Health & Place 66 (2020) 102428

Contents lists available at ScienceDirect

Health and Place

journal homepage: http://www.elsevier.com/locate/healthplace

Objective scoring of streetscape walkability related to leisure walking:

Statistical modeling approach with semantic segmentation of Google Street
View images
Shohei Nagata a, *, Tomoki Nakaya a, Tomoya Hanibuchi a, Shiho Amagasa b, Hiroyuki Kikuchi b,
Shigeru Inoue b
a
Graduate School of Environmental Studies, Tohoku University, 468-1 Aoba, Aramaki, Aoba-ku, Sendai, 980-0845, Japan
b
Department of Preventive Medicine and Public Health, Tokyo Medical University, 6-1-1 Shinjuku, Shinjuku-ku, Tokyo, 160-8402, Japan

A R T I C L E I N F O A B S T R A C T

Keywords: Although the pedestrian-friendly qualities of streetscapes promote walking, quantitative understanding of
Neighborhood walkability streetscape functionality remains insufficient. This study proposed a novel automated method to assess street­
Walking behavior scape walkability (SW) using semantic segmentation and statistical modeling on Google Street View images.
Google street view
Using compositions of segmented streetscape elements, such as buildings and street trees, a regression-style
Deep learning
model was built to predict SW, scored using a human-based auditing method. Older female active leisure
Semantic segmentation
walkers living in Bunkyo Ward, Tokyo, are associated with SW scores estimated by the model (OR = 3.783; 95%
CI = 1.459 to 10.409), but male walkers are not.

1. Introduction abundance of data and the ready availability of GIS methods. Although it
has been proven in many contexts that micro-scale walkability, as
Physical inactivity is a leading risk factor for non-communicable determined by pedestrian perception, including the attractiveness of
diseases, including coronary heart disease, type 2 diabetes, and breast streetscape and the condition of the sidewalks, is effective for promoting
and colon cancers (Lee et al., 2012; World Health Organization, 2017). walking (Cain et al., 2014; Ewing et al., 2016; Kim et al., 2014),
Guthold et al. (2018) reported that 27.5% of the total population measuring these attributes with traditional methods (audits/systematic
worldwide are insufficiently active. Walking is a physical activity that social observation) requires time and resources (Duncan et al., 2018).
most can do relatively easily (Handy et al., 2002) and contributes to risk Thus, it is difficult to use traditional methods to assess micro-scale
reduction of chronic disease, including cardiovascular disease (Hu et al., walkability over a large area.
2001), type 2 diabetes (Hu et al., 1999), and hypertension (Williams and In recent years, several studies have effectively used Google Street
Thompson, 2013). It has been found that the built environment of one’s View (GSV) to evaluate micro-scale walkability, a method that achieves
neighborhood, including distances to non-residential destinations, street cost reduction, remote evaluation of neighborhoods over a wide range of
connectivity, and the condition of pedestrian infrastructure, affects global contexts, and automated assessment (Rzotkiewicz et al., 2018).
walking behavior (Christiansen et al., 2016; Saelens and Handy, 2008; Pliakas et al. (2017) demonstrated that GSV-based audits can provide
Sugiyama et al., 2014). Therefore, understanding the construction of significant reduction in time costs compared with foot-based audits
walkable or unwalkable places is crucial for urban planners and while maintaining audit quality. In addition, Hanibuchi et al. (2019)
epidemiologists. developed a simple checklist for virtual audits of streetscape walkability
It has been well documented that macro-scale walkability, based on (SW) and proved inter-source (between in-person and virtual audits) and
objective measurements of land use mix, street connectivity, and dis­ inter-rater (between two trained auditors and between trained auditors
tance to facilities, is related to the choice of walking as a mode of and untrained crowd-sourced workers) reliability for GSV-based evalu­
transportation (Owen et al., 2007; Saelens and Handy, 2008) due to the ation. In the automated approach, several studies have extracted

* Corresponding author.
E-mail addresses: shohei.nagata.r7@dc.tohoku.ac.jp (S. Nagata), tomoki.nakaya.c8@tohoku.ac.jp (T. Nakaya), info@hanibuchi.com (T. Hanibuchi), amagasa@
tokyo-med.ac.jp (S. Amagasa), kikuchih@tokyo-med.ac.jp (H. Kikuchi), inoue@tokyo-med.ac.jp (S. Inoue).

https://doi.org/10.1016/j.healthplace.2020.102428
Received 29 February 2020; Received in revised form 19 July 2020; Accepted 18 August 2020
Available online 22 September 2020
1353-8292/© 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
S. Nagata et al. Health and Place 66 (2020) 102428

