How Urban Morphology Relates to the Urban Heat Island Effect: A Multi-Indicator Study

Round 1

Reviewer 1 Report

Thank you for extending the invitation to review the manuscript titled "How Does Urban Morphology Relate to the Urban Heat Island: A Multi-indicator Study" by Liu et al. I have carefully examined the manuscript and find the motivation and methodology to be satisfactory. The manuscript presents insightful findings that can be of interest to scholars in the field of urban studies. The authors have employed numerous 3D and 2D urban morphological variables, combined with remotely-sensed data products, and have implemented the XGBoost regressor and the SHAP (SHapley Additive exPlanations) method for interpreting the relationship between these indicators and the occurrence of the urban heat island in different seasons. The manuscript is well-written, but I have some comments that could enhance its quality:

1) Lines 86-94: The introduction should provide a more substantial emphasis on the use of the SHAP method, which is one of the widely recognized techniques in explainable artificial intelligence (XAI). Since the application of XAI techniques is not particularly common in this domain, it requires a more detailed explanation. Although your sub-section 2.3.2 is well-written, the paper lacks a clear rationale for employing SHAP in urban studies. I recommend consulting existing XAI-based applications in urban studies, such as [1, 2, 3].

2) Lines 238-243: The information provided here and in Table 3 appears to belong to the Materials section.

3) Table 5: This table requires further clarification. Firstly, it is unclear why the variables were grouped differently compared to the correlation table in Table 4, where the coefficients for each variable were presented individually. Secondly, it is unclear why R2 scores were chosen to explain the variance of UHII, considering the use of a decision tree-based regressor, which is inherently explainable through SHAP or permutation feature importance. While using the R2 score as a regression performance metric is appropriate, explaining the regressor using the R2 score may lead to confusion.

4) I have concerns about the SHAP dependence plots in Figures 5, 6, and 8. The approach of annotating the plots with text to show three distinct categories based on the values of FAR, SVI, and NDVI lacks a clear rationale. It is important to explain how you determined the intervals for these categories based solely on the values. If you can provide a scientific explanation for selecting these intervals, it would be beneficial. Alternatively, if you are unable to justify the intervals, I recommend plotting the SHAP values without interpreting them into categories and reorganizing the presentation of results accordingly. For instance, following the approach used for HIGH and DBS in Figure 7, where only the values are considered. I hope this clarifies my point.

5) It would be valuable to include information about cloud cover and snow cover in the Landsat 8 imagery, as these factors are crucial for understanding LST and NDVI. Please incorporate these details into the text.

[1] https://doi.org/10.1016/j.habitatint.2022.102660

[2] https://doi.org/10.3390/app12189169

[3] https://doi.org/10.1016/j.scs.2023.104443

Author Response

General comments:

R: Thank you for taking the time to read our manuscript "The relationship between urban morphology and urban heat island: a multi-indicator study". We appreciate your valuable feedback and the time and effort you put into providing us with insights and suggestions. We have revised the manuscript based on your suggestions. Therefore, we expanded the explanation of the SHAP method in the Introduction section (Note 1), moved the data preprocessing to the Materials section (Note 2), added more explanation between Tables 4 and 5 (Note 3), and enhanced Figures 5, 6 and 8 (Note 4); finally, we supplemented the information on cloud cover and snow cover in the Landsat 8 satellite images (Comment 5).

Below are point-by-point responses to each comment. We will be happy if you are satisfied with the revised manuscript.

1. Lines 86-94: The Introduction should more substantively emphasize the use of the SHAP method, one of the widely recognized techniques in Explainable Artificial Intelligence (XAI). Since the application of XAI techniques is not particularly common in this field, a more detailed explanation is required. While your subsection 2.3.2 is well written, the paper lacks a clear justification for using SHAP in urban studies. I recommend looking up existing XAI-based applications in urban research, such as [1, 2, 3].

R Thanks for pointing this out. We have provided more details to explain this issue in the revised manuscript (pages 2-3).

