Cross-Dimension Attention Guided Self-Supervised Remote Sensing Single-Image Super-Resolution

Round 1

Reviewer 1 Report

This paper proposed a cross-dimension attention-guided self-supervised learning method for single-image super-resolution in remote sensing, which is novel and reasonable. Only a single image is needed for training in the proposed method, which does not require prior training in the dataset, unlike the existing image SR methods. It can better adapt to remote sensing super-resolution tasks in different situations, and a large number of experiments show that the performance of the proposed method is superior to that of the most advanced methods at present. The cross-dimensional attention module proposed in this paper is a novel module that combines the weights of learning spatial and channel to improve their interaction.

This paper is innovative and easy to understand. There are several minor issues regarding the description and experiment of this paper, which would better be addressed before publication.

1. Do the letters ‘S’ and ‘s’ in Figure 2 mean the same thing? It is different from the description in the paper. Please verify and correct it further.

2. It is strongly recommended that the author adjust the font in Figure 3 to an appropriate size.

3. In the application of super-resolution reconstruction, the mean square error loss is generally used, but the L1 loss is used in this article. Why use the L1 loss? Can it achieve a better reconstruction effect than the mean square error loss? Give the corresponding explanation.

4. The authors highly suggest discriminating the mathematical notations for matrix, vector, and scale using capital or small letters, or Italic or bolded versions. Current mathematical representations are difficult to follow.  
 
5. In the literature review part, it would be good to provide more related works for the remote sensing super-resolution method and then provide the added values of your own proposed method

6. The authors are advised to read the full text carefully and correct the writing errors. 

7. The paper shows that this method can save a lot of computing resources, which is one of the highlights of this paper. Therefore, it is suggested that the authors explain this viewpoint in detail.

8. It is encouraged to include more discussions about state-of-the-art blind super-resolution algorithms, such as "Learning to super-resolve blurry face and text images", ICCV 2017, and "Zoom to learn, learn to zoom", CVPR 2019.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors proposed a self-supervised SR method for remote sensing dataset. Several concerns need to be clarified in order for this manuscript to be published:

1. please describe the difference between the proposed CDAM and the CBAM method. The concept of two methods are quite similar. A fair comparison of integrating CDAM or CBAM to CDAN is recommended.
2. In terms of creating the non-ideal datasets, the authors used anisotropic Gaussian kernels. Is it a standard protocol of creating non-ideal dataset? Some citations should be added if this is a standard protocol. If not, the authors should discuss why they only applied anisotropic Gaussian kernels in generating the non-ideal datasets.
3. In section 4.3, did the authors retrain other models shown in Table 2 with the non-ideal dataset? Or just tested them using the non-ideal dataset?  
4. In the experiments, the authors tried two scale factors: 2 and 4. It is an interesting question to know the "limitation" (i.e. the maximum value of the scale factor) of the proposed method.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors had replied my concerns so I don't have further questions.

A minor issue is that I think the dataset "RSCNN7" described in line 241 should be "RSSCN7".  This is what I found on internet. Please ignore this comment if I am wrong.

Author Response

Thanks for the reviewer’s careful reading and suggestions. This is our spelling error, and we have modified it in the revised version ("RSCNN7" to "RSSCN7"), see line 241.