A Self-Supervised Learning Approach for Extracting China Physical Urban Boundaries Based on Multi-Source Data

Round 1

Reviewer 1 Report

In this manuscript, authors proposed a self-supervised learning approach for PUB extraction, which consists of three main steps, including data preprocessing, pretext task, and downstream task. China physical urban boundary data are produced by experiments which demonstrate the performance of the proposed method. The paper is organized well. Here are some issues that I hope the authors could considered before publication.

1) The abbreviations of NTL, OSM, ISA, etc., should appear with their full names at the first time.

2) How to ensure the accuracy of sampling generation? And if the points are mistakenly learned, I think it will mislead the Self-Supervised learning strategy.

3) How to evaluate the effectiveness of the CPUB product? I mean the product is compared with GUB, then why don’t we directly use GUB products.

4) The proposed method is comprised by a series of individual steps. It is interesting. But could you please provide the performance comparisons while each step is omitted or substituted by other methods?

5) In my opinion, it is better to compare with some state-of-the-art methods for some key steps of the processing.

6) Multi-source data are used as input, then it is necessary to compare the performance with single source data.

7) More references that use “self-supervised” strategy should be analyzed.

Minor revisions can be made for English Language.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

In this research, the authors proposed a novel method using a self-supervised learning approach for the extraction of CPUB. The topic is interesting however, some comments need to be clarified.

1.    The authors presented the use of nighttime and OpenStreetMap to initiate the urban boundary from the city center of each city. Nevertheless, why it has to be the city center? Will it be better if the authors use a central part of a predefined city as the initiating point? Or maybe the center of some criteria such as the point that has the highest population, GDP, and land price?

2.    The reviewer does not understand Line 126, What is the verification here, and how to quantify it?

3.    The results need to be verified more quantitatively. What is the benchmarking data? What is the performance of your system such as F1 score or Intersection over Union?

4.    On Line 350, if the authors give some information on GUB, it would be beneficial to the readers who have lower experience in this field to understand better.

5.    Why did the authors use CPUB of 2020 to evaluate with GUB of 2018? Will not the difference in years makes the evaluation poorer?

6.    On Line 377, how much is “slightly higher” and how do the authors quantify it?

7.    References about websites should also be written in the format of other current references, not in a blanket.

8.    Table 1 should be sorted from smaller years on the left to the higher years on the right.

The English quality is fine but further checking would also be appreciated.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The paper presents an interesting technique that extract physical urban boundaries (PUBs) of several Chinese cities. The paper is well organized and the results are interesting. However, some major issues emerge:

1. the aim of self-supervised learning is to obtain fruitful representations of the data to be used in some downstream tasks. In the proposed paper, the authors name “self-supervised learning” a procedure to obtain some labels to be used in a further processing step. This, in my opinion, is misleading. I strongly encourage the authors to rework all the passages that refers to self-supervised learning. On the other hand, it is still reasonable to refer to the task used to obtain labels as “pretext task”.

2. The procedure is interesting, but it is important to quantify the gain in performance w.r.t. the application of only the Step 1 to extract PUBs. In fact, one can say that if the gain is not considerable, two more steps are not worth. Please add some comparison.

Other issues:

1. I strongly encourage the authors to directly release the code and the dataset, in order to support open science.

2. Some passages should be reworked for clarity and/or better justified with some references (e.g. “annual impervious surface dataset in China” p. 3 line 123; Section 3.3; etc…)

Given these considerations, I think that the paper could be accepted only if all the raised issues are solved.

Minor editing needed

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have properly addressed my issues. It can be published in this form. Thank you.

Author Response

Thank you very much for your constructive comments. Those comments and suggestions have significantly improved our manuscript.

Reviewer 2 Report

The reviewer would like to thank you the authors for your very in-depth responses. Here are the further comments from the reviewer. The comments that the reviewer did not mention here mean that the authors have already explained or edited clearly.

Comment 2: So, the authors proposed the verification method in another separate reviewing paper. The reviewer thinks that it is pretty dangerous to do because if the reviewing paper is rejected by the reviewing journal, that means this paper will cite a non-existent paper, which is unacceptable. For the best practice, you should cite your paper once it is already published. If it has not been published at the time that our current manuscript will be accepted, you have to write what you want to publish in that reviewing paper. Thus, in this case, you should explain how can you quantify the accuracy. It may not be the entire process because the reviewer believes that it may be some novel research gap that your other manuscript will propose to the academic society. However, it should be enough for other readers to understand where this accuracy came from and calculated, without going further to read the cited paper.

Comment 3: There are some minor issues with this comment.

-        Line 410-412, “The middle part of the city is not sampled because that is very difficult to be 410 misidentified. However, the fringe parts of the city are areas that can be easily 411 misidentified.”, needs to be modified. It is very hard to read.

-        The authors may consider providing a small explanation of the Producer’s accuracy, User’s accuracy, Overall accuracy, and how to quantify them for lower expertise readers.

-        Furthermore, because Table 1 has a lot of abbreviations that will be explained on Line 450++, Table 1 should not come before Line 450. Readers will be confused about what are UA, PA, and OA.

-        Moreover, Table 1 looks very difficult to understand. The authors may separate the value inside into two tables, one is the confusion matrix, and the other is the comparison of PA, OA, UA, and F1 scores between CPUB and GUB.

The reviewer believes that after this revision, this manuscript will be ready to be published. Congratulations to all of the authors.

Some minor editing would be required such as what mentioned in the comments.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

I think the authors gave interesting answers to the raised issues. Two only issues in my opinion remain:

1) as already stated, "self-supervised learning" is misleading. Some similarities with the proposed work are clear, but the framework is different. I encourage the authors to change the title.

2) I would add all the interesting insights, that the authors gave me in the rebuttal letter, directly in the paper (at least in the Appendix). This can help people who is having my same doubts.

A swift proofreading will solve minor issues

Author Response

Please see the attachment

Author Response File: Author Response.docx