Although machine learning models have proven effective in classification and regression, understanding the interrelationships between these factors and these "black box" models remains challenging. (page 2, lines 93-95, section 1)

The SHAP model is one of the methods based on the Explainable Artificial Intelligence (XAI) framework, which aims to improve the interpretability and explainability of AI models. (page 3, lines 99-101, section 1)

Thanks again for the valuable references, which we have included in the revised references [1-5].

2. Lines 238-243: The information presented here and in Table 3 appears to belong in the Materials section.

R We fully agree with your suggestion. We moved the revised lines 238–243 and the data preprocessing of Table 3 to subsection 2.1.1 of the Materials section (page 4 lines 140–148, section 2.1).

3. Table 5: This table needs further clarification. First, it is not clear why the variables are grouped differently compared to the correlation table in Table 4, where the coefficients for each variable are shown individually. Second, it is unclear why the R2 score was chosen to explain the variance in UHII, which could essentially be explained by SHAP or permutation feature importance, given the use of decision tree-based regressors. While it is appropriate to use R2 scores as regression performance metrics, interpreting regressors using R2 scores can lead to confusion.

R Thank you very much for pointing out these issues.

We mainly want to show the significant relationship between the independent variable and the dependent variable through the correlation analysis performed in Table 4. This is because correlation analysis is usually performed before performing regression analysis. In other words, Table 4 focuses on the interpretability of individual indicators, so these factors are ranked individually. To avoid ambiguity, we added 'The purpose of this analysis was to identify potential correlations between UHII and multiple indicators.' (page 9, lines 304-305, section 3.2)

The five groups of independent variables were analyzed separately using the XGBoost regression model in Table 5. These five groups of indicators divide the above indicators into two-dimensional building indicators, three-dimensional building indicators, all building indicators, ecological infrastructure indicators and all multi-indicators. With them, we tried to examine to what extent each type of indicator is crucial for explaining UHI.

As for the regression performance metric, due to its metric-specific nature, it is not feasible to compare feature importance across different groups. We chose explained variance (R2) to more clearly explain differences in performance between model groups. This statement can be found in the revised manuscript (page 10, lines 327-330, section 3.3).

We apologize for the ambiguity in Table 5 and have made the following changes: “These five sets of independent variables were analyzed separately using an XGBoost regression model to examine the relative prominence of each indicator in a particular season. The above explanations are included in the text Already computed (page 10, lines 322-324, Section 3.3).

4. I am concerned about the SHAP dependency graphs in Figures 5, 6 and 8. The approach of annotating the plot with text to show three different categories according to the values of FAR, SVI and NDVI lacks a clear justification. Be sure to explain how the intervals for these categories are determined based on the values alone. It would be helpful if you could provide a scientific explanation for choosing these intervals. Alternatively, if you can't justify the intervals, I recommend plotting the SHAP values without interpreting them as categories and reorganizing the representation of the results accordingly. For example, follow the approach for HIGH and DBS in Figure 7, where only values are considered. I hope this clarifies my point.

R Your suggestions have been very helpful in improving the manuscript. In fact, our aim is to pick three representative points (from low to high) in the scatterplot to show the local urban structure with aerial photos accordingly. We apologize for the misleading information in the previous episode, a revised version can be found in the attachment.

5. It is valuable to include information about cloud cover and snow cover in Landsat 8 imagery, as these factors are critical to understanding LST and NDVI. Please include these details in the text.

R Your suggestions are very valuable to us. Please find the revised section in the article (page 4):

  The selection of Landsat 8 satellite images was based on choosing four images with cloud cover less than 10% during the period from December 2017 to October 2018. (Page 4 Line 143-145 Section 2.1).

Thank you again for your patience and insightful comments for improving this paper.

References

[R1] Guidotti, R.; Monreale, A.; Ruggieri, S.; Turini, F.; Giannotti, F.; Pedreschi, D., A survey of methods for explaining black box models. ACM computing surveys (CSUR) 2018 , 51, (5), 1-42.

