Using deep learning to examine the correlation between transportation planning and perceived safety of the built environment

Abstract

In this study, we attempt to estimate the effects of various transportation policies on the perceived safety of the built environment. We train a convolutional neural network on a dataset of safety perception scores for Google Street View images taken in Boston, MA . We then apply the trained neural network to a large set of Google Street View images of coordinates in Montreal and Toronto to generate their respective safety perception scores. We estimate probit, logit, and ordinary least squares regression models using our cross-sectional dataset consisting of safety perception scores, as well as transportation policy variables and a set of control variables, by regressing the safety perception scores on the remaining set of variables. We answer our research question by observing the direction, magnitude, and statistical significance of the coefficient estimates associated with the policy variables across all regression models. We studied and cataloged transportation policies planned for over the past 10 years in both cities. We found that those census tracts with the poorest safety scores were the same places where planners focused their transportation investments. The study makes an important contribution to transportation planning methodologies by drawing on the novel data source of Google Street View images, to understand the safety of an area.

Keywords

Transportation policy machine learning Google Street View Canada

Introduction

The built environment has various physical features from which its inhabitants gain a variety of experiences directly or indirectly.¹ Glaeser (2012) writes that local governments hold considerable authority over the built environments under their jurisdiction. The built environment affects the well-being of people living within a given jurisdiction and is impacted by services and activities performed by a local government. Local government planners develop strategies to determine where to make investments in the built environment. In determining where to add new subway lines or bicycle paths, planners need to obtain knowledge about the effects of certain properties of the built environment on health (Guo and Gandavarapu, 2010; Peacock-McLaughlin et al., 2018). Since the aforementioned mechanism is complex and multifaceted, it is important to approach the issue by breaking it down and isolating specific well-defined relationships. For this reason, the scope of our study is narrow—we measure the relationship between various transportation policies and the perceived safety of the built environment.

The extent to which a location seems safe is directly determined by the physical characteristics of the surrounding area. Our focus is the perception of safety. Although perceived safety is not strictly a component of the built environment, a critical assumption made in our study is that perception of safety of a location has an impact on the utility of its inhabitants that interact with the built environment. For this reason, perceived safety may also be a factor driving planning and other investment decisions. From this logic, it follows that policies will be developed and targeted to places where safety is perceived to be poor.

Aiming to verify our aforementioned assumption, this study estimates the relationship between various transportation policies and the perception of safety using data from the cities of Montreal and Toronto. We intend to find out how different modes of transportation mentioned in transportation policies correlate to safety perception.

To do so, we create a dataset with location, policy, and non-visual attribute variables for a large sample of Google Street View images in both cities. A binary safety perception score is then assigned to each image using a neural network algorithm trained on a similar dataset of Boston, MA, in which safety perception scores are generated for images using a large set of survey responses. Lastly, we apply a battery of simple cross-sectional regression models to estimate the effects of various policies by testing the significance of their respective coefficients.

The literature review section in this study explores past research which analyzes the importance of perception of safety in transportation planning, how to measure the perception of safety, and the context of Canadian planning and urban development in Montreal and Toronto; Methodology section describes the process through which our dataset was generated, along with the statistical analysis methodology used to estimate policy effects; Results section presents and interprets the results of the analysis; lastly, Discussion section discusses the implications of the results of the study and recommendations for future research.

Literature review

A review of the literature shows that the perception of safety is associated with different dimensions of transportation planning and community development. Various studies have investigated how perceived safety is related to transportation mode choice. Transportation plans and reports released in the past 10 years in Canada show that Canada was in a transition of transportation planning in regard to the degree of attention to diverse travel modes.

Perception of safety and transportation planning

The built environment is the combination of all physical aspects of an urban territory that can be either directly or indirectly experienced by its inhabitants. Local governments hold considerable authority over the built environments under their jurisdiction (Glaeser, 2012). In turn, the built environment affects the well-being of people living within a given jurisdiction. There are several studies which highlight the relationship between built environment and personal well-being including mental health, of which some focused on the human perceptions of their surroundings and the quality of that built environment (Croucher et al., 2007; Dalgard and Tambs, 1997; Gong et al., 2016; Lavin et al., 2006; Wainwright and Surtes, 2003; Weich et al., 2002; Won et al., 2016). At the end of the last century, Dalgard and Tambs (1997) found that neighborhoods with poor environments had higher number of residents with higher environmental stress, of which the characteristics of the built environment would have association with emotional states. Lavin et al. (2006) wrote that streets, routes, and buildings are the particular means by which the built environment determines the well-being in neighborhoods. Stathi et al. (2012) had found that the availability of public transportations and the accessibility to open space were the keys for improving the happiness of local neighborhoods, especially for groups like the elderly. On the contrary, poor traffic and rebuilding and construction of the built environment could have the opposite impact on mental health (Croucher et al., 2007; Stafford et al., 2007; Whitley and Prince, 2005). Although there were various studies presented above about the relationship between built environment and individual well-being, none of these studies adopted Google Street Views as the data source for analyzing the built environment.

