Machine Learning-Based Urban Renovation Design for Improving Wind Environment: A Case Study in Xi’an, China

Round 1

Reviewer 1 Report

The article is well-written but the authors need to provide the research question, aim or objective at the end of the introductory section.   

Comments for author File: Comments.pdf

Author Response

Responses to the Comments from the Reviewer 1

Comment 1: The article is well-written but the authors need to provide the research question, aim or objective at the end of the introductory section. 

Response 1:

First of all, we are grateful to the referee for the review. The comment is insightful and has a positive effect on improving the quality of our manuscript. Adding the research objective is an important modification in the revised manuscript. The modified paragraph at the end of the Introduction section is shown below. We highlight the new sentences in red.

“In this study, we explore a method for applying machine learning techniques to achieve urban renovation design. With the objective to improve the wind environment at the block scale, a group of urban fabric images are generated by the computer program. There are three fundamental steps. First, we apply region growing and k-means clustering to separately identify roads and buildings. To obtain accurate training instances, the urban fabric image is masked according to the land use map. Second, a multiple-point statistics program is performed to create new urban fabric images. Our program continuously reproduces the spatial structures in the target area. The similarity between training images and the variability between generated images are simultaneously increased. Third, we conduct Hausdorff distance and multidimensional scaling to examine the generated images. The distance between image patches becomes an indicator of the morphological consistency. We test the proposed method with an example from Xi’an, China. The building density and arrangement in the residential block are updated. A computational fluid dynamics program is launched to simulate the wind environment. The results imply that the proposed method significantly reduces the calm wind area. Rising wind speed becomes a key factor for facilitating air circulation and mitigating pollutant concentrations. There are two main objectives of this paper. (1) A machine learning-based framework is proposed to create new urban fabric images. Three basic components include image preprocessing, image generation and quality examination. (2) We fulfill an urban renovation design in Xi’an, China. The building density, building arrangement and open space are upgraded in the research area in order to solve the wind environment problem.”

 

Please see the attachment for more information.

Author Response File: Author Response.docx

Reviewer 2 Report

The methodology and results seem to be promising however in conclusions authors state:  The building  density and arrangement specified by our program play a key role in improving wind environment.

We can read in the text that 

we perform MPS to generate new urban fabric images. The computer  automatically  understands  the  spatial  relationship  between  roads,  buildings  and  open space. ...

In my opinion wider explanation is needed. Generating new fabric image is fine however there many ways of generating of such a design and urban renovation can not be designed only by the means for improving wind environment. Please add a statement reg. ways of design for urban renovation and indicate that authors' approach is focused on the wind aspect. Also my guess is that  former historical layout of  Xujiazhuang was different and this is important in urban renovation.

The middle school is at the top right

In my opinion such a consluion is pretty obvious and there is no need for use machine learning technique. The conclusions sections should be corrected. The aspect of computing urban mophology in 4 different scales should be highlighted, since that is the potential of machine learning, which in large scale can be problematic for non computional methods.

Note: Please do not put statements like 'the computer collects the urban fabric in the residential area' - in my opinion i t is definitely not the computer itself.. I assume that this is authors' program/ intention.

 

Author Response

Responses to the Comments from the Reviewer 2

Comment 1: The methodology and results seem to be promising however in conclusions authors state:  The building density and arrangement specified by our program play a key role in improving wind environment.

We can read in the text that 

We perform MPS to generate new urban fabric images. The computer automatically understands the spatial relationship between roads, buildings and open space. ...

In my opinion wider explanation is needed. Generating new fabric image is fine however there many ways of generating of such a design and urban renovation can not be designed only by the means for improving wind environment. Please add a statement reg. ways of design for urban renovation and indicate that authors' approach is focused on the wind aspect. Also my guess is that former historical layout of  Xujiazhuang was different and this is important in urban renovation.

 

Response 1:

Prior to the point-by-point response, we are grateful to the reviewer for his/her careful reading and valuable suggestions. Each comment plays a substantial role in improving our manuscript.

Based on the reviewer’s comment, the Conclusion section is rewritten. First, we explain that there are many ways to generate urban fabric images. The core contribution of the proposed method is discussed. Second, we emphasize that our target is to tackle the urban microclimate and wind environment problem with an urban renovation design. Third, the comparison between the building layout of Xujiazhuang and our design is highlighted. The modifications in the Conclusion section are shown below. The new content is described by red.

“In this study, we propose a machine learning-based framework to achieve an urban renovation design. There are two primary research objectives. On one hand, a three-step program is suggested to create new urban fabric images. Although there are many ways to draw fabric images, the core contribution of our method is to enable the computer program to automatically understand the spatial relationship between roads, buildings and open space. On the other hand, we achieve an urban renovation design in Xi’an, China. In particular, the urban microclimate and wind environment problems attract our attention. Aiming at speeding up air circulation, urban form is carefully upgraded in the research area.”