[R2] Gao, Y.; Zhao, J.; Han, L., Quantifying the nonlinear relationship between block morphology and the surrounding thermal environment using random forest method. Sustainable Cities and Society 2023, 91, 104443.

[R3] Kim, M.; Kim, G., Modeling and Predicting Urban Expansion in South Korea Using Explainable Artificial Intelligence (XAI) Model. Applied Sciences 2022, 12, (18), 9169.

[R4] Antoniadi, A. M.; Du, Y.; Guendouz, Y.; Wei, L.; Mazo, C.; Becker, B. A.; Mooney, C., Current challenges and future opportunities for XAI in machine learning-based clinical decision support systems: a systematic review. Applied Sciences 2021, 11, (11), 5088.

[R5] Iban, M. C., An explainable model for the mass appreciation of residences: The application of tree-based Machine Learning algorithms and interpretation of value determinants. Habitat International 2022, 128, 102660.

Author Response File: Author Response.pdf

Reviewer 2 Report

Initially, I would like to congratulate the authors for the excellent work submitted to the journal. The importance of discussing the discussed topic is of unique relevance to environmental science as a whole. However, the work is lacking in some aspects, which end up weakening it too much.

·        The research problem is poorly defined in the introduction. What is the research problem? This should be a clear question in the text. The research objectives are clear.

·        Lines 139-141: Are the previous surveys from the same group of researchers? If so, they are part of a larger research project on this type of investigation. If applicable, mention these projects more in the writing of the article.

·        Enlarge figures 1 to 9 to the full width of the page, this is fine. It will make it easier for readers to see. Even if the size of the final article gets bigger.

·        Insert spaces between the titles and subtitles of the article stages, so the work is more visually presentable.

·        Item 4 of the manuscript lacks bibliographical references, as this is the part of the research that discusses the results with the bibliography. Expand the review of articles for discussion in this item. Suggestions must be inserted in the final considerations / conclusions of the article.

·        In the conclusions, include a paragraph that mentions the main difficulties of the research, and which new directions the data found can follow in future research.

Research of a complex nature, all suggestions made here are to improve the work. Congratulations to the authors.

Author Response

General comments:

 

The authors sincerely appreciate your constructive feedback and valuable comments, and we are committed to addressing these concerns in our revised manuscript.

Accordingly, we further clarified the research problem in the Introduction section, added more information for background field calculation in the Methodology section, and included future works in the Conclusion section (Comment 1, 2 &6); Format polishing has been conducted to the titles and figures (Comment 3&4); In the Discussion section, more viewpoints from experts have been referred to enhance ^ (Comment 5).

In response to your assessment, we have outlined the specific actions we intend to take:

 

1. The research problem is poorly defined in the introduction. What is the research problem? This should be a clear question in the text. The research objectives are clear.

R. Thanks for pointing out this problem. After revision, we have clarified our research problem:

 As part of this work, the following research questions are addressed:1) How do different types of indicators impact on UHI? 2) Are the effects of indicators on UHI consistent across different seasons? (Page 3 Line 115-117, Section 1)

 

2. Lines 139-141: Are the previous surveys from the same group of researchers? If so, they are part of a larger research project on this type of investigation. If applicable, mention these projects more in the writing of the article.

R. We apologize for the ambiguity. We have fixed this issue and clarified as:

 In this paper, we used the method proposed by Li to detect urban growth and extract complete urban region (Figure 1). This method utilizes two valuable remote sensing data sources: nighttime light data and impervious surface data. The nighttime light data was obtained from the NCEI National Centers for Environmental Information, while the impervious surface data was acquired from Liu's team. The area outside the urban boundary with the same area as the urban area was determined as the urban background temperature field, and the average land surface temperature within this area was regarded as the urban background temperature. (Page 5 Line 167-175, Section 2.2)

 Accordingly, we added references [1-3] introducing background field calculation method. The modified Figure 1 could be found in the attachment.

 

3. Enlarge figures 1 to 9 to the full width of the page, this is fine. It will make it easier for readers to see. Even if the size of the final article gets bigger.