Table 1 “enclosure,” measured by the proportion of street wall or sky, indicates

Participant characteristics by sex. an area’s room-like quality that can make pedestrians feel comfortable
Overall Male Female and secure (Ewing, 1996). “Human scale” is evaluated by the proportion
of buildings whose first floors have windows, the number of small
n = 312 n = 164 n = 148
planters, and other similar factors presenting the size, texture, and
Walking for leisure articulation of physical elements corresponding to the speed at which
≥150 min/ n (%) 98 61 37
(31.4%) (37.2%) (25.0%)
humans walk. Although multiple components of streetscapes influence
<150 min/week n (%) 214 103 111 walkability, the number of streetscape components found by automated
(68.6%) (62.8%) (75.0%) methods and their relationship to walking behavior are still limited.
min/week Mean/ 127.3/ 150.2/ 101.9/ Wang et al. (2019) produced an automatic classification of cityscape
SD 188.8 208.2 161.7
images including multiple components such as trees and grasses into six
Living arrangements types (wealthy, safe, lively, depressing, boring, and beautiful) based on
With others n (%) 252 143 109 human perception and clarified the relationship between each category
(80.8%) (87.2%) (73.6%)
Alone n (%) 59 20 39
of perception and types of associated physical activity. However, how
(18.9%) (12.2%) (26.4%) these streetscape components, such as vegetation, affect walkability or
Missing n (%) 1 (0.3%) 1 (0.6%) 0 (0.0%) walking behavior has not been sufficiently discussed. Moreover,
Working status although several studies have indicated that micro-scale walkability,
Working with income n (%) 111 76 35 such as sidewalk conditions and the streetscape aesthetics, affects
(35.6%) (46.3%) (23.6%) physical activity for leisure (De Bourdeaudhuij et al., 2005; Inoue et al.,
Not working n (%) 201 88 113
2010; Saelens and Handy, 2008; Witten et al., 2012), little attention has
(64.4%) (53.7%) (76.4%)
been paid to the relationship between SW, as determined from GSV
Educational attainment images, and leisure walking.
≥13 years n (%) 175 109 66
This study examines the relationships among multiple streetscape
(56.1%) (66.5%) (44.6%)
<13 years n (%) 137 55 82 components, micro-scale walkability, and older people’s leisure walking
(43.9%) (33.5%) (55.4%) using deep learning and a statistical approach to GSV images. A sys­
Car driving tematic review describes the positive relationship between a neighbor­
Routine car driving n (%) 64 59 5 (3.4%) hood’s physical environment and older people’s physical activity (Yen
(20.5%) (36.0%) et al., 2010). Thus, management to ensure pedestrian-friendly neigh­
Not routine n (%) 243 101 142 borhoods is important to promote older people’s well-being. The study
(77.9%) (61.6%) (95.9%)
consists of two phases: building a prediction model for the SW scores
Missing n (%) 5 (1.6%) 4 (2.4%) 1 (0.7%)
based on the evaluation of streetscapes by manual audits and the
Physical limitation
component elements of the streetscape, and analyzing the relationships
Limited (somewhat, quite a n (%) 74 30 44
lot, could not do physical (23.7%) (18.3%) (29.7%) between leisure walking by older people and neighborhood SW. In the
activity) first phase, we extract component elements of the streetscape from GSV
Not limited (not at all, very n (%) 232 131 101 images using a semantic image segmentation method of deep learning.
little) (74.4%) (79.9%) (68.2%) The segmented components are used as independent variables in the
Missing n (%) 6 (1.9%) 3 (1.8%) 3 (2.0%)
prediction model for SW scores. In the second phase, we predict SW
Age Mean/ 74.3/2.9 74.3/3.0 74.3/2.8 scores for each intersection in Bunkyo ward in central Tokyo and
SD
analyze the relationship between the SW score and older people’s
walking for leisure. This study contributes to the evaluation of the
component elements of streetscapes through assessment of images from effectiveness of the automated assessment of neighborhood environ­
GSV and similar services (e.g., Tencent Street View) using machine ments by using quantitative methods to examine the relationship be­
learning or deep learning methods. These studies analyzed the re­ tween older people’s leisure walking and micro-scale walkability based
lationships among each component element to SW, physical activity, or on multiple components of the streetscape.
walking behavior (Lu, 2018; Villeneuve et al., 2018; Wang et al., 2019;
Yin et al., 2015; Yin and Wang, 2016). For example, methods of 2. Methods
detecting pedestrians from GSV images with deep learning showed
acceptable accuracy for automated audits (Yin et al., 2015). In studies 2.1. Data collection
that examined the relationship between specific component elements of
the streetscape extracted from GSV and SW or walking behavior, Yin and 2.1.1. SW score using manual audits
Wang (2016) ascertained a significant correlation between visual To build an SW score prediction model, the results of an SW evalu­
enclosure, determined by sky visibility, and WalkScore® or pedestrian ation by manual audits using GSV were used as the dependent variable.
volume. Further, Lu (2018) found that neighborhood greenness is The evaluation was based on a checklist developed by Hanibuchi et al.
significantly related to walking behavior. Cai et al. (2018) published an (2019) and is related to SW as regards physical conditions, safety, and
open-source library called “Treepedia” that allows calculation, using aesthetics. This checklist was developed to be simple and easy to use for
GSV images, of the amount of vegetation cover along a street, and Vil­ practical purposes (i.e., reducing time cost and enabling many untrained
leneuve et al. (2018) showed the relationship between promoting auditors to participate over large study areas) and included only 14
summer leisure physical activity and neighborhood greenness calculated items with dichotomous responses. The checklist was reported to have
by Treepedia. good inter-rater and inter-source reliability (Hanibuchi et al., 2019). The
However, several issues continue to have importance to fulfill 14 items included the presence or absence of the following features:
automated assessment study based on streetscape images and machine/ sidewalk, wide sidewalk, obstructions, steep slopes, street parking,
deep learning approach. Previous studies have proposed protocols, such heavy traffic, heavy foot traffic, crosswalk, traffic mirrors, streetlights,
as imageability, enclosure, human scale, transparency, and complexity, street trees, attractive streetscapes, graffiti and litter, and abandoned
as quantitatively measurable aspects of urban design for walkability buildings. These were selected by considering frequency of use in
(Ewing et al., 2006, 2016; Ewing and Handy, 2009). For instance, existing audit tools, multiple aspects of micro-scale streetscape, and the
Asian and Japanese urban context. Each item is evaluated with a binary

2
S. Nagata et al. Health and Place 66 (2020) 102428

Fig. 1. Example of the API request to obtain GSV images.

Sources: Map: Esri, HERE, Garmin, FAO, NOAA, USGS, © OpenStreetMap contributors, and the GIS User Community. Photo: Author’s photo.

value (Yes/No) and converted into a numerical point. In the conversion including Bunkyo-ward. The details of the questionnaire are given in
process, for items assumed pedestrian-friendly (i.e., presence of side­ different publications (Amagasa et al., 2019; Inoue et al., 2011; Kikuchi
walk, wide sidewalk, crosswalk, traffic mirrors, streetlights, street trees, et al., 2018). The questionnaire includes questions on weekly frequency
and attractive streetscape), “Yes” is valued at 1 point, and “No” at 0. For of walking (days/week) and average walking duration each day
items assumed to create a pedestrian-unfriendly environment (i.e., (min/day) for the purposes of daily errands, leisure, commuting to work,
presence of obstructions, steep slopes, street parking, heavy traffic, working, etc. The walking measure’s validity based on the questionnaire
heavy foot traffic, graffiti and litter, and abandoned buildings), “Yes” is was ascertained by Spearman correlation coefficient between walking
valued at 0, and “No” at 1. The SW score at each assessed location is the time and step counts per day assessed by an accelerometer (R = 0.3; p <
sum of all items’ points, and the maximum score is 14. Using the 0.001) (Inoue et al., 2010). Participants were also asked about their
checklist, from August to October 2018, we assessed the streetscape in lifestyles and health conditions such as living arrangements, physical
Bunkyo-ward, Tokyo, and the surrounding neighborhoods. After initial limitations, and perception of neighborhood environments. In this
instruction with brief material provided by Hanibuchi et al. (2019) and study, we calculated participants’ walking time for leisure from the
repeated consultation, a trained auditor (a freelance research assistant) frequency (days/week) and duration (min/day) of leisure walking. We
evaluated 2842 street segments connected to 854 intersections using defined participants as “active leisure walkers” if they walked more than
“walking through” target streets on GSV. These target intersections were 150 min/week for leisure, referring to a guideline by Nelson et al.
randomly selected, and they amounted to 10% of all intersections in the (2007). Inoue et al. (2011) described the sampling methods they used in
study area. We calculated the means of the scores for the connecting this survey. In 2010, the authors of this study used stratified random
streets with the intersection as the SW score assessed by the manual sampling, selecting a total of 900 residents in Bunkyo-ward from a
audit. residential registry to acquire the baseline data. In 2015, we conducted a
follow-up survey of 373 respondents who had participated in the base­
2.1.2. Walking time data line survey in 2010 and agreed to enroll in the follow-up survey in 2015.
For the walking time source data, we used results of a questionnaire Of these, 312 residents responded, and their data were used for this
survey conducted to analyze the relationship between the neighborhood study (Table 1). We used information provided on age, educational
environment and the health of older people in three locations in Japan, attainment, living arrangements, working status, car driving, and

3
S. Nagata et al. Health and Place 66 (2020) 102428

Fig. 2. Structure of encoder–decoder networks of DeepLab v3+ (Chen et al., 2018).

