Analysis of Road Safety Perception and Influencing Factors in a Complex Urban Environment—Taking Chaoyang District, Beijing, as an Example

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Review. ijgi-3065867. Quantifying human perception of environmental security and identifying the streetscape elements that underpin this perception is pivotal for enhancing urban environments and fostering a heightened sense of safety among residents. Despite the significance, large-scale quantitative research on safety perception in Chinese cities remains underdeveloped. This study, therefore, focuses on Chaoyang District, Beijing, leveraging deep learning (CNN models), image semantic segmentation, and object detection technologies, coupled with LightGBM and SHAP interpretability frameworks, to analyze streetscape elements and their influence on road safety perception. The findings contribute to the interdisciplinary application of methods and data-driven safety perception analysis, though certain limitations necessitate further exploration.

Firstly, while the methodological innovations are commendable, the theoretical framework underlying the study could be strengthened. Incorporating theories from crime geography and environmental psychology would enrich the analysis, providing a more nuanced understanding of the mechanisms shaping road safety perception. This integrated approach would facilitate a deeper exploration of the interplay between physical features and human behavior in shaping safety perceptions. In this regard, the authors can refer to the earliest papers in this field, such as He et al., 2017, Built environment and violent crime: An environmental audit approach using Google Street View. CEUS; and Hwang, J., & Sampson, R. J. (2014). Divergent Pathways of Gentrification: Racial Inequality and the Social Order of Renewal in Chicago Neighborhoods. American Sociological Review, 79(4), 726-751. https://doi.org/10.1177/0003122414535774

Secondly, although the study identifies pivotal factors affecting road safety perception, it falls short of articulating concrete policy recommendations or improvement measures grounded in these findings. To bridge this gap and enhance the practical applicability of the research, targeted policies should be formulated. These might encompass optimizing road designs to reduce congestion and improve visibility, refining traffic management strategies, and intensifying safety education programs. By implementing such measures, the research outcomes can translate into tangible improvements in residents' road safety perception and the overall safety of the urban environment.

Author Response

respond  to reviewer1

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This study uses CNN method to obtain the values of road safety perception and then models the impacts of street view elements on the safety perception through LightGBM and SHAP. The topic is innovative. However, there are still some key defects needed to be solved.

First, In literature, the paper does not review on the factors that influence people’s safety perception. It introduces a lot about the perceptual safety dataset in previous studies.

Second, what does the street view elements stand for in the perspective of theory? Why these elements are important? What is new? Moreover, there are abundant description about the results, while there is seldom explanation for the machine learning results.

Third, the theoretical contribution of this study needs to be enhanced. 

Comments on the Quality of English Language

The writting needs to be improved.

Author Response

Respond  to reviewer

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

This is an interesting paper using machine learning to develop safety perceptions using street view in China. Whilst the paper has merit, I think it needs revisions before I can recommend it for publication. It lacks an inclusion of crime and place/fear of crime and place literature that has evolved from crime prevention though environmental design, environmental criminology explanations of crime opportunity and crime patterns (and associated fear), and ideas of defensible space and eyes on the street.

Literature review:
I think the literature on perceptions of safety/fear of crime is driven mostly from psychological literature but there is a strong research evidence within criminology that has not been covered. I think it is essential to recognise two different concepts of fear – dispositional fear and situational fear/perceptions.

Dispositional fear relates to the differences between individuals' propensities to experience fear of crime in different settings (for example age and gender) – so two individuals with different dispositional fear will perceive the street view images used in this study differently. Situational fear refers to a transitory state of experiencing fear based on the environment a user experiences. This is more akin to this study – although both are relevant in practice. Finally more recent studies have moved from cross sectional victimisation surveys to ecological momentary assessment (EMA) – recognising that safety perceptions may vary over the time of day, as users experiences place differently (for example daytime/nighttime, or even weekday/weekends. The model used here is in my mind a part hybrid between EMA and situational assessment – based on how person perceives a place at the time the street view image was taken. I recommend this systematic review be considered (https://doi.org/10.1177/0013916520947114).

This temporal dynamic is also a limitation of this study that should be discussed and recognised – fear/perceptions of safety are not a static concept

Methodology

I am not quite clear on how you trained your model to determine high/low perceptions of safety. For example, the MIT dataset you refer to used user inputs of images (which to you feel safer) to develop its indicators of safety of places. However, you acknowledge this may not be applicable to China. Is this based on the Li model (page 3) – please can you specify how the variables were selected to train the model.

Is there any user verification of perceptions (either in thr Li study which I think you have developed) or within your study. The MIT data was based on user input, but I am not sure how/if your model has been validated by real users.

For Figure 4 – is there an error with labelling. The arrows follow a, b, c, f, e, d. I think (d) and (f) just need swapping. It is difficult to read CNN model text – it is blurred.

I applaud the use of Bayes models – this is great to see.

Results

In your study the ‘busyness of place’ seems to be important in terms of how many people, how many cars on roads etc. I think this is important. On page 14 you draw from Halls work on spatial distance (personal space types). However for crime I think this is complex. For example certain environments where it is very crowded are more risky for pocket-picking and sexual touching (eg a busy train) but there tends to be less violent assaults. In places were there are low densities of people (you should also consider how ‘capable guardianship’ might influence this then people would be more fearful or rape and robbery for example. I disagree with the discussion about (lines 437-439) - when the number of pedestrians is large, the proportion of pedestrian area has no obvious relationship with the number of pedestrians. This may be because when the number of pedestrians reaches a certain scale, people's sense of safety will not be affected by the proportion of pedestrian area.

