by a factor of two, instead of directly jumping to the target resolution with a large ratio. This coarse-to-fine fashion not only facilitates the network to better fuse the LR-HSI and HR-MSI, but also makes the sharpening process more stable. Therefore, in this work, the cascade up-sampling method is performed to make the UCFNet better handle large up-scale ratios. 

We have added this clarification to section 3.4. Please check it in the revised manuscript.

3) Line 230: The sentence "sharpening. Specifically, our method only learns to improve the spatial resolution of the LR-HSI by fusing the spectral" seems grammatically incorrect. Please revise it.

A: Sorry for the ambiguous expression. The sentence "Specifically, our  method only learns to improve the spatial resolution of the LR-HSI by fusing the spectral details from LR-HSI with the spatial information from the HR-MSI." (Line 401) is changed to " With the spectral information provided by the LR-HSI and the spatial details offered by the HR-MSI, our method is able to effectively improve the spatial resolution of the full-spectrum LR-HSI while maintaining high radiometric accuracy." 

Please check it in the revised manuscript.

 4) The Section 4.6 needs to be improved/made elaborate by reasoning a bit on the effects of up-sampling to initialize the HR-VNIR-HSL.

A: The main difference between using bilinear and GLP-HS for initializing the HR-VNIR-HSI is that bilinear upsampling is a simple mathematic interpolation algorithm that only utilizes the spatial information of the LR-HSI, while GLP-HS is designed for hyperspectral sharpening, which fully exploits the spectral and spatial correlations between HSI and MSI. Moreover, regarding conventional sharpening algorithms, the experiment results in work^[1] show that GLP-HS has the most effective and stable performance with small spatial resolution ratios. Therefore, we chose GLP-HS to initialize the HR-VNIR-HSI in this paper.

We also added this clarification to section 4.6. Please check it in the revised manuscript.

Reference

[1]Huang, Sihan, and David W. Messinger. "An Unsupervised Laplacian Pyramid Network for Radiometrically Accurate Data Fusion of Hyperspectral and Multispectral Imagery." IEEE Transactions on Geoscience and Remote Sensing 60 (2022): 1-17.

Reviewer 3 Report

Main remarks

         The manuscript “An Unsupervised Cascade Fusion Network for Radiometrically-Accurate Vis-NIR-SWIR Hyperspectral Sharpening” is carefully written and deals with a very important subject, namely Hyperspectral sharpening.

The manuscript presents high-quality remote sensing analyses, namely unsupervised cascade fusion network (UCFNet) which focuses on improving the spatial resolution of the full-spectrum HSI, while maintaining high per-pixel spectral accuracy. The proposed approach as well as the network architecture is thoroughly explained in detail. Also, the validation of the results is done, appropriately. The article contains many elements of the scientific novelty mentioned, for example, in the abstract. The obtained results are so interesting that it is undoubtedly worth publishing. The whole article is very thorough, I did not notice any significant weaknesses in it. However, I would suggest that the authors remove from the abstract the sentence "The experimental results show that the proposed sharpening method excels the state-of-the-art methods in terms of radiometric accuracy, spatial accuracy, and classification performance. ", which I find too categorical and too general.

        

Summary

I recommend the manuscript for publication in RS in the present state. My very minor suggestion could be addressed in the editorial phase.

 

                                                                  Sincerely yours,

 

                                                                          Reviewer

Author Response

Thank you for your valuable comments on our manuscript. We have removed the sentence from the abstract section. Please check it on the revised manuscript.