physical limitations, in addition to weekly frequency of walking and the intersection, we used ArcGIS Geo Suite Douromo (road network)
average walking duration each day for leisure. (Esri Japan Inc.). The digital elevation model (DEM) (Geospatial Infor­
Before the survey in 2015, ethical approval for this study was ob­ mation Authority of Japan) was used to obtain elevation data to calcu­
tained from the Tokyo Medical University Ethics Committee (No. 2898). late slope of the road segments and to build a prediction model for the
Before completing the questionnaire, all participants signed an informed SW score. Each intersection’s slope value was calculated as follows:
consent form. First, road segments connected from an intersection were split into links
at every vertex; second, the elevation value of DEM was given to the
2.1.3. Streetscape images vertices; third, the slope was calculated by the elevation value of the
This study estimated the SW score based on the streetscape images vertices and the length of the link between the vertices; fourth, the mean
taken at intersections in Bunkyo-ward. Because the SW score assessed by of the slope values was calculated for the road segment, weighted by the
manual audit was summarized at intersections to which streets connect, length of the links; and fifth, the steps above were repeated for every
the estimation of the SW score using GSV images at the intersection can intersection. The National Land Numerical Information (Ministry of
simplify the calculation process. Furthermore, because the distance be­ Land, Infrastructure, Transport and Tourism, Japan) and telephone
tween intersections is short in the study area (mean distance = 45.6 m), directory data as of June 2017 (Eins Inc.) were used as destination data
it is possible to evaluate the SW there using GSV images from in­ to calculate objective walkability to analyze the relationship between
tersections. We requested GSV images for 5321 intersections using leisure walking and SW. Moreover, population density, based on the
Street View Static API (Google Inc.) and collected 17,276 images for population census of 2015 (Statistics Bureau of Japan) was used for the
5317 intersections. GSV images were taken from 2009 to 2019 and more objective walkability score as well. Objective walkability was defined as
than 90% after 2016. These intersections are at the same locations where the sum of z scores for population density, intersection density, and
SW evaluation was completed by manual audits or covered in 1000 m variety of destinations, all calculated in 1000 m network buffers from
network buffers from the residence of the participants in the follow-up participants’ residences, in accordance with a previous study (Kikuchi
survey. At each intersection, we obtained street images in each direc­ et al., 2018). We defined the variety of destinations as the sum of z scores
tion toward the intersections to which it connects. The procedure for for the number of destination facilities (station, post office, elementary
obtaining the GSV images for each intersection was as follows: obtain school, community center, bank, bookstore, convenience store, restau­
latitude and longitude for the intersection; calculate the bearing rant, supermarket, department store, and sports and fitness club) in the
(heading) from the intersection to the street; set API parameters, 1000 m network buffer. We used ArcGIS Pro 2.3.2 (Esri Inc.) to conduct
including location and the bearing; and download outdoor images. An all GIS processing.
example of this procedure and its result is shown for the Koishikawa­
korakuen-iriguchi intersection in Fig. 1. The latitude and longitude were 2.2. Streetscape segment detection
set to 35.704824 and 139.747541, and the bearings in the A, B, and C
directions were 3.23, 110.89, and 183.55, respectively. Images showing To detect the component elements of each intersection’s streetscape,
each street from the intersection were obtained as a response to the API we used DeepLab v3+ (Chen et al., 2018), a deep learning model
request. developed for semantic image segmentation. DeepLab v3+ architecture
characteristically adopts atrous convolution in encoder–decoder net­
2.1.4. GIS data and software works (Chen et al., 2018), which are used in semantic image segmen­
In this study, published GIS data were used as data sources for road tation models such as U-Net (Ronneberger et al., 2015) and SegNet
segments and geographic variables to construct the SW score prediction (Badrinarayanan et al., 2017). Typically, pooling and convolution pro­
model and analyze the relationships between leisure walking and the cesses included in an encoder module reduce the resolution of feature
SW. For the source of data on the network segments and the locations of maps and obtain high semantic information using filters based on an

4
S. Nagata et al. Health and Place 66 (2020) 102428

Table 2 Table 3
Definition of segment classes provided by Cityscapes Dataset. Detected segments from all obtained GSV images.
Group Class Description Segments % Segments %

Flat Road Area where cars usually drive, e.g., lanes, directions, Building 45.00 Person 0.74
and streets. Areas only delimited by markings from Road 23.49 Bicycle 0.50
the main road are also roads, e.g., bicycle lanes, Vegetation 9.94 Traffic sign 0.34
roundabout lanes, or parking spaces. This label does Sky 6.65 Terrain 0.24
not include curbs. Fence 3.02 Train 0.10
Sidewalk Area located at the side of a road and delimited from Wall 2.83 Bus 0.03
the road by some obstacle for pedestrians or cyclists, Sidewalk 2.29 Motorcycle 0.03
e.g., curbs or poles (perhaps small). This label Car 2.26 Rider 0.03
includes a possibly delimiting curb, traffic islands Truck 1.26 Traffic light 0.01
(the walkable part), or pedestrian zones (where Pole 1.24
usually cars are not allowed during daytime).

Human Person A human walking, standing, or sitting on the ground,

on a bench, or on a chair. This class includes toddlers
et al., 2018). Fig. 2 shows the structure of the encoder–decoder networks
and someone pushing a bicycle or standing next to it in DeepLab v3+ (Chen et al., 2018). In the encoder module, Atrous
with both legs on the same side of the bicycle. This Spatial Pyramid Pooling (ASPP) investigates the convolutional features
class also includes anything carried by the person, e. at multi-scales using atrous convolution with different rates (A) and
g., backpack, but not items touching the ground, e.g.,
generates a feature map including abundant semantic information (B).
trolleys.
Rider A human using some device to move, e.g., riders/ In the decoder module, the feature map (B) is up-sampled four times by
drivers of bicycles, motorbikes, scooters, bilinear sampling and concatenated with low-level features from the
skateboards, horses, roller-blades, wheel-chairs, road network backbone to recover spatial information (C). The feature map
cleaning cars, cars without a roof. Note that a visible
before processing ASPP is reused as a low-level feature to improve
driver of a car with a roof can be seen only through
the window. Since holes are not labeled, the human is
spatial information’s recovery accuracy. Finally, predicted segmenta­
included in the car label. tion results are exported after refinement by convolution processes and
four times bilinear up-sampling (D).
Vehicle Car Automobile, jeep, SUV, van with continuous body
shape, caravan, but no trailers. In this study, we used DeepLab v3+ trained on streetscapes using
Truck Truck, box truck, pickup truck, including their Cityscapes Dataset (Cordts et al., 2016). The model is published at
trailers. Back part/loading area is physically “Supervise.ly” (Deep System Inc.). The Cityscapes Dataset is an image
separated from the driving compartment.
dataset with annotation of streetscape segments. Annotation is defined
Bus Bus for 9+ persons, public transport, or long distance
transport.
as 30 classes based on groups recognized as streetscape components
Train Vehicle on rails, e.g., tram, train. such as flat, human, vehicle, construction, object, nature, sky, and void.
Motorcycle Motorbike, moped, scooter without a driver (a The dataset provides 19 classes for training (Table 2), while the other 11
rider—see above). classes are excluded from the dataset due to rare segments in street­
Bicycle Bicycle without a driver (a rider—see above).
scapes (Cordts et al., 2016). Regarding model validation, the agreement
Construction Building Building, skyscraper, house, bus stop building, of segment recognition between the model and a human at random lo­
garage, car port.
cations of the GSV images was 82% (kappa coefficient = 0.76, p <
Wall Individual standing wall. Not part of a building.
Fence Fence including any holes. 0.001) (see Appendix).
Every pixel of all GSV images was classified into one of 19 segments
Object Pole Small, mainly vertically oriented pole or having a
diameter (in pixels) of at most twice the diameter of
by DeepLab v3+. We calculated the percentages of segments for each
the pole, e.g., sign pole, traffic light poles, intersection.
streetlights.
Traffic sign A sign without a pole, showing information of the 2.3. Building the SW score prediction model
driver/cyclist/pedestrian in an everyday traffic
scene, e.g., traffic signs, parking signs, direction
signs. This label counts only the front side of a sign To build the model based on the linear relationship between SW and
containing information. No ads/commercial signs. streetscape components, we used a regression analysis on all the SW
Traffic A traffic light box without its poles. scores from manual audits as a dependent variable and the percentage of
light streetscape segments detected by the deep learning model, slope, and
Nature Vegetation Tree, hedge, all kinds of vertical vegetation. Plants road width as independent variables. All variables were calculated for
attached to buildings are usually not annotated each intersection. The road width was given in ranges (3–5.5, 5.5–13,
separately and are labeled “building” as well. If
and 13 m or more) for each road segment attribute. For the value of the
growing at the side of a wall or building, it is marked
as vegetation if it covers a substantial part of the road width, we chose the largest width values for the streets connected
surface (more than 20%). to the intersection. For the model selection, first, the stepwise method
Terrain Grass, all kinds of horizontal vegetation, soil, or sand. was applied to select the independent variables; second, the genetic
This label includes a possibly delimiting curb. algorithm (GA) (Holland, 1975) was used to select the model with
Sky Sky Open sky, without tree leaves. Includes thin electrical interaction effects based on the Akaike Information Criterion (AIC) to
wires visible in skyscape. avoid overfitting; and lastly, the final model was calibrated as a spatial
Note: This table is retrieved from the website of Cityscapes Dataset. More simultaneous autoregressive error (SAR) model (Anselin, 2003). In cal­
detailed information is available at the website (https://www.cityscapes-dataset ibrating the final model, we adopted Gabriel graph neighbors (Matula
.com/). and Sokal, 1980) to create the adjacency matrix between the in­
tersections. All processes for the regression analysis were completed in R
object’s feature to be extracted; a decoder module gradually recovers 3.6.1. The glmulti library was used for model selection including the
spatial information lost during encoding processes (Chen et al., 2018). stepwise and GA to select statistically meaningful variables and inter­
While retaining the feature map’s resolution calculated by Deep Con­ action effects, and the spdep library was used for spatial regression
volutional Neural Networks (DCNN), atrous convolution can capture modeling.
multi-scale information by adjusting the filter’s field-of-view (Chen

