Semi-/Weakly-Supervised Semantic Segmentation Method and Its Application for Coastal Aquaculture Areas Based on Multi-Source Remote Sensing Images—Taking the Fujian Coastal Area (Mainly Sanduo) as an Example

Round 1

Reviewer 1 Report

I consider that the authors have correctly answered my observations and requests. Nothing else to add. 

Reviewer 2 Report

This paper proposed a novel semi-/weakly supervised method Semi-SSN, adopts 3 data sources: GaoFen-2 image, GaoFen-1(PMS) image and GanFen-1(WFV) image with 0.8m, 2m and 16m spatial resolution respectively.  I think this paper has been well revised based on my comments.  It is suitable for journal publication.

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

The paper title "Semi-/Weakly-Supervised Semantic Segmentation Method and
Its Application for Coastal Aquaculture Areas Based on MultiSources Remote Sensing Images ——Taking Fujian coastal area
(mainly Sanduo) as an example" provided sufficient information to reader and would be publish in this journal.

I have a minor concern.

GAN model is highly sensitive model in training and produced unstable behavior and sometime collapse the discriminative and generative loss function.

How do you overcome or tackle the situation when training is unstable?

GAN sometimes produced better results on one dataset and did not perform well on another dataset.

Did you try your proposed model in other image modality and you must provide an example on other dataset to show the better performance if possible?

Did you try other segmentation models and compare the results with GAN.

Reviewer 2 Report

This paper introduced  semi-/weakly-supervised sematic segmentation
method and its application for coastal aquaculture areas based on multi-sources remote sensing image. Authors used CGAN to solve some problems. I have the following comments:

• Authors mentioned "semi-/weakly-supervised semantic segmentation networks based on GAN." but just one paper has been introduced. As my check, there are many papers based on GAN. I think more survey is needed. In this aspect, authors should check on "https://www.springerprofessional.de/en/semi-supervised-semantic-segmentation-of-multiple-lumbosacral-st/17617040".  In this paper, already semi-CGAN has been introduced. I think very similar structure and concept.
• The designed method is reliable and well evaluated. In experiment results section, I think authors should take more experiments to compare the performance with open datasets for remote sensing. 
• It would be better to add some deep learning-based papers such as:

        . Deep BLSTM-GRU Model For Monthly Rainfall Prediction A case study of Simtokha, Bhutan, Remote Sensing (MDPI), vol. 2020, No. 12(19), pp.3174;

        . https://www.springerprofessional.de/en/semi-supervised-semantic-segmentation-of-multiple-lumbosacral-st/17617040 

     - For conclusion section, it should be re-written with the results and the performance of the proposed scheme. Also the future work should be addressed. 

• There are many typos. For example, "this paper proposed a novel semi-/weakly-supervised method SemiCGAN, adopted 3 data sources~" should be corrected as "this paper propose a novel semi-/weakly-supervised method SemiCGAN, adopted 3 data sources~".  In Abstract section, the current tense should be used. Authors should re-check all sentences very carefully.

Reviewer 3 Report

The paper proposes an effective new approach for semantic segmentation using CGAN framework. The application is very interesting and useful. The paper has substantial results and written nicely. I have no comments to review.

Reviewer 4 Report

After reading the work, my decission is "Accept after major revision (control missing in some experiments". Here are the main reasons for the before mentioned Overall Recommendation.

A. Does the introduction provide sufficiente background and include all relevant references? Yes

Nothing to report. The introduction provides enough background, including relevant references.

B. Is the research design appropiate? Yes

Nothing to report.

C. Are the methods adequately described? Can be improved

C.1. There are several flow chart figures that can be improved: 

- Figure 1. Is the discriminator trained with the real images? In that case, the "real images" has to have its arrow pointing to the discriminator. The figure depicts some king of operation with the "fake images" before reaching the discriminator.

- Figure 2 and 3. There are double head arrows. It cannot be followed the data flow without any ambiguity.

C.2. In addition to this, there are several places where the variables are not clearly explained. For example, in the subsection 2.1 the authors write G(z, theta_g) and D(x, theta_d) without explaining the meaning of these variables. There are other places where similar omissions take place. All the variables have to be described. A table containing all the variables could be included at the begining of the text in order to avoid these accidental omissions. 

C.3. Also, the algorithm 1 could be better written. A formal notation have to be used. Also, there are varaibles that could be better written. For example, {x1,...,xn} unlabelled imaged could be written using the "u" superscript. There is a problem of ambiguity. Several variable names are using for describing different objetcts.

C.4. There is no clear differentation between "semi-supervised" or "weakly-supervised". Please, include somekind of figure for explaining the differentes in the use of the training, validation and test sets in both scenarios.

C.5. Two different metrics are used Accuracy and IoU. The Accuracy is not described anywhere.

C.6. There are two different metrics used. Accuracy (Table 9) and mIoU (Table 10). Why?

C.7. This is one of the main flaws. It has to be seriously considered. The authors talk about the training, validation and test set. However, the data was only divided in  training and validation. The last one is "used to evaluate the robustness and generalization ability of the method". However, the tables describe the results on validation set. The authors have not used a test set. Is there any reason for this omission?

C.8. There is a mix between the methodology and results. The authors have to rework both sections and separate them. In the methodology, it is shown results of the different neural networks approachs. These results must be moved to the Results sections. Please, do not write any results in the Methodology section.

D. Are the results clearly presented? Yes

Nothing to report.

E. Are the conclusions supported by the results? Yes

Nothing to report.

F. In addition to all the recommendations written, the lines of the document are not numbered. This difficult the citations to differents part of the present work. In the next revisions of the work, the authors must number the lines.