“Our method is tested by a residential block in Xujiazhuang, Xi’an, China. Based on the neighboring areas which are successfully modernized, the proposed program creates new urban fabric images. Compared with the historical layout in Xujiazhuang, the building density and arrangements are substantially improved in our design proposal. Computational fluid dynamics program is conducted to simulate air circulation in the research area. The result indicates the percentage of calm wind area is considerably reduced. Rising wind speed has a positive effect on air circulation and ventilation.”

 

 

 

 

Comment 2: The middle school is at the top right

 

Response 2:

We sincerely appreciate the reviewer for the careful reading. The modified sentence in Section 3.2 is shown in the following.

“It should be noted that the buildings in the top-right area belong to a middle school.”

 

 

 

Comment 3: In my opinion such a consluion is pretty obvious and there is no need for use machine learning technique. The conclusions sections should be corrected. The aspect of computing urban mophology in 4 different scales should be highlighted, since that is the potential of machine learning, which in large scale can be problematic for non computional methods.

 

Response 3:

The reviewer provides a constructive suggestion. In this paper, a set of machine learning techniques are applied to achieve an urban renovation design at the block scale. Moreover, there are four research scales investigating the effect of urban morphology on the microclimate. It would be interesting to further develop the machine learning technique in the field of the large-scale research.

Based on the suggestion, we add a new paragraph in the Conclusion section. The future directions are discussed. The relevant content is listed as follows.

“It is worth noting that the investigation in this work is an ongoing project. As mentioned in the introduction section, there are four research scales studying the effect of urban morphology on the wind environment. An applicable progression of this paper is to explore the urban renovation design at regional and urban scales. In the large-scale urban morphology study, there is a complicated relationship between human settlements and natural environment. A wide range of physical and social factors should be taken into account. Further research in the development of machine learning techniques would be of great help in understanding the spatial structure and characteristics of a metropolitan area, city, town and village.”

 

 

 

Comment 4: Note: Please do not put statements like 'the computer collects the urban fabric in the residential area' - in my opinion i t is definitely not the computer itself.. I assume that this is authors' program/ intention.

 

Response 4:

Thanks for this useful suggestion. In order to better explain the proposed method, we modify the sentences whose subject is the computer.

The related sentence in the Abstract is modified as follows.

Line 19: “Viewing the training image as a prior model, our program constantly reproduces morphological structures in the target area.”

 

In the Introduction section, we update the following content.

Line 110: “Given training images as the input, the computer program automatically outputs numerous realistic images.”

Line 114: “With the objective to improve the wind environment at the block scale, a group of urban fabric images are generated by the computer program.”

Line 119: “Our program continuously reproduces the spatial structures in the target area.”

 

There are many modified sentences in Section 2.2.

Line 190: “Though it does not impact the manual inspection, the noise brings difficulty to the computer program.”

Line 198: “First, the program selects one pixel in the original image as the seed point.”

Line 202: “If the difference is lower than a threshold, the program would add the neighboring point into a dataset.”

Line 205: “Third, we iteratively perform the neighbor checking step until the point dataset does not grow.”

Line 214: “Similar samples are allocated into one group.”

Line 220: “We repeatedly conduct steps (2) and (3) until the iteration number reaches a threshold.”

Line 229: “Based on the image masking technique, we retain the urban form in the residential areas.”

Line 257: “First, we visit an unknown point in the target area and collect informed points in the neighborhood.”

Line 261: “MPS creates a location vector .”

Line 263: “The program visits a point in the training image.”

Line 298: “Based on the similarity matrix, we carry out MDS to create a scatter plot.”

Line 332: “The spatial relationship between roads, buildings and open space is automatically understood by the computer program.”

Line 345: “The building arrangement is viewed as a prototype by the downstream procedures.”

 

 

We correct the relevant sentences in Section 3.

Line 377: “The program outputs the road network image using the region growing technique.”

Line 401: “Based on the semantics of land use property, we collect the urban fabric in the residential area.”

Line 408: “The proposed program automatically understands the spatial relationship between roads, buildings and open space.”

Line 424: “Moreover, our program can reserve the entrances, roads and square space in the city block.”

 

In Section 4, the improved content is displayed below.

Line 522: “The building density and arrangement are automatically determined by the computer program.”

 

Please see the attachment for more information.

Author Response File: Author Response.docx

Reviewer 3 Report

The authors proposed a machine learning-based framework to process the urban map and come up with a renovation design according to the analysis results. I have the following comments

1. In line 190, the author mentioned “the computer selects one pixel in the original image as the seed point”, how exactly the seed be selected?  Does the algorithm select the seed randomly? Or the seed is selected by following some rules?  If the algorithm randomly selects the seed, then how can you make sure the initial seed point belongs to the road?  

2. What is the difference used in line 191? Is it an RGB difference? 

3. The author has applied K-means to identify the buildings from the map. What is the K value in the paper? How did you decide the K value?

4. From figure 5(a), it seems that the road, open space, and building are marked with different colors in the urban fabric image. Instead of using the region growing algorithm and k-means, we can simply use this information to separate the road, open space, and buildings. Is there any benefit using those methods?

 

Author Response

Responses to the Comments from the Reviewer 3

Comment 1: The authors proposed a machine learning-based framework to process the urban map and come up with a renovation design according to the analysis results. I have the following comments

1. In line 190, the author mentioned “the computer selects one pixel in the original image as the seed point”, how exactly the seed be selected?  Does the algorithm select the seed randomly? Or the seed is selected by following some rules?  If the algorithm randomly selects the seed, then how can you make sure the initial seed point belongs to the road?

 

Response 1:

Before presenting our response to the comments, we would like to show our gratitude to the reviewer. Each comment is beneficial to improve our manuscript and motivate the reproducibility of the proposed work.

As the comment 1 states, the seed point is a key factor in the region growing program. Aiming at creating an accurate road network, we manually select the seed point. The statement is added in Section 2.2.1.

“First, the program selects one pixel in the original image as the seed point. To ensure the initial pixel belongs to the road, we manually specify the seed point in the urban fabric image.”

 

 

 

Comment 2: What is the difference used in line 191? Is it an RGB difference? 

 

Response 2:

Thanks for this helpful comment. The difference between two image pixels is an important concept in the region growing program. Since the original urban fabric image is grayscale, we define the difference as the Euclidean distance between two image pixels. The modified paragraph is shown as follows.

“Second, the difference between the seed point and a neighboring point is calculated. If the difference is lower than a threshold, the program would add the neighboring point into a dataset. Otherwise, the mismatching point is not collected. In our application, the original urban fabric image is grayscale. Therefore, the region growing program compares the intensities of two image pixels and adopts Euclidean distance to calculate the difference.”

 

 

 

Comment 3: The author has applied K-means to identify the buildings from the map. What is the K value in the paper? How did you decide the K value?

 

Response 3:

We appreciate the reviewer for providing this insightful comment. As a traditional machine learning method, k-means clustering focuses on dividing the instances into several groups. Examples with intensive similarity are allocated into one group. As a key parameter in the k-means clustering, the value of k represents the number of groups. In our method, k is configurated as 2 because our target is to distinguish buildings and open space. Accordingly, there are two pixel groups in the clustering result. In Section 2.2.1, we add the description of the k value.

“As the only one parameter in k-means clustering, the value of k represents the number of clusters. We configurate k=2 because the target of the clustering program is to distinguish buildings and open space. There are two pixel groups in the clustering result.”

 

 

 

Comment 4: From figure 5(a), it seems that the road, open space, and building are marked with different colors in the urban fabric image. Instead of using the region growing algorithm and k-means, we can simply use this information to separate the road, open space, and buildings. Is there any benefit using those methods?

 

Response 4:

The reviewer raises a valuable question. In the original urban fabric image, different colors are used to express roads, buildings and open space. It is intuitive to identify roads and buildings with color information. Therefore, the grayscale difference between two image pixels plays an important role in our program. Aiming at creating a high-quality result, we individually apply the region growing and k-means clustering to preprocess the urban fabric image. There are two key advantages. First, one key characteristic of the road network is its long-range connectivity. In order to find the connected component, region growing program pays attention to the similarity between one pixel and its neighboring pixels. Therefore, the road network associated with strong connectivity can be easily detected. By comparison, directly checking the color information is not a reasonable choice in this case. At the edge of buildings, many pixels have similar grayscale values to the road structure. A set of pixels would be misclassified by the color comparison program.

The second benefit of our method is that k-means clustering is able to automatically find the optimal threshold to partition buildings and open space. An adaptive program has a positive effect on yielding an objective result and saving computation time.

With the intention of better explaining the advantage of our method, we add a new paragraph in Section 2.2.1. The related content is shown below.

“In summary, the original urban fabric image contains many noise points and disturbances. The grayscale difference between the image pixels becomes a core factor in the preprocessing program. At first, we perform the region growing procedure to detect the road network. An important benefit is capturing the road structure with long-range connectivity. Next, k-means clustering is conducted to identify buildings and open space. Our program automatically finds the optimal threshold to distinguish image pixels. Finally, the program carries out the image masking to further improve the fabric image. With the intention of providing a high-quality image to the subsequent steps, we remove uncorrelated areas based on the land use map.”

 

Please see the attachment for more information

Author Response File: Author Response.docx