R. Many thanks. Following your suggestion, we have increased the width of the images. In addition, we slightly modified these images to make the results easier to comprehend. We pick the revised version to Figure 5 as a representative. Please see the attachment.

 

4. Insert spaces between the titles and subtitles of the article stages, so the work is more visually presentable.

R. We have revised the spacing between the titles, taking into consideration the submission guidelines as well as your feedback.

 

5. Item 4 of the manuscript lacks bibliographical references, as this is the part of the research that discusses the results with the bibliography. Expand the review of articles for discussion in this item. Suggestions must be inserted in the final considerations / conclusions of the article.

R. Thank you very much for your suggestions. In the Discussion section, we have incorporated additional viewpoints from literature and included more relevant references [4-8] to support our arguments:

 Different shapes of buildings primarily affect the UHI effect through their impact on ventilation [4]. A homogeneous cluster of buildings of a single type may have better ventilation effects, thus mitigating the UHI [5]. (Page 15 Line 488-491, Section 4.1.1)

 To clarify, the evaluation of the cooling impact of vegetation can be conducted through spatial aggregation [6]. (Page 16 Line 532-533, Section 4.1.2)

 Hence, in densely populated urban areas where there is a scarcity of vegetation cover that is also fragmented, the most efficacious approach to mitigate the UHI effect would be to plant rapidly growing tall trees [7]. (Page 17 Line 550-553, Section 4.1.2)

 Initially, erecting high-rise structures amidst verdant landscapes and aquatic envi-ronments is a proficient tactic for alleviating the UHI effect[8]. (Page 17-18 Line 563-564, Section 4.2)

 

6. In the conclusions, include a paragraph that mentions the main difficulties of the research, and which new directions the data found can follow in future research.

R. Thanks for pointing out these important issues. After revision, we have included our future works (Page 18) as:

 Due to the limitations of the study scope and data availability, there are several unexplored directions in urban morphology research that can enhance our understanding of the UHI effect. These factors include different vegetation types, biomass, as well as factors such as precipitation, wind direction, and wind speed that influence the daily UHI. (Page 18 Line 611-615, Section 5)

 

Once again, the authors would like to gratefully thank the reviewer for his/her insightful comments and recommendations for improving the paper.

 

 References

[R1] Li, K.; Chen, Y.; Wang, M.; Gong, A., Spatial-temporal variations of surface urban heat island intensity induced by different definitions of rural extents in China. Science of the total environment 2019, 669, 229-247.

[R2] Zhang, X.; Liu, L.; Wu, C.; Chen, X.; Gao, Y.; Xie, S.; Zhang, B., Development of a global 30 m impervious surface map using multisource and multitemporal remote sensing datasets with the Google Earth Engine platform. Earth System Science Data 2020, 12, (3), 1625-1648.

[R3] Peng, S.; Piao, S.; Ciais, P.; Friedlingstein, P.; Ottle, C.; Bréon, F.-M.; Nan, H.; Zhou, L.; Myneni, R. B., Surface urban heat island across 419 global big cities. Environmental science & technology 2012, 46, (2), 696-703.

[R4] Gao, Y.; Yao, R.; Li, B.; Turkbeyler, E.; Luo, Q.; Short, A., Field studies on the effect of built forms on urban wind environments. Renewable Energy 2012, 46, 148-154.

[R5] Jiang, Y.; Wu, C.; Teng, M., Impact of residential building layouts on microclimate in a high temperature and high humidity region. Sustainability 2020, 12, (3), 1046.

[R6] Armson, D.; Stringer, P.; Ennos, A., The effect of tree shade and grass on surface and globe temperatures in an urban area. Urban Forestry & Urban Greening 2012, 11, (3), 245-255.

[R7] Tan, J. K.; Belcher, R. N.; Tan, H. T.; Menz, S.; Schroepfer, T., The urban heat island mitigation potential of vegetation depends on local surface type and shade. Urban Forestry & Urban Greening 2021, 62, 127128.