Lin and Ting (2014) find a significant link between built environment factors and transportation and leisure activities of adolescents. Similarly, Durand et al. (2011) performed a meta-analysis of a large set of studies and deduced the presence of a positive relationship between the level of physical activity and the diversity of housing types, mixed land use, housing density, and levels of open space. Such evidence demonstrates the opportunity for local governments to attempt to maximize social benefit through residents’ interactions with the built environment.

Perception of traffic safety is highly associated with mode choices of active transportation (Aziz et al., 2018; Handy et al., 2006; Winters et al., 2010). Accident occurrences have been found to have direct impacts on pedestrians and bicyclists’ perception of safety (Dill and Carr 2003; Ferdous et al., 2011). Also, Nevelsteen et al. (2012) found that both location and infrastructure characteristics play an important role in people’s perceived safety. Students’ travel mode is also impacted by their parent’s perceptions of safety on the built environment (Guliani et al., 2015; Rothman et al., 2015).

The actual levels of crime vary based on its perceived safety of an environment and overall well-being of community (Okunola and Amole, 2012; Pain, 2000). Pain (2000) showed how the design of subways and streets directly influenced people’s perception of a particular place. There is evidence to believe that an improvement in the perceived safety of a location has a positive effect on the physical activity of nearby residents (Shenassa et al., 2006), as well as on the extent of interactions between members of a community (Baum and Palmer, 2002).

Canadian planning and urban development

Context of Canadian planning

Priority issues in Canada revolve around downtown areas and ways in which they can be restructured to be better serviced by public transportation and provide public amenities and the linked question of suburbs (Simmins, 2011). The jurisdiction over transportation systems fall under three levels of government: federal, provincial, and municipal governments. On a federal level, transportation policy-making and regulatory activities are managed by Transport Canada and other transportation planning and strategies are typically threaded throughout municipal plans in Canada (Lawson, 2015). Transportation policies are encouraged to promote density and land use mix that reduce vehicle trips and diversify the modes of public transit (Lawson, 2015).

Background on Montreal

Between the years of 1970 and 1981, Montreal experienced a decline in population density. The drop in density was due, in part, to the decrease in the number of households and the rise of suburban development and demographic shift from the urban center to the edges (Fischler and Wolfe, 2000). As the urban fringe started to expand, there was a perceived need to extend accessible transportation to these areas in order to continue economic growth and connectivity within the city (Nazarnia et al., 2016).

In the early 2000s, the merging of the municipalities of island of Montreal into a super government body shifted the population trend from the suburbs, resulting in 53% of the residents living in the city center (Nazarnia et al., 2016). The city of Montreal itself has a population of 1.7 million (Statistics Canada).

Background on Toronto

The city of Toronto is located in the Province of Ontario, Canada. At 2.7 million, Toronto has the largest population compared to any other city in Ontario and Canada (Statistics Canada). The city boundaries cover approximately 630 km² of land and, like Montreal, is a super government body subsuming scores of formerly independent municipalities within the region.

Since 2008, the median housing price in the city of Toronto has exceeded that of suburban areas connected to Toronto by highways and GO train, especially in York and Halton regions (RBC and the Pembina Institute, 2013). Will and Young (2015) found that housing in suburban areas was more affordable for several groups of people who sought cheaper living expenses. The number of new condos grew on account of the increasing share of transit modal share in Toronto Metro region beginning in the 2000s (Searle and Filion, 2011).

Methodology

We begin this section by describing the process using which we generated the cross-sectional dataset used in our analysis. The following is a brief overview of the involved steps:

We first trained a convolutional neural network on safety perception data on a large set of Google Street View images of the streets of Boston created by Raskar et al. (2015). We then applied the trained neural network to a set of Google Street View images of streets in Toronto and Montreal to assign the images their respective safety perception scores.

We created a dataset containing a list of all census tracts in Toronto and Montreal, as well as associated values of all relevant transportation policies and control variables.

Lastly, we created a master dataset containing the Google Street View images of Toronto and Montreal streets as observations, and then matched each observation with associated census tracts to assign their respective transportation policy and control variable values.

Step 1: trained convolutional neural network

While designed to assist in navigational purposes, Google Street View has been widely used for a variety of building and neighborhood audit purposes and has been found in previous studies to be a reliable and practical tool to measure the characteristics of built environments (Badland et al., 2010; Curtis et al., 2013; Kelly et al., 2013; Odgers et al., 2012; Rundle et al., 2011; Vanwolleghem et al., 2016; Wilson et al., 2012). Raskar et al. (2015) created a scene-understanding algorithm called “Streetscore” to predict the safety perception of a street view. By using “Streetscore”, the maps of perceived safety from Google Street View were created for 21 cities around the United States (Raskar et al., 2015). Salesses et al. (2013) conducted research where their study participants’ perceptions of safety were only driven by the difference in the images’ visual attributes but not by biases in participants’ personal attributes (such as age or gender). It is feasible to conduct a computational modal for analyzing perceived safety based on image features (Raskar et al., 2015).