5
S. Nagata et al. Health and Place 66 (2020) 102428

Fig. 3. Examples of semantic segmentation results.

Sources: Author’s photo and Supervise.ly.

2.4. Relationship between leisure walking and SW the segments in the GSV images was calculated. Fig. 3 shows the results
of segmentation at the same intersection as Fig. 1. For instance, in the B
To ascertain whether any relationship between leisure walking and direction, the largest segment shows vegetation (45.32%), the second is
neighborhood SW exists, we estimated the odds ratios (ORs) and 95% the road segment (32.55%), and the third is the building segment
confidence intervals (CIs) of the neighborhood SW using the logistic (9.92%). Likewise, for the mean value of the segments at the intersec­
regression analysis. The dependent variable was defined as “1” if the tion, the largest segment consists of vegetation (48.39%) (Fig. 3). By
participant was an active leisure walker or as “0” otherwise. The comparison with the mean values for the segments in all images
neighborhood SW was used as the independent variable. As control (Table 3), the streetscape at the intersection shows abundant vegetation.
variables, age, physical limitation, educational attainment, living ar­ We excluded 1313 unsuitable images from the streetscape evaluation
rangements, working status, routine car driving, and objective walk­ due to massive inclusion of the platform of a subway station, the exterior
ability were selected by referring to previous studies (Amagasa et al., of a house, or other unsuitable items. The criterion for this exclusion was
2019; Inoue et al., 2011; Kikuchi et al., 2018). Because several studies road segment less than 5%. After this, the mean values for the proportion
have shown that physical activity data features different trends per segment were calculated at 5293 intersections.
depending on the walker’s sex (Amagasa et al., 2017; Caspersen et al.,
2000), regression models were estimated for each sex. In calculating the
neighborhood SW, first, the SW scores for all intersections within 1000 3.2. Building a prediction model for SW score
m network buffers from the participant residence were predicted with
the SW score prediction model; next, we calculated the median values Table 4 shows the optimal combination of independent variables,
for the predicted SW score in 500 or 1000 m as values for neighborhood including interaction effects, based on stepwise GA–SAR modeling, and
SW in 500 m or 1000 m. Adjusted p-values were calculated by Bonfer­ we used this as the prediction model for the SW score (R2 = 0.51; MAE =
roni adjustment (Bland and Altman, 1995) to control Type I Error 0.66). Fig. 4 shows the streetscapes predicted by the model to have high
because we compared the four models (two sex groups × two neigh­ SW scores. These locations are well-designed and well-maintained
borhood definitions of the SW in 500 m and 1000 m). streetscapes and include sophisticated road facilities, such as wide
sidewalks or trees lining the street. The locations for which low SW
3. Results scores were predicted (Fig. 5) consist of pedestrian-unfriendly street­
scape components such as narrow and dark streets, no sidewalks, or
3.1. Streetscape segment detection steep slopes.
From the model results (Table 4), nine types of segments (road,
The deep learning model was applied to all images (n = 17,276), and building, sky, terrain, pole, sidewalk, vegetation, traffic light, and rider),
all of the 19 segments were detected (Table 3). Then, the percentage of including their interaction effects, are associated with SW scores. For a
streetscape segment with a positive coefficient, the estimated SW score

