A Multi-Objective Semantic Segmentation Algorithm Based on Improved U-Net Networks

Round 1

Reviewer 1 Report

The research content of the paper has a certain significance, but the overall structure of the paper is not reasonable:

1.        The introduction structure is not reasonable and fails to clarify the research background and current situation. The fourth paragraph of the introduction is not closely related to the research of this paper and is unnecessary. In the 52nd line of the introduction, the introduction of classification is not rigorous.

2.        When explaining the method of this paper, too much research status is included, which leads to confusion in reading.

3.        The experimental design is unreasonable, and the expression is inappropriate. The experiment did not clearly compare with other existing algorithms, including the DeepLabV3+ model mentioned in the paper. The description of the four models in the experiment is not clear, and it is recommended to use a table for comparison display. Why does the experiment show the results of SSTFM-1 and SSTFM-3?

4.        There are many grammatical issues in the paper that need to be modified, such as the 187th line, the 267th line, etc.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper presents a novel segmentation algorithm, based on U-Net architecture, for classification of urban land close to the construction sites of transport facilities. The research question is well motivated. The task and background of segmentation algorithms is nicely explained. Also, the methodology and results are adequately presented. However, I have noticed some quite notable shortcomings, listed below.

Major issues, which should be addressed before the final paper acceptance:

1)      The paper’s contribution is not clear. I have found several similar studies, e.g., [1] and [2]. The authors should thoroughly differentiate their problem from the similar ones, and state what are the key points why their problem should be solved separately. Alternatively, the authors should compare their results to the one of other studies and provide proof of an improvement.

2)      The authors should explain why the presented work is strictly related to the transportation facility construction sites. I would rather see this as a general land cover segmentation work.

3)      The “novel virtual data augmentation method” is announced as a contribution per se. However, the corresponding section focuses on the general theory of augmentation, rather than on the novel method. Furthermore, the mentioned method is not contrasted to the state-of-the-art, so its scientific merit is unknown. The authors should address these shortcomings or present the augmentation method only as a step of the whole segmentation algorithm.

4)      The technical aspects, such as the details about applied image preprocessing (L168-175), data annotation quality (L190-197), and annotators’ labor cost (L204-209), have been quite emphasized. Also, discussing hardware resources without any quantification (L198-203) has no scientific significance. IMHO, these parts deviate the reader from the main paper’s message by introducing unnecessary and subjective information.  

5)      Section 2.4 is almost entirely on the general theory of transfer learning, without presenting any contribution. Only last two sentences mention transfer learning usage in the presented study. It is indeed required for a paper to be self-contained, but the text should focus on the author’s work and achievements rather then on the general knowledge.

6)      Similarly, in Section 2.5, the whole text is about evaluation metrics, which are well known and are not the paper’s contribution.

7)      The selected values for (hyper-)parameters, given in Section 3.1, should be justified. For example, the authors can present the model accuracy while varying certain parameter values.

Minor shortcomings:

1)      L42-43 and L44: “high-risk risks” – Please avoid close word repetition.

 

Because of the listed weaknesses, I suggest a major revision for this manuscript.

 

References:

[1] Nurhanisah Hashim and Juazer Rizal Abdul Hamid: Multi-Level Image Segmentation for Urban Land-Cover Classifications. 2021 IOP Conf. Ser.: Earth Environ. Sci. 767 012024. DOI 10.1088/1755-1315/767/1/012024

[2] Lilay, M.Y., Taye, G.D. Semantic segmentation model for land cover classification from satellite images in Gambella National Park, Ethiopia. SN Appl. Sci. 5, 76 (2023). https://doi.org/10.1007/s42452-023-05280-4

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper, A Multi-objective Semantic Segmentation Algorithm Based on 2 Improved U-Net Networks is proposed. This paper is well written. There are some flaws in this paper.

1.     In section 2, the details of the algorithm should be given, for example, what type of “Filtering” should be given in Figure 1, and Figure 3 should also be introduced in more detail.

2.     In experimental section, The parameter settings of the proposed algorithm should be given as much as possible. This is conducive to the experimental recovery of readers.

3.     It is suggested to add classical semantic segmentation algorithm in the experimental part for comparison.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Dear Authors, 

Thank you kindly for addressing my remarks. Hope they helped you improve your valuable manuscript.

Reviewer 3 Report

The revised version is acceptable.