Former-CR: A Transformer-Based Thick Cloud Removal Method with Optical and SAR Imagery
Round 1
Reviewer 1 Report
- The authors propose a Transformer-based Cloud Removal model called Former-CR, which expands the input channel, allowing SAR images and RGB images to be inserted simultaneously.
- The approach is compared with three other methods: SAR-Opt-cGan (2018), DSen2-CR (2020) and GLF-CR (2022)
- by the results presented, the approach largely surpass in performance SAR-Opt-cGAN and DSense2-CR, both visualy and in objective metrics (SSIM, PSNR and MAE)
- also considering the results presented, the approach is a little superior in performance than GLF-CR
- Authors also conducted some ablation experiment, showing the advantage of using the proposed loss funcion
- As a general impression, the text is well-structured, well-written and clear for the reader
Author Response
Please see the attachment
Author Response File: 
Author Response.docx
Reviewer 2 Report
After carefully reading this paper，some major revisions should be performed.
1.Are there any considerations on their characteristics of optical and SAR imageries for the proposed network？For a better performance，the features should be involved，which is different from others.
2.For the limited samples，how is the performance of proposed network？
3.The relevant parameters may have impacts on the results. Please analyze the parameter relationship on them.
4.Turn our attention to the state-of-the-art methods，I suggest the authors implement some advance methods within 3 years.
5.The comparison results will haveshould be more riched.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 3 Report
The paper is interesting well prepared some additional experimental results are needed and motivation behind using the proposed architecture of CNN. Comments are the following:
1. Why did you use the U-shaped network or encoder-decoder design. In the literature this is the most used approach in denoising, ship detection, etc.
2. The IPP block using (1) and (2) is not described or referenced within the text.
3. Why did you use LTDS structure using down-sample and up-sample and then the LTL structure in the series again. Would not be better to have much more simple network to down-sample and up-sample the features and then additional extract features.
4. How did you determine number of down sampling blocks. Why did you use 4 of them?
5. What is motivation of using LTL and LeFF structures? They are widely used in denoising approaches. How this system affects on accuracy?
6. Details regarding datasets are missing. You used Sentinel 2 and 1 data. Images what is a resolution of data. Sentinel 1 provides resolution of 25 meters, images depicted have much lower resolution.
7. Time needed to train the network is missing.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
It can be accepted.
Reviewer 3 Report
Authors answered to all of my questions, paper can be accepted in present form.
