Semi-Supervised Remote Sensing Image Semantic Segmentation via Consistency Regularization and Average Update of Pseudo-Label

Round 1

Reviewer 1 Report

Upon checking, the authors have revised the article based on almost all the comments given in the initial submission. 

Author Response

    Thank you very much for your guidance and help in our paper. 

    We also learned a lot of knowledge when revising the paper, and had a better understanding of the writing method.

Reviewer 2 Report

The proposed reserch "Semi-Supervised Remote Sensing Image Semantic Segmentation via Consistency Regularization and Average Update of Pseudo-label" seems a really interesting idea. The methods proposed are based on Consistency Regularization training for semi-supervised training. Later on, make use of the new learned model for Average Update of Pseudo-label, and finally, combine pseudo labels and strong labels with training semantic segmentation network.

Some aspects need to be addressed before further processing of the article:
The proposed methods make use of UNet, DeepLabV3 and DeepLabV3+ CNN models. Table 1, 2, and 3 present the MIoU scores of the proposed experiments. Apparently, the proposed methods seem effective and achieve better performance in comparison to base models. There is a considerable improvement in results when CR and AUP are used along with the base model. Though in this type of study, we also assess the proposed method's performance in terms of trainable parameters (time and space). This article didn't show any analysis or comparison addressing this aspect. I would suggest presenting an analysis of how much accuracy is achieved at the expense of how much trainable parameters. The proposed methods add several additional parameters to the running system. Please show the base and proposed model' parameters analysis through table or charts to better show the effectiveness of your results. 

 

Author Response

    Thank you very much for your guidance and suggestions.

    We have replied to the questions raised one by one and made appropriate modifications in the paper.

    The specific content is in the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The topic is interesting. Even though the article is interesting in its current format, some aspects should be improved for possible publication and for a better understanding by the readers. Comments formulated during my review are presented below. These are as follows:

1) As the image segmentation is the significant technique in this paper, the authors should briefly provide several groups of image segmentation methods, namely:

*superpixel segmentation methods [a,b]
*watershed segmentation methods [c,d]
*level set segmentation methods [e,f]

[a] "Superpixel Segmentation of Polarimetric Synthetic Aperture Radar (SAR) Images Based on Generalized Mean Shift". Remote Sensing, 2018, vol 10(10), 1592.
[b] "Superpixels: An evaluation of the state-of-the-art". Comput. Vis. Image Underst. 2018, vol. 166, 1-27.
[c] "River channel segmentation in polarimetric SAR images: watershed transform combined with average contrast maximisation." Expert Systems with Applications, 2017, vol. 82, 196-215.
[d] "Watershed cuts: thinnings, shortest path forests, and topological watersheds", IEEE Trans. Pattern Anal.Mach. Intell. 2010, vol. 32, 925–939.
[e] "A Median regularized level set for hierarchical segmentation of SAR images". IEEE Geoscience and Remote Sensing Letters, 2017, vol. 14(7), 1171-1175.
[f] "Level Set Segmentation Algorithm for High-Resolution Polarimetric SAR Images Based on a Heterogeneous Clutter Model."IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2017, vol. 10(10), 4565-4579.

The above segmentation methods offer some alternatives to machine learning based segmentation.

2) In the related work section, a more rigorous investigation on the existing methods, such as comparison of previous approaches in terms of pros and cons, should be given. A summary table can be used in this regard.

3) In a separate paragraph it is required to provide some including remarks to further discuss the proposed methods, for example, what are the main advantages and limitations in comparison with existing methods?

4) Please give a frank account of the strengths and weaknesses of the proposed research method.

5) The authors need to present and discuss several solid future research directions.

Author Response

     Thank you very much for your guidance and suggestions.

    We have replied to the questions raised one by one and made appropriate modifications in the paper.

    The specific content is in the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

The Authors have addressed all the comments.

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

Dear authors,

The manuscript has many positive aspects however it needs improvement in many areas for further processing.

Areas to improve:

1. lack of state-of-the-art literature review - please show a depth of review and then formulate research questions to address the knowledge gap. The literature on theory and application should be widely reviewed.
2. The experimental design is limited to certain image data sets given the semantic segmentation is a foundation of image processing, I advise to select the data set that capture the varieties of images and can be considered as sensible representations. 
3.  the discussion section needs further elaboration and contrast with previous literature in the field.
4. Please rewrite the sections and sub-sections considering the logical flow.

Regards

Author Response

        Thank you very much for your suggestions. We provide a point-by-point response and make appropriate modifications to our paper.       We try our best to explain your questions and then make corrections to our problems.         Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper presents a semi-supervised segmentation method using consistency regularization and average update of pseudo-label features. This method only requires a comparably smaller amount of labeled data and combines with unlabeled data to improve segmentation performance. This paper also demonstrates the effectiveness of the proposed method on different remote sensing image datasets. This paper is well written, but there are some comments for improvements:

1. When the authors summarize the contribution of this paper in line 52, they claim that they propose an ‘improved’ consistency regularization training method. It is unclear what they have improved in consistency regularization. In my opinion, they have ‘modified’ some pre-processing steps so that the input is suitable for the semi-supervised training purpose.
2. The authors mentioned in line 73 that they cannot translate and crop the input image. Please clarify it by telling us the reasons.
3. The first sentence in Section 3 or line 109 is unclear.
4. The follow-up question is that, if the authors classify each pixel, we would like to know how efficient the method is.
5. Please describe the random noise model used for generating the perturbed input.
6. Please check the subscript and superscript of the symbols, e.g., ux1 and ux2 in the paragraph before Equation (2), and 1e-4 in line147.
7. Please add a space before K in Equation (5).
8. Please use ‘x’ instead of ‘*’ and include the unit when describing the resolution of the image, e.g., 5000 x 5000 pixels.

Author Response

            Thank you very much for your suggestions. We provide a point-by-point response and make appropriate modifications to our paper.       We try our best to explain your questions and then make corrections to our mistakes.         Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear authors,

The literature you have cited majorly from conference papers are not relevant to remote sensing science. 

The logical flow to the contribution to knowledge is not justified. Unfortunately, the research outcome of this manuscript is not providing a compelling contribution to knowledge to the semantic segmentation within the remote sensing science/application. In this ground, perhaps it should be submitted to a more algorithm focused journal.

I recommend resubmitting to other relevant journals.

Regards,