Building on Raskar et al. (2015), we developed a score-allocating model that assigned safety perception scores to inputted images and links those results with transportation planning priorities. A safety perception score for an image may take on a value of either 0 or 1, where a score of 0 represents an unsafe site, while a score of 1 represents a safe site. The score-allocating model is a convolutional neural network (LeCun and Bengio, 1995) that was trained on datasets of safe and unsafe sites (we continue to refer to this as the training dataset). The training dataset, discussed and generated by Raskar et al. (2015), is made up of 229,564 images of neighborhoods in Boston, with a safety score assigned to each image. The safety scores range from 0 to 40, with 0 being the least safe and 40 being the safest.

We trained the convolutional neural network using transfer learning (Tan et al., 2018). We used a pre-trained Resnet-50 (He et al., 2015) model and fine-tuned the last couple of layers with our training dataset. We then used a concentric image-collection algorithm to gather Google Street View images of sites across Toronto and Montreal (20,103 images for Toronto and 29,117 images for Montreal). Depictions of the process for Montreal and Toronto are presented in Supp Figures 5 and 6 in the Appendix, respectively, along with a more detailed account of the data generation process and a brief explanation of the concept of neural networks. Finally, we ran our trained score-allocation model on the gathered images. We believe that the model we trained on images of neighborhoods in Boston generalizes well to neighborhoods in Montreal and Toronto because of the similarity in city planning and building structure of these cities (Dear and Scott, 2018).

We trained the model using a Tesla K80 GPU with 11 GB of memory. The CPU used was an Intel(R) Xeon(R) CPU @ 2.30 GHz. We trained the model to a prediction accuracy of 84.3%. Out of the 29,117 images of sites in Montreal, 11,165 images had a score of 1 and 13,012 images had a score of 0.² Out of the 20,103 images of neighborhoods in Toronto, 10,530 had a score of 1 and 9,573 had a score of 0. Supp Figure 1 presents examples of sites in Montreal with an assigned safety score of 0, and Supp Figure 2 presents examples of sites in Montreal with an assigned safety score of 1. Supp Figures 3 and 4 present the same examples as Supp Figures 1 and 2, but of sites in Toronto instead of Montreal.

Step 2: identification of transportation policies and other variables

In the next step, we gathered transportation policy data for the two cities through systematic policy tracking. Brooks (2002) and Hopkins (2001) report that older plans and reports were not compelling enough for policy analysis because they diverge from current societal conditions (Brooks, 2002; Hopkins 2001). In addition, our focus was on the reports and plans having on-going or recent-finished goals which would happen around the year of our research. Thus, we extracted transportation planning data from urban planning reports of Toronto and Montreal created in the past 10 years. The policy plans selected for Montreal are the following:

“Sustainable Development Plan 2025—Societe de Transport de Montreal (STM)”;

“2008 Transportation Plan”.

The former was released in 2017 by STM, a public transportation enterprise in Montreal. By targeting a time horizon of 10 years, the plan aimed to achieve sustainable development through several activities, such as delivering multiple transit services such as bus and metro. The latter is a transportation-focused plan for the entire City of Montreal; it involves the assessment of major investments on road infrastructure and public transit networks.

The policy plans chosen for Toronto are the following:

“The Big Move: Transforming Transportation in the Greater Toronto and Hamilton Area (GTHA)”;

“Toronto Official Plan”;

“Building Ontario 2012”.

The Big Move contains guides and blueprints of all modes of transportation and transit systems that have been put in place since 2008. The Toronto Official Plan is a comprehensive plan focused on improving Toronto’s city connections, diversity, and beauty. The final plan is an infrastructure investment plan focused on building public facilities including roads and public transit in Ontario, including Toronto.

We determined which census tracts in Montreal and Toronto were focused on by each plan, as well as what types of transportation policies were being recommended for each census tract. Since words are the smallest unit of analysis that can be used to analyze frequency (Gaber and Gaber, 2007), we pre-read each document to determine our list of policy variables.³ First, we categorized the types of transportation mentioned by the plan to be the following: tramway, subway, station, road, pedestrian path, parking, ferry service, commuter lines, bus (BRT), bus, bike, underground PATH system, light rail, and rail. Underground PATH system refers to the walkways underground which connect office towers, especially in Downtown Toronto (City of Toronto, 2020).The system connects more than 30 km (19 mi) of tunnels, walkways, and shopping areas. Then we manually counted the numbers of different policies into the above categories. Additionally, we focused on the corresponding streets or specific addresses of each transport investment mentioned in each policy or plan. We identified the locations that were impacted by each transportation policy via Google Maps and used Online ArcGIS map to determine the census tract IDs to which those locations corresponded. By grouping the above information into Excel, we counted the numbers of categorized transportation policies in each census tract. Ultimately, we created a binary variable for each policy, where a value of 1 indicated the presence of the associated policy in a census tract, while 0 indicated the lack of such a policy. Supp Tables 1–3 show the percentage of observations across both cities, as well individually for Montreal and Toronto, respectively, which are located in census tracts that enforce each given policy. Figure 1 presents the maps of safety scores perceptions score in its corresponding census tract in each city, along with the corresponding locations of total numbers of policies mentioned in reports.