6
S. Nagata et al. Health and Place 66 (2020) 102428

Table 4 × 0.047). In contrast, a streetscape segment with a negative coefficient

Summary of the SW score prediction model. decreases the estimated SW score. For instance, since a sky segment has
β 95% CI p a negative coefficient (β = − 0.153; 95% CI = − 0.209 to − 0.096), the
existence of the sky segment relates to a decreased SW. When the sky
(Intercept) 8.297 7.211 to 9.382 <2.2e-16
Road: Building − 0.003 − 0.004 to − 0.002 <0.001*** segment occupies 30% of pixels in the GSV images, the sky segment
Road 0.147 0.098 to 0.195 <0.001*** causes a decreased SW score at the intersection by 4.59 points (30 ×
Building: Road Width 5.5–13 m − 0.033 − 0.045 to − 0.021 <0.001*** − 0.153).
Sky − 0.153 − 0.209 to − 0.096 <0.001*** The member, as the combination of two streetscape segments shown
Road: Vegetation − 0.002 − 0.002 to − 0.001
in Table 4, is the interaction effect when these segments in the combi­
<0.001***
Building: Road Width 3–5.5 m − 0.030 − 0.043 to − 0.017 <0.001***
Slope: Traffic light 3.740 2.147 to 5.333 <0.001*** nation appear in the GSV image. The interaction effect between a road
Terrain: Rider − 1.712 − 2.580 to − 0.845 <0.001*** segment and a terrain segment is positively associated with SW (β =
Building 0.047 0.022 to 0.071 <0.001*** 0.006; 95% CI = 0.002 to 0.011). When the road segment occupies 30%
Road: Road Width 5.5–13 m 0.029 0.013 to 0.046 <0.001***
of pixels and the terrain segment occupies 5% of pixels of GSV images,
Pole − 0.864 − 1.368 to − 0.360 <0.001***
Pole: Vegetation 0.013 0.005 to 0.022 0.002** the interaction of the two segments increases the SW score at the
Slope: Rider 0.584 0.172 to 0.995 0.005** intersection by 2.85 points (30 × 5 × 0.019). The interaction effect
Building: Pole 0.011 0.003 to 0.018 0.006** between the road segment and the building segment is negatively
Pole: Sky 0.022 0.006 to 0.038 0.008** associated with SW (β = − 0.003; 95% CI = − 0.004 to − 0.002).
Road: Sidewalk 0.006 0.002 to 0.011 0.008**
Building: Sky 0.002 0.000 to 0.003 0.008**
Regarding the variable other than streetscape segments, slope is
Road: Terrain 0.019 0.002 to 0.035 0.026* marginally associated with a decrease in SW score (β = − 0.239; 95% CI
Road: Rider − 0.061 − 0.116 to − 0.006 0.029* = − 0.478 to 0.000). When the mean value of the slope of the street
Building: Slope 0.003 0.000 to 0.006 0.039* connected from an intersection is 5%, the SW score decreases by 1.195
Pole: Rider 0.815 0.032 to 1.597 0.041*
points (5 × − 0.239). The interaction effects between a building and road
Sidewalk: Terrain 0.059 0.002 to 0.116 0.043*
Slope − 0.239 − 0.478 to 0.000 0.050. width 5.5–13 m or a building and road width 3–5.5 m relate to a
Terrain − 0.459 − 0.928 to 0.01 0.055. decrease in SW score (β = − 0.033; 95% CI = − 0.045 to − 0.021, and β =
Sidewalk: Rider 0.297 − 0.033 to 0.627 0.078. − 0.030; 95% CI = − 0.043 to − 0.017). When the building segment oc­
Sky: Slope 0.009 − 0.001 to 0.020 0.091. cupies 30% of pixels of the GSV images taken at an intersection with the
Road: Slope − 0.005 − 0.012 to 0.001 0.121
road width of 5.5–13 m, the interaction effect of the two variables de­
Pole: Slope 0.026 − 0.008 to 0.060 0.128
Sidewalk − 0.092 − 0.220 to 0.035 0.157 creases the SW score at the intersection by 0.99 points (30 × 1 ×
Sidewalk: Road Width 5.5–13 m 0.045 − 0.038 to 0.128 0.286 − 0.033). In case of the road width of 3–5.5 m, the SW score decreases by
Sidewalk: Road Width 3–5.5 m − 0.037 − 0.123 to 0.049 0.404 0.9 points (30 × 1 × − 0.03). While the building segment positively
Road: Road Width 3–5.5 m 0.003 − 0.017 to 0.024 0.753
relates to the SW, the influence of the building segment on the SW will
be small on a narrow street.
R2 0.51
MAE 0.66
AIC 2198
3.3. Relationship between walking time and SW
λ 0.14 LR test value: 8.381, p value: 0.004

Note: CI: confidence interval; MAE: Mean Absolute Error; AIC: Akaike Infor­ Table 5 presents the estimated ORs and 95% CI of the neighborhood
mation Criterion; and λ: Simultaneous autoregressive error coefficient. ’***’, SW for active leisure walkers. The neighborhood SW in 500 m is asso­
’**’, ’*’, and ’.’ denote statistical significance at the 0.1%, 1% 5% and 10%
ciated with female active leisure walkers, and female participants more
levels, respectively.
actively walked for leisure if the neighborhood SW is high (OR = 3.783;
95% CI = 1.459 to 10.409). When the neighborhood SW in 500 m in­
increases when the percentage of the segment in the GSV image in­ creases by one point, the odds of older females walking more than 150
creases. For instance, since a building segment has a positive coefficient min/week for leisure are 3.78 times higher. There is no relationship
(β = 0.047; 95% CI = 0.022 to 0.071), the existence of the building between the neighborhood SW in 500 m and male active leisure walkers,
segment increases the SW. When the building segment occupies 30% of and the neighborhood SW in 1000 m is not associated with both male
pixels in the GSV images taken at an intersection, the building segment and female active leisure walkers.
causes an increase in the SW score at the intersection by 1.41 points (30 We applied multilevel logistic regression models with a random

Fig. 4. Examples of locations with high predicted SW scores.

Source: Author’s photo at the same location as the GSV images used for the analysis.

7
S. Nagata et al. Health and Place 66 (2020) 102428

Fig. 5. Examples of locations with low predicted SW score.

Source: Author’s photo at the same location as the GSV images used for the analysis.

prediction model for the SW score that can evaluate micro-scale walk­
Table 5 ability in relation to streetscape images and examined the relationship
Estimated odds ratios and 95% confidential intervals of the neighborhood SW between older females walking for leisure and predicted neighborhood
for active leisure walkers.
SW.
OR 95% CI p Adjusted Through the construction of the prediction model for SW scores, the
p relationships that we found between each component of the streetscape
Male (N = 156) SW 500 m 1.534 0.696 to 0.290 1.000 and micro-scale walkability support quantitative evaluation for pedes­
3.438 trian-(un)friendly streetscapes. Previous studies attempted extracting
SW 1000 0.810 0.228 to 0.739 1.000
features associated with walkable urban design using objective or
m 2.787
Female (N = SW 500 m 3.783 1.459 to 0.007** 0.028* automated methods quantitively to understand pedestrian-friendly
144) 10.409 streetscapes (Purciel et al., 2009; Yin, 2017; Yin and Wang, 2016).
SW 1000 3.373 0.747 to 0.110 0.440 Purciel et al. (2009) translated urban design variables into GIS mea­
m 15.277 sures, but there are still no data for several types of features, such as the
Notes: OR: odds ratio; CI: confidence interval; and Adjusted covariates are age, proportion of sky and number of small planters. Furthermore, Yin and
educational attainment, living arrangements, working status, routine car Wang (2016) demonstrated an automated measurement of proportion of
driving, physical limitation, and objective walkability. Participants who had at sky using a machine learning approach, and Yin (2017) suggested that
least one missing variable were excluded. Active leisure walker is defined as a GSV and computer vision can assist in the evaluation of urban design.
participant who walks for leisure more than 150 min/week. Adjusted p-values This study supports the assertion that GSV and the computer vision
were calculated by Bonferroni adjustment. ’***’, ’**’, ’*’, and ’.’ denote statis­
approach can produce useable results, including semantic image seg­
tical significance at the 0.1%, 1% 5% and 10% levels, respectively.
mentation performed by deep learning approach, for walkable urban
design.
intercept to the data using the sampling neighborhood unit (chocho-aza) Supporting the results of Ewing and Handy (2009) and Yin and Wang
to consider a possible clustering tendency due to the area stratified (2016) on visual enclosure, the results of our regression show that
sampling design; however, it did not improve the model performance building segments lead to increases in SW, and sky segments are nega­
and produced the same estimates of coefficients and their standard tively associated with SW. This suggests that locations enclosed by
errors. buildings increase pedestrian comfort, unlike streets with low-rise
buildings, such as those found in suburban residential area. However,
4. Discussion the model also indicates that several interaction effects, including
building segment, have negative effects on the SW. For instance, inter­
This is the first study to examine the relationship among leisure action effects between road and building lead to decreases in SW. It is
walking, micro-scale walkability, and multiple components of street­ conceivable that locations mostly occupied by road segments and
scape using GSV images and a deep learning approach. We built a building segments, such as shown in Fig. 6, are situated on highly

Fig. 6. Examples of locations mostly occupied by road and building segments.