I would argue when the number of pedestrians is large there may be some crimes people are fearful of. There are limited studies here – some of this is covered in this chapter (https://doi.org/10.1093/oxfordhb/9780190279707.013.16) which discusses ‘Density, proximity and busyness’ and draws from work of Angel’s (1968), Clark, Belanger and Eastman (1996), and Loukaitou-Sideris (1999) to suggest their may be optimal levels of intensive whereby different crime types are more likely – which would impact on perceptions of safety at these different ‘intensities’ of pedestrians.

Conclusions

Your four key findings (p18) need to be related back to potential theoretical explanations – in particular you should draw from CPTED and environmental criminology/opportunity theories that have been developed – I think these may be as important as the psychological studies you discuss.

Again I question the findings on lines 551-555 – perceptions of safety/crime are nuanced – so what makes you fearful of pocket-picking, robbery, sexual groping, or assault are all different due to situational fears experienced. Trying to evaluate all crime (as single category) and high/low perceptions of safety misses some key nuances. This will be especially important for further research study 3 you identify (comparing with actual crime).

Policy Implications

What are the policy implications for the study – there is some learning that has practical value for crime prevention/increasing perceptions of safety.

Comments on the Quality of English Language

Some sentences needed reading 2-3 times to understand.

The flow of text was sometimes difficult to follow.

Suggest some moderate revision of flow of text to improve readability.

Author Response

Respond  to reviewer

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

In order to quantify how streetscape elements affect the level of human perception of environmental safety, the article calculates and presents the distribution of cyber safety perception in Chaoyang District, Beijing through a deep learning model, and then extracts the streetscape elements from the streetscape images and identifies and analyzes the elements that affect the perception of road safety by utilizing the image semantic segmentation and target detection techniques. The results show that the overall safety perception level is high in the study area, and the four factors that have the greatest influence on the environmental safety perception are motor vehicle ownership, road, sky, and the percentage of sidewalk area, and the different streetscape elements, the proportion and number of streetscape elements have an interactive effect on safety perception, and the degree of the image of the streetscape elements on the perception of safety changes. I think this paper makes a valuable contribution, providing sufficient methodological detail and justification, as well as sufficient analysis to support the conclusions reached. Before it is in good shape to be published, I do have a few points that should be addressed.

The main points that require edits would be the following:

1. Only using Google Street View API to obtain street view images may have data bias, consideration can be given to combining other data sources such as local street view data, social security data, etc. for comprehensive analysis and comparison to improve the comprehensiveness of the study

2. Whether the use of 50 meters as a spacing scale in the selection of sample points can adequately represent the streetscape situation of the entire Chaoyang District, the impact of different sample point spacing on the results of the study can be added to select the optimal sample point spacing.

3. While the main streetscape elements that influence perceptions of safety were identified, an in-depth exploration of the interactions between these elements was lacking. The interaction of these streetscape elements could be further analyzed, as well as differences in performance across neighborhoods or time periods.

4. The conclusion section is rather brief and does not fully summarize the research findings and their practical application value. The research findings could be appropriately added, and the applied value of the research and policy recommendations could be illustrated with practical examples.

5. The illustrations in the article are not clear enough and some of the pictures are missing legends.

6. Aspects of the manual evaluation of the streetscape safety perception scores as explanatory variables in this study are not discussed, focusing in exploring deep learning content.

7. Quantitative results obtained through machine learning models (mechanistic explanation-free, data-driven) Mechanistic explanations that are analyzed and then re-interpreted using SHAP are not reliable enough, and human perceptions of road safety can be captured from the perspective of subjective awareness, e.g., through questionnaires.

8. The reasons mentioned in the article for the differentiated results in safety perception scores in different areas relate only to the level of economic development and the construction of streets, which could be explored in greater depth in relation to the other combined influencing factors.

Author Response

Respond  to reviewer

Author Response File: Author Response.docx

Reviewer 5 Report

Comments and Suggestions for Authors

The article takes Chaoyang District in Beijing as an example and measures urban road safety perception using existing methods and data. It employs a variety of machine learning techniques, including semantic recognition and target detection, to extract multiple elements from a large amount of street view data. Based on the LightGBM and SHAP interpretation framework, it explores the impact and characteristics of these elements on urban road safety perception. In today's era of increasingly abundant multi-source big data, studying safety perception based on street view images and utilizing machine learning methods reflects the latest research trends. This is particularly necessary for research conducted in important and typical megacities like Beijing. The research is well-structured, with clear writing, specific content, appropriate method selection, expected conclusions, and reasonable explanations, making it a well-executed paper that is recommended for publication. Here are some minor revision suggestions to help the authors further improve the paper:

(1)On line 89, the initial letter of "With" should be lowercase.

(2)On line 168, the authors mention "Street view safety perception score is the explanatory variable of this study." However, the street view safety perception score seems to be the dependent variable.

(3)On line 173-174, the authors state "and its parameters and structure have been proved to be reliable in many aspects of image feature extraction." A reference should be added here.

(4)On line 382-383, the author writes "Figure 11 shows our analysis of the street scene elements with obvious interaction in Figure 10." This sentence is unclear and seems to contain errors.

(5)In reference number 28, the journal name is incorrect.

(6)The formatting of cited references in the article should be revised according to the ijgi style to improve readability.

(7)The images are too blurry and need to be replaced with clearer ones.

(8)It is suggested that in the section of the data source, the authors provide a thorough account of how the data on road safety perceptions were obtained, including the channels and methods used.

Author Response

Respond  to reviewer

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed all my concerns.

Comments on the Quality of English Language

Minor editing of English language required.