Figure 1.

Policy density maps with the safety scores in Montreal and Toronto.

Step 3: matching of street view and transportation data

For the sake of parsimony of our regression models, we use a smaller set of control variables measured in 2016 at the census tract level and come from the Canadian Census. The following is a list of the control variables actually used in the analysis:

Percentage of private dwellings occupied by usual residents;

Population density;

Percentage of population that is male;

Percentage of population that holds a Bachelor’s degree or higher;

Percentage of population between the ages of 15 and 64;

Median household income.

We approach our research question with a cross-sectional framework that attempts to estimate the significance of the effects of individual policies on the perception of safety in the cities of Montreal and Toronto. More specifically, we regress the safety perception score variable on a set of policy and control variable. We estimate the regression model using three different estimation methods: (1) probit, (2) logit, and (3) ordinary least squares (OLS) with robust standard errors (McCullagh and Nelder, 1989). All three of the aforementioned models are estimated over three different samples: (1) total sample containing observations from both Montreal and Toronto, (2) sample of observations restricted to Montreal, and (3) sample of observations restricted to Toronto. For the estimates on each sample, we pick a set of appropriate policy variables. These variables are chosen based on the frequency of positive values over each restricted sample—variables with mostly zero values are excluded as a way of avoiding biased coefficient estimates, while those with over 5% positive values are kept (Supp Tables 4 and 5).

By following the guidelines for the choice of policy variables described above and by observing Supp Tables 6, 8, and 10, we can decide that the regression models over the total sample must contain road, commuter line, bus, underground path system, and light rail policy variables (Supp Table 8). Similarly, we can deduce that the regressions over the sample of observations in Montreal must contain tramway, subway, road, commuter line, and bus policies (Supp Table 10). Lastly, we can deduce that the regressions over the sample of observations in Toronto must contain road, underground path system, and light rail policies (Supp Table 6). We also estimate probit, logit, and OLS models over all three samples with the specification containing all policy variables as regress ands—this allows us to test the robustness of coefficient estimates from the base model, as well as observe the extent to which the skewed distributions of the excluded policy variables potentially generate biased estimates of those included in the base model.

We refer to the model which is specified to have a smaller set of policy and control variables as the base specification, while the model that contains all available policy and control variables is referred to as the robust specification. All available demographic census data were included in the robust specification as control variable, while the base model contains only those controls that turned out to be statistically significant.

The base model was achieved by first estimating the robust specification and ridding the model of all policy and control variables with coefficient estimates that failed to be significant at a 10% level or higher. The base model is more parsimonious, which leads us to focus on it while using the robust model as an additional reference.

The nature of the relationship between policy variables and safety perception we would ideally like to uncover using the above framework is causal. This then begs the question of whether estimating a simple cross-sectional linear regression model is an appropriate approach—specifically, one might be concerned about biased coefficient estimates as a result of confounding variables. Admittedly, such concerns are to some extent warranted, but we try to solve the issue with the inclusion of relevant control variables, as well as using robustness tests presented in the following section. For the time being, this is the most effective method of mitigating potential policy coefficient estimate biases. Over the course of the following few decades, as longitudinal Google Street View data become available, it will be possible to generate policy effect estimates using more appropriate econometric analyses such as difference-in-difference, matching models (Chiappori and Salanié, 2016), synthetic control methods (Abadie et al., 2010; Pinotti, 2015), and other panel data frameworks (Wooldridge, 2002).

Results

The base specification coefficient estimates obtained by the regression models on the total sample are presented in Table 1. According to these estimates, the presence of a road policy is associated with a negative effect on the perceived safety of the built environment. This result is confirmed by models (1), (2), and (3), which represent probit, logit, and OLS estimates, respectively. All of the model estimates are statistically significant at a 1% level. The same result is associated with the presence of commuter line and light rail policies at a 1% significance level across all models. In the case of commuter line policy, a smaller magnitude is assigned relative to those of both road and light rail policies. In other words, road, commuter line, and light rail policies have an estimated negative effect on the perceived safety of the built environment, but the effect of commuter line policy is least in magnitude. The robustness test results presented in Supp Table 7 match the coefficient estimates of the road policy variable closely. The direction of the effect for the commuter line policy variables is also negative; however, the probit estimates show a statistical significance level of 10%, while the logit and OLS estimates show a significance level of 5%. Similarly, the magnitude of the commuter line policy effect is estimated to be lower in the robustness tests relative to those of the base specification estimates. The light rail policy variable was removed from all of the models in Supp Table 7 due to multicollinearity with other added policy variables.⁴

Table 1.

Base specification regression on total sample.