[R8] Deilami, K.; Kamruzzaman, M.; Liu, Y., Urban heat island effect: A systematic review of spatio-temporal factors, data, methods, and mitigation measures. International journal of applied earth observation and geoinformation 2018, 67, 30-42.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper “How Does Urban Morphology Relate to the Urban Heat Island: A Multi-indicator Study” explores in an empirical manner the relationship between urban morphology based on quantitative indicators and the intensity of the urban heat island in central Beijing (China). The research topic is very significant bearing in mind the growth of cities and urban population at the global level and the consequent modification of environmental conditions. In this sense, the study seems affirmative with potentially practical outcomes and implications. Anyway, several things are completely unclear and further clarifications are needed.

Many explanations are missing and there are many ambiguities in the methodological part. Authors should explain the used urban morphology indicators in detail (especially 3d indicators), how they were calculated, what was used as a data source, and the obtained scale of indicator values as well as the spatio-temporal resolution of the used datasets (Table 1 and 3 should be in the same section with all important information). Spatial resolutions should be discussed additionally. Although the authors refer to the literature when calculating the UHII, the concept, and calculations need more explanation (e.g., clarify suburban/rural, background field area, etc.). In the present work, the Machine Learning Models are discussed only in a technical sense not related to the “nature” of researched problem. Also, section 3.3 should be explained in more detail, especially the “categorization of indicators into five groups of the independent variable”. Figures 1-4 are of poor quality and visibility. Consider putting just one scale (with colors) in the legend of Figure 4.

A major revision is recommended.

Author Response

General comments:

 

We would like to express our gratitude for your time and effort in reviewing our work, which obviously improves the quality of this research. We have revised the manuscript carefully by taking into account your concerns. We have provided more details for the calculation of urban morphology indicators (Comment 1) and the background field (Comment 2). Additionally, we have supplemented the rationale for why a machine learning model is more suitable for our experiment (Comment 3). In section 3.3, we have provided a more detailed explanation of the grouping of indicators (Comment 4). Lastly, we have improved the quality of the figures in the article to ensure they are clearer and more visible (Comment 5).

A point-by-point response to each comment is given below. We would be happy if you are satisfied with the revised manuscript.

 

1. Authors should explain the used urban morphology indicators in detail (especially 3d indicators), how they were calculated, what was used as a data source, and the obtained scale of indicator values as well as the spatio-temporal resolution of the used datasets (Table 1 and 3 should be in the same section with all important information). Spatial resolutions should be discussed additionally.

R The authors greatly thank this reviewer for his/her positive comments. We provide more details in the revised manuscript (Page 3):

 The building form indicators are all derived from building vector data, including diversity of building shapes (DBS), floor area ratio (FAR), sky view factor (SVF), and building height (HIGH). These indicators normalized difference built-up index (NDBI), modified normalized difference water index (MNDWI), and normalized difference vegetation index (NDVI) are calculated using Landsat 8 satellite imagery data. The population density (PD) indicator is derived from population raster data. (Page 3 Line 128-134, Section 2.1)

 

2. Although the authors refer to the literature when calculating the UHII, the concept, and calculations need more explanation (e.g., clarify suburban/rural, background field area, etc.).

R Thanks for pointing out these important issues. We made a more detailed explanation to UHII, clarifying concepts of urban area and urban background temperature field. We have fixed this issue and clarified as:

 In this paper, we used the method proposed by Li to detect urban growth and extract complete urban region (Figure 1). This method utilizes two valuable remote sensing data sources: nighttime light data and impervious surface data. The nighttime light data was obtained from the NCEI National Centers for Environmental Information, while the impervious surface data was acquired from Liu's team. The area outside the urban boundary with the same area as the urban area was determined as the urban background temperature field, and the average land surface temperature within this area was regarded as the urban background temperature. (Page 5 Line 167-175, Section 2.2)

 The modified figure could be saw in the attachment.

 

 Accordingly, we added references [1-3] introducing background field calculation method.

 

3. In the present work, the Machine Learning Models are discussed only in a technical sense not related to the “nature” of researched problem.