Source: Author’s photo at the same location as the GSV images used for the analysis.

8
S. Nagata et al. Health and Place 66 (2020) 102428

Fig. 7. Examples of locations with nature-related segments.

Source: Author’s photo at the same location as the GSV images used for the analysis.

enclosed streets, and such locations could be dark and make pedestrians walkability scores based on destinations’ density and accessibility,
feel unsafe. A highly enclosed streetscape may have a negative influence several studies have found that walkability in the 1000 m buffer affects
on the psychology of pedestrians that would carry over to their sense of physical activity (Arvidsson et al., 2012; de Sa and Ardern, 2014).
the street. Further research is needed to clarify the optimality of quan­ Furthermore, a longitudinal study has specified that walkability in the
titative enclosures and to advance the semantic segmentation methods 1000 m network buffer has a greater effect on physical activity than
that can help measure the amount of enclosure. walkability in the 500 m network buffer (Kikuchi et al., 2018).
Besides small planters and street trees in urban design variables Conversely, this study’s results reveal that although a relationship does
(Ewing and Handy, 2009; Yin, 2017), our model indicated interaction appear between the neighborhood SW in 500 m and older females
effects between the road and the terrain segments or between the side­ walking for leisure, the neighborhood SW in 1000 m is not related to
walk and terrain segments cause increases in the SW. Although the plant walking behavior for leisure. In brief, the area in which micro-scale
factor increases the perception of human scale of the urban design walkability based on pedestrian perceptions affects walking behavior
protocol (Ewing and Handy, 2009), it is difficult to obtain the distri­ is smaller than objective walkability. This suggests a functional differ­
bution of this factor from objective data sources. The results of semantic ence between objective walkability and micro-scale walkability.
segmentation indicate that a small planter or street tree may exist at Neighborhoods that have abundant destinations, such as stations, shops,
locations where there is a combination of a terrain segment and a road and restaurants, include purposes for walking behavior, and its area is
segment or of a terrain segment and a sidewalk segment (Fig. 7). often stipulated using a range of about 1000 m, which matches the
Although further research is needed to improve accuracy, it is certain distance that most people are willing to walk (Arvidsson et al., 2012; Lee
that GSV and the deep learning approach can objectively extract plant and Moudon, 2006). On the other hand, neighborhoods including safe
components related to human scale. In case of the vegetation segment, streets, well-maintained pedestrian infrastructure, and attractive
contrast to terrain segment, the interaction effects with the road streetscapes encourage older females to go out. Conditions around one’s
segment have a negative relationship to SW. However, locations with residence are particularly important.
high proportions of road and vegetation segments include both main­ This study’s result was adjusted by objective walkability. Considered
tained street plants and dense vegetation in yards and parks (Fig. 7). together with the foregoing discussion, it is found that older female
Although greenness is commonly recognized as an effective positive residents could lose interest in leisure walking when nearby streets are
factors promoting physical activity and mental health (Bell et al., 2008; pedestrian-unfriendly, despite an abundance of destinations in the
Lu, 2018; Sugiyama et al., 2008), dense vegetation that creates a blind neighborhood. Maintaining a pedestrian-friendly environment and
space could potentially lead to an increase in fear of crime (Bogar and street on a micro-scale (e.g., for each residential unit) is a subject of
Beyer, 2016) and perceptions of unsafety (Jansson et al., 2013). Clari­ concern, in addition to enhancing accessibility to destinations on the
fying the best arrangement of greenness to increase physical activity by macro-scale. The results of our study enable us to envisage an automatic
objective methods, such as computer vision, would be needed. detection of streetscapes that should be maintained or repaired in
From the analytical results on the association between leisure detailed spatial resolution to the level of the intersection.
walking and the neighborhood SW, improving the neighborhood SW This study has several limitations. First, GSV images were not taken
based on the condition of pedestrian infrastructure, safety, and aes­ in the same years or seasons. Although more than 90% of the GSV im­
thetics facilitates older females’ active walking for leisure although ages we used were taken from 2016 to 2019, some images might not
there is no association with male leisure walking. The positive rela­ reflect the current streetscapes. How the changes of streetscapes by year
tionship between micro-scale walkability and leisure walking has been or season influence micro-scale walkability should be studied further.
clarified by previous studies (De Bourdeaudhuij et al., 2005; Inoue et al., Second, the deep learning model used is to segment the general street­
2010; Saelens and Handy, 2008; Witten et al., 2012). Moreover, several scape but not specifically based on pedestrian perceptions. Although the
studies also proved that street safety and condition encourage females’ regression model based on the semantic segmentation of the streetscape
physical activity (Richardson et al., 2017; Suminski et al., 2005). To the can predict the SW score, there are other components that encourage or
best of our knowledge, however, little work has been done to analyze discourage walking behavior, such as the presence of historic buildings,
this relationship based on walkability evaluated by an automatic color, graffiti, and noise. Further studies are needed to increase the ac­
method. Saelens and Handy (2008) found that measurement of the built curacy of the model based on walkability-specific components. Third,
environment related to aesthetics varies widely across studies. That is, there is bias due to study design. Although active older people might
the quantitative understanding of suitable streetscapes for leisure select more walkable neighborhoods for their residences, the cross-
walking is not yet complete. Therefore, automated method of the SW sectional study has difficulty explaining whether the walkable neigh­
evaluation that this study demonstrated can help enable the evaluation borhood causes older people to be more active or not. Additionally, since
of the locations that comprise leisure walking-friendly streetscapes with leisure walking time is self-reported, we need to consider recall bias.
a unified criterion. Future studies should reduce this bias through a longitudinal approach
In terms of neighborhood distance, for GIS-based objective and objective measurement of walking behavior. Fourth, the study area