	(1)	(2)	(3)
Road policy	–0.751***	–0.466***	–0.168***
	(–18.75)	(–19.43)	(–20.08)
Commuter line policy	–0.287***	–0.183***	–0.068***
	(–6.64)	(–6.79)	(–6.91)
Bus policy	0.240***	0.154***	0.056***
	(5.86)	(6.03)	(5.91)
Underground path system policy	0.957***	0.583***	0.213***
	(13.75)	(13.96)	(14.87)
Light rail policy	–0.652***	–0.393***	–0.142***
	(–11.39)	(–11.50)	(–12.31)
Total population	–0.000***	–0.000***	–0.000***
	(–8.48)	(–8.42)	(–8.59)
% of dwellings occupied by usual residents	1.057**	0.676***	0.231**
	(2.57)	(2.71)	(2.52)
Population density (2016)	0.000***	0.000***	0.000***
	(12.17)	(12.66)	(13.48)
% of population, male	3.407***	2.150***	0.817***
	(4.68)	(4.86)	(4.93)
% of population w/bachelors or beyond	2.122***	1.319***	0.516***
	(15.50)	(15.87)	(17.02)
% of population ages 15–64	–4.429***	–2.700***	–1.018***
	(–15.03)	(–15.18)	(–15.53)
Median household income	0.000***	0.000***	0.000***
	(11.10)	(11.35)	(11.80)
Observations	31,259	31,259	31,259
R ²			0.097
Adjusted R²			0.096
F			326.4

(1) is a probit model, (2) is a logit model, and (3) estimated using OLS. T-statistics are in parentheses.

*10% significance level.

**5% significance level.

***1% significance level.

The presence of bus and underground path system policies are both associated with a positive effect on the perceived safety of the built environment according to the estimates presented in Table 1. The magnitude of the effect of the underground path system is considerably greater across the estimates of all three models in comparison to those of bus policy. For both bus and underground path system policy, all of the model estimates are statistically significant at a 1% level. The robustness tests presented in Supp Table 7 do not confirm these results, however. According to Supp Table 7, the coefficient estimates of the bus policy are negative and statistically insignificant in the probit model, while its magnitude is near-zero in the logit and OLS estimates while simultaneously being statistically insignificant. The underground path system policy variable is removed during the estimation process due to multicollinearity with other policy variables.⁵

The base specification coefficient estimates obtained by regressions on the sample restricted to Montreal observations are presented in Supp Table 8. These estimates show that the presence of all included policies (tramway, subway, road, commuter line, and bus) has a negative effect on the safety perception of the built environment in all estimates models. Subway and road policy coefficient estimates are significant at a 1% level in all models. The tramway policy coefficient is significant at a 1% level only in the model estimated using OLS, while it is significant at a 5% level in the probit and logit models. The commuter line and bus policy effects are statistically insignificant in all model estimates. The directions and magnitudes of the aforementioned coefficient estimates match those of the robustness tests presented in Supp Table 9 to a large extent, with only one notable exception. Specifically, the commuter line policy coefficient is statistically significant at a 10% level in the probit model, while being significant at a 5% level in the logit and OLS models. Also, the magnitudes of coefficient estimate of the commuter line policy in models (1)–(3) of Supp Table 9 are noticeably greater than those of Supp Table 8.

The base specification coefficient estimates obtained by regressions on the sample restricted to Toronto observations are presented in Supp Table 10. According to these estimates, the presence of road and light rail policies has a negative effect on the perceived safety of the built environment. The coefficient estimates for both policy variables are statistically significant at a 1% level in the probit, logit, and OLS models. Both the directions and magnitudes strongly match those of the estimated models presented in Supp Table 11. Additionally, according to the model estimates presented in Supp Table 10, the presence of underground path system policy is associated with a positive effect on the perceived safety of the built environment. This is confirmed by the probit, logit, and OLS model estimates presented in both Supp Tables 10 and 11 at a 1% significance level.

Discussion

Our analysis was an observational study that attempted to estimate the effects of various transportation policies on the perceived safety of the built environment.

Here, we restate our results regarding the effects of the following policies: (1) road, (2) commuter line, (3) bus, (4) underground path system, (5) light rail, (6) tramway, and (7) subway. The results are as follows:

Road. The road policy variable is included in the regressions run on all samples. The base specification shows the road policy to be associated with a statistically significant negative effect on the safety perception of the built environment across all samples. This relationship is further confirmed by the robust specifications included in all of the tables for entire samples.

Commuter line. The commuter line policy variable is included in the regressions on the total and Montreal samples, while being excluded from the Toronto-only sample. The base specification shows the commuter line policy to have a negative effect on safety perception as well. This result is not statistically significant in the base specification of the regressions on the Montreal sample but is statistically significant in the robust specification, as well as both specifications on the total sample.

Bus. The bus policy variable is included in the regressions on the total and Montreal samples only. Its average effect remains inconclusive, however, due to the fact that the coefficient estimates are inconsistent across both the base and robust specification regressions on the Montreal sample.

Underground path system. The underground path system policy variable is included in the regressions on the total and Toronto samples, while being excluded from the Montreal sample. The base specification shows the given policy to have a positive effect on the safety perception of the built environment according to the regressions on both samples. Cross-referencing these estimates with those of the robust specifications of both models is not possible, due to the variable being removed because of multicollinearity.

