Block2vec: An Approach for Identifying Urban Functional Regions by Integrating Sentence Embedding Model and Points of Interest

Round 1

Reviewer 1 Report

The study aims to present the “Block2vec: An Approach for Identifying Urban Functional Re-gions by Integrating Sentence Embedding Model and Points of Interest.” The paper is well organized, and I would like to suggest minor comments before publishing.

1. The purpose of the study need to add to the abstract
2. The Introduction is well organized.
3. Figure 5 – X-axis unit is missing
4. Results, discussion, and conclusions are ok
5. Did you use any limitations? If so, explain them.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

I would like to thank the authors for this interesting work. This work has a good potential for urban planning, however, the content of the work has not been reported in an organized manner that makes it useful for general readers. A lot of steps are involved in defining urban functional regions but they do not converge to a single point to help readers understand how they have contributed to fulfilling the research objectives. Also, research objectives or the research questions need to be clearly outlined. Please find my detailed comments below:

Line 30: Clarification required on which information the authors are referring to

Line 31: It is important that the urban functional regions are clarified in the very first part of the introduction. Also, the use of “urban functional” and “urban function” terms needs to be consistent. Use either one of them for consistency.

Lines 37 to 38: Use better terms instead of “more and more”

Line 39: What is meant by urban spatial structure requires clarification. What spatial structures in urban areas are you referring to?

Line 46: Sentence appears to be incomplete

Lines 44-51: Requires further clarification. It appears that the authors are trying to say that the remote sensing approaches are often unable to capture the anthropogenic land use in a particular land cover category such as the impervious surface. The sentences do not reflect that.

Lines 56- 58: Meaning unclear. Consider directly saying that POI data are inexpensive and fast to acquire from the internet. Also, consider introducing the POI data in brief for non-technical readers.

Line 113: Why was 2016 selected as the study year?

Table 1: Requires better tabulation of information. The flow from left to right is a bit confusing.

Methodology: Each and every component of the workflow requires to be clearly outlined:

1. What information is comprised in POI semantic sequence?
2. How have you grouped the parcels and how were the parcel groups integrated with POI semantic sequence?
3. What is the structure of the latent semantic feature model?
4. As per figure 2, if you are generating training samples from the latent semantic feature of parcels and that is used to train the RF classifier, why was a separate result analysis for the RF models? Why wasn’t the region aggregation from the latent semantic feature of parcels the same as that from the RF models?

Line 142: A bit confusing here, isn’t the road located inside the land parcel? If yes, what is meant by “located closer to the road and other parts located in the parcel?”

Line 146: Spatial distance means Euclidean distance?

Line 153: “various lengths” – what lengths are you referring to?

Line 155: Is there any effect of removing excess POI on the feature/information extraction process?

Section 3.2 and 3.3: These two sections need to be revised. At present state, both these sections pose considerable difficulties in understanding how LSTM and RF had each served their purpose. It poses significant difficulties in understanding that the LSTM was used as a form of feature extraction process for the RF classifier. The entire LSTM structure need not to be explained in such details in the main body of the text and could be referred to as supplementary information. For example, the tuning and parametrization information of RF was provided in brief. Similarly, LSTM information could be condensed to remove information that, in general, are not necessary for readers. The most important part is explaining how these two ML algorithms converge to fulfill the study objectives.

Line 247: What do you mean by repeated 100 times? 100 iterations? And what “average values” are referred to, given RF was used for classification purposes?

Results: Even at this stage, it is unclear as to what “POI sentence length” entails. Also, the use of terms needs to be consistent. For example, how is “POI sentence length” different from “POI statement length”? If they are the same, why was a different terminology used here? For highly technical manuscripts such as this, it is essential to maintain uniformity/consistency.

Lines 263 to 271: Should be in the methods section.

Line 306: Where is class 2 after reclassification?

Lines 319 to 320: From where were these functions or categories derived? Doesn’t seem to be have been derived from the unsupervised method.

Lines 325: Please use the full terms for the first use.

Figure 8: (d) our method? Or the Block2vec method?

Figure 9: Consider using per class accuracy instead of the confusion matrix. Also, the confusion matrix does not establish the superiority of the proposed model. As a matter of fact, it demonstrates that different methods are suitable for different categories. It is still unclear what useful information this confusion matrix is providing to the readers.

4.2: The content of the section does not indicate that a sensitivity analysis was performed. Please revise

Discussion: The use of term “spatial interactions” should be used with caution. Interaction occurs both ways and from Figure 3, it appears that a unidirectional interaction was covered here. I could not find any neighborhood weight assignments to capture the spatial interactions between the POI or the parcels. Furthermore, interactions exist between POI from which the parcels were classified and then on the broader scale, the parcels influence each other. The study seemed to have captured only the interactions taking within a parcel for extracting information. This needs to be outlined clearly in the limitations.

Have tried to feed in the information from POI directly to an RF model without the extraction of latent semantics using LSTM?

Limitations not sufficiently discussed. For example, if it does conflict with standard urban land use classification, what problems does it pose from a policy-making perspective? Will this limit its use?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The authors have addressed 95% of the areas I have mentioned in the report. Figure 1, the left side map with admin divisions, labels are hard to read. Readers can not find where Wuhan is. I recommend this paper be suitable for publishing after another check for possible minor English errors based on overall improvements. If the editors expect, figure 1 can be improved a bit more.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

I accept the authors' answers to my questions.

The article has been much improved by the extension of subsection 4.1.2. 

I recommend the publication of the manuscript in its present form.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Thank you for the revised version and for addressing the comments. The revised version reads much better than the original version.