9
S. Nagata et al. Health and Place 66 (2020) 102428

and the ages of the participants that we examined are limited. In relation Christiansen, L.B., Cerin, E., Badland, H., Kerr, J., Davey, R., Troelsen, J., van Dyck, D.,
Mitáš, J., Schofield, G., Sugiyama, T., Salvo, D., Sarmiento, O.L., Reis, R., Adams, M.,
to the analysis of the relationship between micro-scale walkability and
Frank, L., Sallis, J.F., 2016. International comparisons of the associations between
leisure walking behavior, this study found that SW based on automated objective measures of the built environment and transport-related walking and
evaluation is positively related to active leisure walking by older fe­ cycling: IPEN adult study. J.Transp.Health 3, 467–478.
males. It is possible that pedestrian perceptions of streetscapes vary Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U.,
Roth, S., Schiele, B., 2016. The cityscapes dataset for semantic urban scene
depending on other stipulations, such as age and location. Therefore, understanding. In: Proceedings of the IEEE Conference on Computer Vision and
extensive tests of multiple aspects, such as age, other sociodemographic Pattern Recognition, pp. 3213–3223.
characteristics, and other cities, are also needed. De Bourdeaudhuij, I., Teixeira, P.J., Cardon, G., Deforche, B., 2005. Environmental and
psychosocial correlates of physical activity in Portuguese and Belgian adults. Publ.
Health Nutr. 8, 886–895.
5. Conclusion de Sa, E., Ardern, C.I., 2014. Neighbourhood walkability, leisure-time and transport-
related physical activity in a mixed urban-rural area. PeerJ 2, e440.
Duncan, D.T., Goedel, W.C., Chunara, R., 2018. Quantitative methods for measuring
Micro-scale walkability in relation to pedestrian perceptions is an neighborhood characteristics in neighborhood health research. In: Duncan, D.T.,
important factor in walking behavior. This study found that the deep Kawachi, I. (Eds.), Neighborhoods and Health, second ed. Oxford University Press,
learning approach is a useful tool to quantify streetscapes. Furthermore, New York, pp. 57–90.
Ewing, R., 1996. Pedestrian-and Transit-Friendly Design. Report Prepared for the Public
we found a relationship between leisure walking by older females and Transit Office. Florida Department of Transportation.
micro-scale walkability based on the quantified streetscape. The pro­ Ewing, R., Handy, S., Brownson, R.C., Clemente, O., Winston, E., 2006. Identifying and
posed automated method allows assessment of the micro-scale aspects of measuring urban design qualities related to walkability. J. Phys. Activ. Health 3,
s223–s240.
the neighborhood environment with a unified objective criterion and the Ewing, R., Hajrasouliha, A., Neckerman, K.M., Purciel-Hill, M., Greene, W., 2016.
detection of leisure walking-(un)friendly streetscapes over large areas. Streetscape features related to pedestrian activity. J. Plann. Educ. Res. 36, 5–15.
This may be of widespread benefit to urban planning and studies on the Ewing, R., Handy, S., 2009. Measuring the unmeasurable: urban design qualities related
to walkability. J. Urban Des. 14, 65–84.
urban environment and human health.
Guthold, R., Stevens, G.A., Riley, L.M., Bull, F.C., 2018. Worldwide trends in insufficient
physical activity from 2001 to 2016: a pooled analysis of 358 population-based
surveys with 1⋅9 million participants. Lancet Global Health 6, e1077–e1086.
Declaration of competing interest Handy, S.L., Boarnet, M.G., Ewing, R., Killingsworth, R.E., 2002. How the built
environment affects physical activity: views from urban planning. Am. J. Prev. Med.
None. 23, 64–73.
Hanibuchi, T., Nakaya, T., Inoue, S., 2019. Virtual audits of streetscapes by
crowdworkers. Health Place 59, 102203.
Acknowledgments Holland, J.H., 1975. Adaptation in Natural and Artificial Systems. MIT press, Ann Arbor.
Hu, F.B., Sigal, R.J., Rich-Edwards, J.W., Colditz, G.A., Solomon, C.G., Willett, W.C.,
Speizer, F.E., Manson, J.E., 1999. Walking compared with vigorous physical activity
This work was supported by JSPS KAKENHI, Japan (Grant Numbers and risk of type 2 diabetes in women: a prospective study. J. Am. Med. Assoc. 282,
17H00947, JP18KK0371, 20500604, and 20H00040). 1433–1439.
Hu, F.B., Stampfer, M.J., Solomon, C., Liu, S., Colditz, G.A., Speizer, F.E., Willet, W.C.,
Manson, J.E., 2001. Physical activity and risk for cardiovascular events in diabetic
Appendix A. Supplementary data women. Ann. Intern. Med. 134, 96–105.
Inoue, S., Ohya, Y., Odagiri, Y., Takamiya, T., Ishii, K., Kitabayashi, M., Suijo, K.,
Supplementary data to this article can be found online at https://doi. Sallis, J.F., Shimomitsu, T., 2010. Association between perceived neighborhood
environment and walking among adults in 4 cities in Japan. J. Epidemiol. 20,
org/10.1016/j.healthplace.2020.102428.
277–286.
Inoue, S., Ohya, Y., Odagiri, Y., Takamiya, T., Kamada, M., Okada, S., Oka, K.,
References Kitabatake, Y., Nakaya, T., Sallis, J.F., Shimomitsu, T., 2011. Perceived
neighborhood environment and walking for specific purposes among elderly
Japanese. J. Epidemiol. 21, 481–490.
Amagasa, S., Fukushima, N., Kikuchi, H., Takamiya, T., Oka, K., Inoue, S., 2017. Light
Jansson, M., Fors, H., Lindgren, T., Wiström, B., 2013. Perceived personal safety in
and sporadic physical activity overlooked by current guidelines makes older women
relation to urban woodland vegetation - a review. Urban For. Urban Green. 12,
more active than older men. Int. J. Behav. Nutr. Phys. Activ. 14, 59.
127–133.
Amagasa, S., Inoue, S., Fukushima, N., Kikuchi, H., Nakaya, T., Hanibuchi, T., Sallis, J.F.,
Kikuchi, H., Nakaya, T., Hanibuchi, T., Fukushima, N., Amagasa, S., Oka, K., Sallis, J.F.,
Owen, N., 2019. Associations of neighborhood walkability with intensity- and bout-
Inoue, S., 2018. Objectively measured neighborhood walkability and change in
specific physical activity and sedentary behavior of older adults in Japan. Geriatr.
physical activity in older Japanese adults: a five-year cohort study. Int. J. Environ.
Gerontol. Int. 19, 861–867.
Res. Publ. Health 15, 1814.
Anselin, L., 2003. Spatial externalities, spatial multipliers, and spatial econometrics. Int.
Kim, S., Park, S., Seung, J., 2014. Meso- or micro-scale ? Environmental factors
Reg. Sci. Rev. 26, 153–166.
influencing pedestrian satisfaction. Transport. Res. Part D 30, 10–20.
Arvidsson, D., Kawakami, N., Ohlsson, H., Sundquist, K., 2012. Physical activity and
Lee, C., Moudon, A.V., 2006. The 3Ds + R: quantifying land use and urban form
concordance between objective and perceived walkability. Med. Sci. Sports Exerc.
correlates of walking. Transport. Res. Part D 11, 204–215.
44, 280–287.
Lee, I.M., Shiroma, E.J., Lobelo, F., Puska, P., Blair, S.N., Katzmarzyk, P.T., Alkandari, J.
Badrinarayanan, V., Kendall, A., Cipolla, R., Member, S., 2017. SegNet : a deep
R., Andersen, L.B., Bauman, A.E., Brownson, R.C., Bull, F.C., Craig, C.L., Ekelund, U.,
convolutional encoder-decoder architecture for image segmentation. IEEE Trans.
Goenka, S., Guthold, R., Hallal, P.C., Haskell, W.L., Heath, G.W., Inoue, S.,
Pattern Anal. Mach. Intell. 39, 2481–2495.
Kahlmeier, S., Kohl, H.W., Lambert, E.V., Leetongin, G., Loos, R.J.F., Marcus, B.,
Bell, J.F., Wilson, J.S., Liu, G.C., 2008. Neighborhood greenness and 2-year changes in
Martin, B.W., Owen, N., Parra, D.C., Pratt, M., Ogilvie, D., Reis, R.S., Sallis, J.F.,
body mass index of children and youth. Am. J. Prev. Med. 35, 547–553.
Sarmiento, O.L., Wells, J.C., 2012. Effect of physical inactivity on major non-
Bland, J.M., Altman, D.G., 1995. Multiple significance tests: the Bonferroni method. BMJ
communicable diseases worldwide: an analysis of burden of disease and life
310, 170.
expectancy. Lancet 380, 219–229.
Bogar, S., Beyer, K.M., 2016. Green space, violence, and crime: a systematic review.
Lu, Y., 2018. The association of urban greenness and walking behavior: using Google
Trauma Violence Abuse 17, 160–171.
Street View and deep learning techniques to estimate residents’ exposure to urban
Cai, B.Y., Li, X., Seiferling, I., Ratti, C., 2018. Treepedia 2.0: applying deep learning for
greenness. Int. J. Environ. Res. Publ. Health 15, 1576.
large-scale quantification of urban tree cover. In: 2018 IEEE International Congress
Matula, D.W., Sokal, R.R., 1980. Properties of Gabriel graphs relevant to geographic
on Big Data (BigData Congress), pp. 49–56.
variation research and the clustering of points in the plane. Geogr. Anal. 12,
Cain, K.L., Millstein, R.A., Sallis, J.F., Conway, T.L., Gavand, K.A., Frank, L.D., Saelens, B.
205–222.
E., Geremia, C.M., Chapman, J., Adams, M.A., Glanz, K., King, A.C., 2014.
Nelson, M.E., Rejeski, W.J., Blair, S.N., Duncan, P.W., Judge, J.O., 2007. Physical activity
Contribution of streetscape audits to explanation of physical activity in four age
and public health in older adults : recommendation from the American college of
groups based on the Microscale Audit of Pedestrian Streetscapes (MAPS). Soc. Sci.
sports medicine and the American heart association. Circulation 116, 1094–1105.
Med. 116, 82–92.
Owen, N., Cerin, E., Leslie, E., duToit, L., Coffee, N., Frank, L.D., Bauman, A.E., Hugo, G.,
Caspersen, C.J., Pereira, M.A., Curran, K.M., 2000. Changes in physical activity patterns
Saelens, B.E., Sallis, J.F., 2007. Neighborhood walkability and the walking behavior
in the United States, by sex and cross-sectional age. Med. Sci. Sports Exerc. 32,
of Australian adults. Am. J. Prev. Med. 33, 387–395.
1601–1609.
Pliakas, T., Hawkesworth, S., Silverwood, R.J., Nanchahal, K., Grundy, C., Armstrong, B.,
Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H., 2018. Encoder-Decoder with
Casas, J.P., Morris, R.W., Wilkinson, P., Lock, K., 2017. Optimising measurement of
atrous separable convolution for semantic image segmentation. In: Ferrari, V.,
health-related characteristics of the built environment: comparing data collected by
Hebert, M., Sminchisescu, C., Weiss, Y. (Eds.), Computer Vision – ECCV 2018.
Springer, New York, pp. 833–851.

10
S. Nagata et al. Health and Place 66 (2020) 102428

foot-based street audits, virtual street audits and routine secondary data sources. Villeneuve, P.J., Id, R.L.Y., Root, A., Ambrose, S., Dimuzio, J., Kumar, N., Shehata, M.,
Health Place 43, 75–84. Xi, M., Seed, E., Li, X., Shooshtari, M., Rainham, D., 2018. Comparing the normalized
Purciel, M., Neckerman, K.M., Lovasi, G.S., Quinn, J.W., Weiss, C., Bader, M.D.M., difference vegetation index with the google street view measure of vegetation to
Ewing, R., Rundle, A., 2009. Creating and validating GIS measures of urban design assess associations between greenness , walkability , recreational physical activity
for health research. J. Environ. Psychol. 29, 457–466. and health in Ottawa , Canada. Int. J. Environ. Res. Publ. Health 15, 1719.
Richardson, A.S., Troxel, W.M., Ghosh-Dastidar, M.B., Beckman, R., Hunter, G.P., Wang, R., Liu, Y., Lu, Y., Yuan, Y., Zhang, J., Liu, P., Yao, Y., 2019. The linkage between
DeSantis, A.S., Colabianchi, N., Dubowitz, T., 2017. One size doesn’t fit all: cross- the perception of neighbourhood and physical activity in Guangzhou, China : using
sectional associations between neighborhood walkability, crime and physical street view imagery with deep learning techniques. Int. J. Health Geogr. 18, 18.
activity depends on age and sex of residents. BMC Publ. Health 17, 97. Williams, P.T., Thompson, P.D., 2013. Walking versus running for hypertension,
Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: convolutional networks for cholesterol, and diabetes mellitus risk reduction. Arterioscler. Thromb. Vasc. Biol.
biomedical image segmentation. In: International Conference on Medical Image 33, 1085–1091.
Computing and Computer-Assisted Intervention. Springer, Cham, pp. 234–241. Witten, K., Blakely, T., Bagheri, N., Badland, H., Ivory, V., Pearce, J., Mavoa, S.,
Rzotkiewicz, A., Pearson, A.L., Dougherty, B.V., Shortridge, A., Wilson, N., 2018. Hinckson, E., Schofield, G., 2012. Neighborhood built environment and transport
Systematic review of the use of Google Street View in health research : major themes, and leisure physical activity: findings using objective exposure and outcome
strengths, weaknesses and possibilities for future research. Health Place 52, measures in New Zealand. Environ. Health Perspect. 120, 971–977.
240–246. World Health Organization, 2017. Physical inactivity: a global public health problem.
Saelens, B.E., Handy, S.L., 2008. Built environment correlates of walking: a review. Med. Last accessed 3rd July, 2020. http://www.who.int/dietphysicalactivity/factsheet
Sci. Sports Exerc. 40, 550–566. _inactivity/en/.
Sugiyama, T., Cerin, E., Owen, N., Oyeyemi, A.L., Conway, T.L., Van Dyck, D., Yen, I.H., Michael, Y.L., Perdue, L., 2010. Neighborhood environment in studies of health
Schipperijn, J., Macfarlane, D.J., Salvo, D., Reis, R.S., Mitáš, J., Sarmiento, O.L., of older adults : a systematic review. Am. J. Prev. Med. 37, 455–463.
Davey, R., Schofield, G., Orzanco-Garralda, R., Sallis, J.F., 2014. Perceived Yin, L., 2017. Street level urban design qualities for walkability: combining 2D and 3D
neighbourhood environmental attributes associated with adults’ recreational GIS measures. Comput. Environ. Urban Syst. 64, 288–296.
walking: IPEN Adult study in 12 countries. Health Place 28, 22–30. Yin, L., Cheng, Q., Wang, Z., Shao, Z., 2015. “Big data” for pedestrian volume: exploring
Sugiyama, T., Leslie, E., Giles-Corti, B., Owen, N., 2008. Associations of neighbourhood the use of Google Street View images for pedestrian counts. Appl. Geogr. 63,
greenness with physical and mental health: do walking, social coherence and local 337–345.
social interaction explain the relationships? J. Epidemiol. Community Health 62 e9 Yin, L., Wang, Z., 2016. Measuring visual enclosure for street walkability : using machine
LP-e9. learning algorithms and Google Street View imagery. Appl. Geogr. 76, 147–153.
Suminski, R.R., Poston, W.S.C., Petosa, R.L., Stevens, E., Katzenmoyer, L.M., 2005.
Features of the neighborhood environment and walking by U.S. adults. Am. J. Prev.
Med. 28, 149–155.

11