Light rail. The light rail policy variable is included in the regressions on the total and Toronto samples, while being excluded from the Montreal sample. The base specification shows the given policy to have a negative effect on the safety perception of the built environment according to the regressions on both samples. Similar to the underground path system policy variable, however, the light rail policy variable was removed from the robust specifications on both samples due to multicollinearity.

Tramway. The tramway policy variable is included in the regressions on the Montreal sample only. The base specification shows the tramway policy to be associated with a negative effect on safety perception. The robust specification estimates on the Montreal sample confirm this relationship.

Subway. The subway policy variable is included in the regressions on the Montreal sample only. The base specification shows the given policy having a negative effect on the safety perception of the built environment. The robust specification estimates on the Montreal sample confirm this relationship.

Unquestionably, dozens of transportation policies implemented over the course of many years by two large cities have resulted in numerous changes to the built environment.

In both Montreal and Toronto, the planning highlighted in this research has often resulted from community-based processes and was the result of political and public concerns around access, mobility, and safety. What is quite extraordinary about our findings is that across both cities, with the single exception of those policies to promote underground paths in Toronto, no other policy was correlated with any measurable positive safety perception in its census tract. Put another way, our analysis shows that transportation planning measures were more likely to be implemented in tracts which are perceived as less safe than others. Furthermore, by including a large set of control variables in our regressions, we can confidently say that the direction of this relationship is likely to be estimated correctly.

The implication for urban planning is important: places that are perceived as unsafe are being targeted disproportionately by transportation planners than other places. While planners may rationalize their investment decisions and provide reams of explanation and justification, all things being equal, they are targeting the neighborhoods in their cities which look unsafe. Our results show that the presence of past policies related to road, commuter line, light rail, tramway, and subway is associated with a negative effect on the perceived safety of the built environment. It might reflect that previous planners not only did not target their planning area by precisely considering the perceived safety of that corresponding surroundings but also did not focus on how to improve the perceived safety while making planning or changes to their targeting areas.

A concern regarding both the base and the extended models is that their R² values are relatively low.⁶ This raises the question of whether our simple model is a proper representation of the true mechanism that determines safety scores of the environment. The low explanatory power likely stems from the fact that the model contains purely macro-level regressors, despite safety scores being a much more micro-measure. Clearly, a much better predictor of safety scores would be a model that contains various physical properties of images of the environment as regressors (for example, the number of trees and buildings in an image).

The goal of this study, however, is not to build a predictive model but rather to estimate the direction of the effects of large-scale transportation policies on safety scores. The noise that leads our model to have a low R² is an aggregate of variation in the many types of micro-level determinants similar to the couple mentioned above. Further research should seek to find out other factors at play.

It is also important to consider that our results do not show a causal relationship, but rather correlations. The identification of causal relationships requires a longitudinal dataset and a more rigorous identification method.

Limitations

In our Methodology section, we introduced the computational modal built by Raskar et al. (2015) as the novel technological method utilized in answering our study. Our study has contributed to Raskar et al. (2015)’s findings that show their method’s feasibility in transportation analysis, but only with respect to a narrow set of socio-economic and demographic dimensions we included in our model.

Given that the number of planning and reports assessed in this study was small, our ability to generalize is limited. Future research could include additional study cities and expand the total universe of plans and reports analyzed. Future research should assess safety across different modes of transportation, building on these findings and analyze different modes of transportation and their corresponding perceived safety.

Conclusion

This paper has uncovered a heretofore invisible motivating factor in the planning process—the role of perceived safety. This study makes a useful contribution to transportation planning practice by introducing an approach for better elucidating the safety of an area and drawing on the increasingly ubiquitous Google Street View data source. By training a neural network, we were able to make sense of vast quantities of image data and introduce a useful framework for how planners can study their communities and utilize planning and policy strategies to shape the built environment.

Supplemental Material

sj-pdf-1-epb-10.1177_2399808320959079 - Supplemental material for Using deep learning to examine the correlation between transportation planning and perceived safety of the built environment

Supplemental material, sj-pdf-1-epb-10.1177_2399808320959079 for Using deep learning to examine the correlation between transportation planning and perceived safety of the built environment by Justin B Hollander, Giorgi Nikolaishvili, Alphonsus A Adu-Bredu, Minyu Situ and Shabnam Bista in EPB: Urban Analytics and City Science

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for research, authorship and/or publication of this article: We wish to acknowledge the Research Support Program on Intergovernmental Affairs and Québec Identity, Government of Québec, Secrétariat aux affaires intergouvernementales canadiennes (SAIC) for their financial support of this research.

Notes

Supplemental material

Supplemental material for this article is available online.

Justin B Hollander is a Professor in the Department of Urban and Environmental Policy and Planning at Tufts University.

Giorgi Nikolaishvili is a PhD student in the Economics Department at the University of Oregon.

Alphonsus A Adu-Bredu is a PhD student in the Robotics Institute at the University of Michigan.

Minyu Situ is a PhD student in the Community and Regional Planning Program at the University of Texas at Austin.

Shabnam Bista is an associate planner with the City of Boulder in Colorado.

References

Abadie

Diamond

Hainmueller

(2010) Synthetic control methods for comparative case studies: Estimating the effect of California’s tobacco control program. Journal of the American Statistical Association 105(490): 493–505.

Aziz

HMA

Nagle

Morton

, et al. (2018) Exploring the impact of walk–bike infrastructure, safety perception, and built-environment on active transportation mode choice: A random parameter model using New York City commuter data. Transportation 45: 1207.

Badland

Opit

Witten

, et al. (2010) Can virtual streetscape audits reliably replace physical streetscape audits? Journal of Urban Health : Bulletin of the New York Academy of Medicine 87(6): 1007–1016. [21104331]

Baum

Palmer

(2002) ‘Opportunity structures’: Urban landscape, social capital and health promotion in Australia. Health Promotion International 17(4): 351–362.

Brooks

(2002) Planning Theory for Practitioners. Chicago, IL/Washington, DC. American Planning Association Press.

Chiappori

P-A

Salanié

(2016) The econometrics of matching models. Journal of Economic Literature 54(3): 832–861.

City of Toronto (2020) PATH—Toronto’s downtown underground pedestrian walkway. City of Toronto.org. Available at: https://web.archive.org/web/20140620023017/http://www1.toronto.ca/wps/portal/contentonly?vgnextoid=f537b454b35a2410VgnVCM10000071d60f89RCRD&vgnextchannel=04708b7a29891410VgnVCM10000071d60f89RCRD (accessed 19 March 2020).

Croucher K, Myers L, Jones R, et al. (2007) Health and the Physical Characteristics of Neighbourhoods: a Critical Literature Review. Technical Report. Glasgow Centre for Population Health, Glasgow, UK.

Curtis

Mapes

, et al. (2013) Using Google Street View for systematic observation of the built environment: Analysis of spatio-temporal instability of imagery dates. International Journal of Health Geographics 12(1): 53.

10.

Dalgard

Tambs

(1997) Urban environment and mental health: A longitudinal study. The British Journal of Psychiatry: The Journal of Mental Science 171: 530–536.

11.

Dill

Carr

(2003) Bicycle commuting and facilities in major U.S. cities: If you build them, commuters will use them. Transportation Research Record: Journal of the Transportation Research Board 1828(1): 116–123.

12.

Durand

Andalib

Dunton

, et al. (2011) A systematic review of built environment factors related to physical activity and obesity risk: Implications for smart growth urban planning. Obesity Reviews 12: 173–182.

13.

Ferdous

Pendyala

Bhat

, et al. (2011) Modeling the influence of family, social context, and spatial proximity on use of nonmotorized transport mode. Transportation Research Record 2230: 111–120.

14.

Fischler

Wolfe

(2000) Regional restructuring in Montreal: An historical analysis. Canadian Journal of Regional Science/Revue Canadienne Des Sciences Régionales, XXIII 1: 89–114.

15.

Gaber

(2007) Qualitative analysis for planning & policy beyond the numbers. In: Content Analysis and Meta-Analysis.

16.

Glaeser

(2012) Urban Public Finance.

17.

Gong

Palmer

Gallacher

, et al. (2016) A systematic review of the relationship between objective measurements of the urban environment and psychological distress. Environment International 96: 48–57.

18.

Guliani

Mitra

Buliung

, et al. (2015) Gender-based differences in school travel mode choice behaviour: Examining the relationship between the neighbourhood environment and perceived traffic safety. Journal of Transport & Health 2(4): 502–511.

19.

Guo

Gandavarapu

(2010) An economic evaluation of health-promotive built environment changes. Preventive Medicine 5, S44–S49.

20.

Handy

Cao

Mokhtarian

(2006) Self-selection in the relationship between the built environment and walking: Empirical evidence from Northern California. Journal of the American Planning Association 72(1): 55–74.

21.

Zhang

Ren

, et al. (2015) Deep residual learning for image recognition. CoRR, abs/1512.03385.

22.

Hopkins

(2001) Urban Development: The Logic of Making Plans. Vol. 166. Washington, DC: Island Press.

23.

Kelly

Wilson

and Baker

(2013) Using Google Street View to audit the built environment: Inter-rater reliability results. Annals of Behavioral Medicine 45(1): 108–112. (

24.

Lavin

Higgins

Metcalfe

, et al. (2006) Health Impacts of the Built Environment. A Review. Dublin: Institute of Public Health in Ireland.

25.

Lawson

(2015) Transportation Policy and Planning in Canada. Buenos Aires, Argentina: Transport Institute at Universidad Nacional de San Martin.

26.

LeCun

Bengio

(1995) Convolutional networks for images, speech, and timeseries. In: Arbib

(ed.) The Handbook of Brain Theory and Neural Networks. Cambridge; MA: MIT Press.

27.

Lin

J-J

Ting

T-C

(2014) Does built environment matter to early adolescents’ physical activity? The Journal of Early Adolescence 34(8): 1005–1032.

28.

McCullagh

Nelder

(1989) Generalized Linear Models (Chapman & Hall/CRC Monographs on Statistics and Applied Probability). London: Chapman and Hall/CRC.

29.

Nazarnia

Schwick

Jaeger

JAG

(2016) Accelerated urban sprawl in Montreal, Quebec city, and Zurich: Investigating the differences using time series 1951–2011. Ecological Indicators 60(January): 1229–1251.

30.

Nevelsteen

Steenberghen

Van Rompaey

, et al. (2012) Controlling factors of the parental safety perception on children’s travel mode choice. Accident Analysis and Prevention 45: 39–49.

31.

Odgers

Caspi

Bates

, et al. (2012) Systematic social observation of children’s neighborhoods using Google Street View: A reliable and cost-effective method. Journal of Child Psychology and Psychiatry, and Allied Disciplines 53(10): 1009–1017.

32.

Okunola

Amole

(2012) Perception of safety, social participation and vulnerability in an urban neighbourhood, Lagos, Nigeria. Procedia—Social and Behavioral Sciences 35: 505–513.

33.

Pain

(2000) Place, social relations and the fear of crime: A review. Progress in Human Geography 24(3): 365–387.

34.

Peacock-Mclaughlin

Largo-Wight

Wlyudka

, et al. (2018) The built environment, transportation policy, and population health: A comparison of two cities. Urban Research & Practice 11(3): 193–199.

35.

Pinotti

(2015) The economic costs of organised crime: Evidence from Southern Italy. The Economic Journal 125(586): F203–F232.

36.

Raskar

Naik

Philipoom

, et al. (2015) Streetscore—Predicting the perceived safety of one million streetscapes. In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops.

37.

RBC and the Pembina Institute (2013) Priced Out: Understanding the Factors Affecting Home Prices in the GTA. Canada: Royal Bank of Canada and the Pembina Institute.

38.

Rothman

Buliung

, et al. (2015) Associations between parents’ perception of traffic danger, the built environment and walking to school. Journal of Transport & Health 2(3): 327–335.

39.

Rundle

Bader

Richards

, et al. (2011) Using Google Street View to audit neighborhood environments. American Journal of Preventive Medicine 40(1): 94–100.

40.

Salesses

Schechtner

Hidalgo

(2013) The collaborative image of the city: Mapping the inequality of urban perception. PLoS One 8(7).

41.

Searle

Filion

(2011) Planning context and urban intensification outcomes: Sydney versus Toronto. Urban Studies 48(7): 1419–1438.

42.

Shenassa

Liebhaber

Ezeamama

(2006) Original contribution perceived safety of area of residence and exercise: A Pan-European study. American Journal of Epidemiology 163(11): 1012–1017.

43.

Simmins G (2011) Urban and regional planning. In: The Canadian Encyclopedia. Historica Canada. Available at: https://www.thecanadianencyclopedia.ca/en/article/urban-and-regional-planning (accessed 10 February 2017).

44.

Stafford

Chandola

Marmot

(2007) Association between fear of crime and mental health and physical functioning. American Journal of Public Health 97(11): 2076–2081.

45.

Stathi

Gilbert

Fox

, et al. (2012) Determinants of neighborhood activity of adults age 70 and over: A mixed-methods study. Journal of Aging and Physical Activity 20(2): 148–170.

46.

Tan

Sun

Kong

, et al. (2018) A survey on deep transfer learning. CoRR, abs/1808.01974.

47.

Vanwolleghem

Ghekiere

Cardon

, et al. (2016) Using an audit tool (MAPS global) to assess the characteristics of the physical environment related to walking for transport in youth: Reliability of Belgian data. International Journal of Health Geographics 15(1): 41.

48.

Wainwright

NWJ

Surtes

(2003) Places, people, and their physical and mental functional health. Journal of Epidemiology & Community Health 58: 333–339.

49.

Weich

Blanchard

Prince

, et al. (2002) Mental health and the built environment: Cross-sectional survey of individual and contextual risk factors for depression. The British Journal of Psychiatry: The Journal of Mental Science 180: 428–433.

50.

Whitley

Prince

(2005) Fear of crime, mobility and mental health in inner-city London, UK. Social Science & Medicine (1982) 61(8): 1678–1688.

51.

Will

Young

(2015) The politics of place: Place-making versus densification in Toronto’s tower neighbourhoods. International Journal of Urban and Regional Research 39(3): 613–621.

52.

Wilson

Kelly

Schootman

, et al. (2012) Assessing the built environment using omnidirectional imagery. American Journal of Preventive Medicine 42(2): 193–199.

53.

Winters

Davidson

Kao

, et al. (2010) Motivators and deterrents of bicycling: Comparing influences on decisions to ride. Transportation 38(1): 153–168.

54.

Wooldridge

(2002) Econometric Analysis of Cross Section and Panel Data. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press.

55.

Won

Lee

Forjuoh

, et al. (2016) Neighborhood safety factors associated with older adults’ health-related outcomes: A systematic literature review. Social Science & Medicine (1982) 165: 177–186.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.76 MB