RS-DARTS: A Convolutional Neural Architecture Search for Remote Sensing Image Scene Classification
Round 1
Reviewer 1 Report
The authors of this manuscript have satisfactorily revised it. I believe it is now in a form worthy of publication.
Author Response
Thank you for your recognition of our work.
Reviewer 2 Report
Dear authors,
I have finished the review of your paper. In my opinion, you have addressed all my initial concerns and recommendations about your work.
Author Response
Thank you for your recognition of our work.
Reviewer 3 Report
This paper proposes a new paradigm was proposed that can automatially design a suitable CNN architecture for scene classification. A more efficient search framework, RS-DARTS, is adopted to find the optimal network architecture. The work is interesting, and the topic is suitable for publishing in Remote Sensing, but some problmes should be carefully solved before further proceedings. 1. The innovation or scientific contributions of this paper should be more clearly explained in the introduction part. What problem this paper aims to solve? What is the disadvantage of current methods? The current three contribution points seem too weak. 2. The symbols in all the equations should be more clearly expalined. 3. The data source of experimental datasets should be given in the paper. 4. How to determine the regularization parameters in the proposed RS-DARTS? 5. More relevant deep learning networks should be compared in the experiments. 6. The bias anaysis should be given in the experiments. 7. What is the limitations of proposed method?Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Round 2
Reviewer 3 Report
The authors carefully revised the paper, and I think the paper can be published after handling the following small issues:
- some relevant papers about remote sensing can be considered to review in the paper to guarantee a complete literature review.
a. R-CNN-based ship detection from high resolution remote sensing imagery
b.A Band Divide-and-Conquer Multispectral and Hyperspectral Image Fusion Method，2021
c. A simple and effective spectral-spatial method for mapping large-scale coastal wetlands using China ZY1-02D satellite hyperspectral images，2021
d. Low-Rank and Sparse Representation for Hyperspectral Image Processing: A Review，2021
2. The figure 2 can be improved for better illustration.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
This is a well-written manuscript. However, some figures seem out of focus or are of low resolution. Moreover, an English language proof-reading is required.
Reviewer 2 Report
Dear authors,
Your proposal seems to get competitive results in the state-of-art. However, I have some recommendations to improve your work.
- Please, provide some examples of which kind of images you are processing.
- Complement the description of training and test sets from the text of your paper, you have proposed a own organization from the benchmark datasets and it is paramount to have a good understanding of that.
- Analyze the impact of the hyper parameters selection with the best generated model.
- Highlight the best results in your tables.
I hope you find these recommendations useful.
Reviewer 3 Report
The paper requires extensive editing of English language and style required as it has a lot of grammatical error throughout the manuscript.
Reviewer 4 Report
This article presents a gradient descent method for designing a CNN architecture for classifying remote sensing scenes.
Specific Comments:
1. Figures appear out of format and are too small to be appreciated and analyzed.
2. All equations require punctuation marks at the end of their writing.
3. Line 239.- Clearly define the variables delta, L and E, as well as the use given to them in the RS-DARTS model.
4. Line 313.- How were the values ​​of the edge of the cell selected? What if the values ​​are further apart? Would the model stop working properly?
5. Equation 11 is mathematically misspelled. The product by (x << 1) is not clear. This must be corrected by the authors.
6. Some variables and symbols in the text must be written in italics.
7. What initial values ​​should epsilon and w have in steps 3 and 4 of algorithm 1?
8. Line 406.- How are the image data selected and eliminated in the established merging process? What geometric structures must be analyzed to carry out the elimination?
9. How were the Hyperparameters of Tables 4 and 5 selected?