R We greatly thank the reviewer for this suggestion. We provide more discussions on the ML models with our research problem in the revised manuscript (Page 2):

 Traditional regression analysis models, such as Multiple Linear Regression and Polynomial Regression, are based on pre-defined linear or nonlinear relationships and still have advantages in certain specific scenarios. However, their ability to handle complex relationships is limited. On the other hand, machine learning models have the capability to automatically learn complex relationships from data, enabling them to better fit complex real-world problems and nonlinear relationships. Random Forests, XGBoost, Support Vector Machines, and Artificial Neural Networks are widely used models in current research. (Page 2 Line 86-93, Section 1)

 In this part, more references [4-7] have been included accordingly.

 

4. Also, section 3.3 should be explained in more detail, especially the “categorization of indicators into five groups of the independent variable”.

R Thank you for pointing out this problem.

 Five sets of independent variables were analyzed separately using XGBoost regression models in Table 5. These five sets group the indicators into 2D building indicators, 3D building indicators, all building indicators, ecological infrastructure indicators, and all multi-indicator. We try to examine to what extent does each type of indicators essential to the explanation of UHI.

 According to your suggestion, we provide more details in the revised manuscript (Page 10) to explain this issue.

 ‘These five sets of independent variables were analyzed separately using XGBoost regression models to examine the relative prominence of each indicator during specific seasons.’ (Page 10 Line 322-324, Section 3.3).

 ‘As for the regression performance metric, it is not feasible to compare feature im-portance across different groups due to its metric-specific nature. We choose explained variance (R2) to offer a clearer explanation for performance variations among model groups.’ (Page 10 Line 327-330, Section 3.3).

 

5. Figures 1-4 are of poor quality and visibility. Consider putting just one scale (with colors) in the legend of Figure 4.

R Many thanks. Following your suggestion, we have increased the width of the images in the article and replaced them with EMF files. The scale bar has also been modified according to your comment. We pick the revised version to Figure 4 as a representative. Please see the attachment.

 

 

Once again, thank you for your patience and insightful comments for improving this paper.

 

 References

[R1] Li, K.; Chen, Y.; Wang, M.; Gong, A., Spatial-temporal variations of surface urban heat island intensity induced by different definitions of rural extents in China. Science of the total environment 2019, 669, 229-247.

[R2] Zhang, X.; Liu, L.; Wu, C.; Chen, X.; Gao, Y.; Xie, S.; Zhang, B., Development of a global 30 m impervious surface map using multisource and multitemporal remote sensing datasets with the Google Earth Engine platform. Earth System Science Data 2020, 12, (3), 1625-1648.

[R3] Peng, S.; Piao, S.; Ciais, P.; Friedlingstein, P.; Ottle, C.; Bréon, F.-M.; Nan, H.; Zhou, L.; Myneni, R. B., Surface urban heat island across 419 global big cities. Environmental science & technology 2012, 46, (2), 696-703.

[R4] Uyanık, G. K.; Güler, N., A study on multiple linear regression analysis. Procedia-Social and Behavioral Sciences 2013, 106, 234-240.

[R5] Ostertagová, E., Modelling using polynomial regression. Procedia Engineering 2012, 48, 500-506.

[R6] Mahesh, B., Machine learning algorithms-a review. International Journal of Science and Research (IJSR).[Internet] 2020, 9, 381-386.

[R7] Molnar, C., Interpretable machine learning. Lulu. com: 2020.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors made the corrections and suggestions given to improve the article. The presentation of the results and the quality of the graphic material has been adjusted and improved. Therefore, I recommend publishing the manuscript. Congratulations to the authors.

Author Response

I am delighted that the reviewers acknowledged the improvements and adjustments we made. I sincerely appreciate your recognition and support.

Reviewer 3 Report

Thanks for considering the comments! Figure 1-3 still require improvement, at least the legends are not visible, and need to be bigger, as well as other text on graphics (map scales, geographical coordinated, etc.).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf