You are on page 1of 18

Urban Analytics and

Article City Science

EPB: Urban Analytics and City Science


0(0) 1–20
Walkability scoring: ! The Author(s) 2020
Article reuse guidelines:
Why and how does a sagepub.com/journals-permissions
DOI: 10.1177/2399808320977871
3-dimensional journals.sagepub.com/home/epb

pedestrian network
matter?

Jianting Zhao and Guibo Sun


Department of Urban Planning and Design, The University of Hong Kong, Hong
Kong

Chris Webster
Faculty of Architecture, The University of Hong Kong, United Kingdom of Great
Britain and Northern Ireland

Abstract
Previous walkability scoring systems are all based on road networks, even though roads are not
designed for pedestrians. To calculate an accurate walking score, we need pedestrian network
data. This is especially the case in cities such as Hong Kong, where pedestrians are separated
from vehicles by footbridges, underpasses or surface sidewalks. In this paper, we investigate why
and how a three-dimensional pedestrian network makes a difference in walkability scoring, using
Hong Kong as a case city. We developed a walkability scoring system based on networks and
amenities, using multiple open-source programming platforms and languages. Separately, we cal-
culated walkability scores (on a scale of 0–100) using the three-dimensional pedestrian network
and road network of the city, comparing the differences between the two. A GIS raster analysis
was conducted to extract walkability scoring differences from the two walkability surfaces,
followed by a univariate linear model to examine how the scores were underestimated if without
using the three-dimensional pedestrian network. Results show that streets were considered
twice as walkable if rated by pedestrian network rather than road network. Walkability scores
were 92% higher on average. The fitted model shows that the mean score underestimations were
significantly different for different three-dimensional network elements. Surface sidewalks had an
average underestimation of 33.75 (p < 0.001), footbridges and underground paths expanded the
underestimations by 3.85 and 2.97 (both p < 0.001), respectively, and the linkages to footbridge
and underground path enlarged the surface sidewalk underestimations by 2.68 and 4.92 (both

Corresponding author:
Guibo Sun, Department of Urban Planning and Design, The University of Hong Kong, Pokfulam Road, HKU, Hong Kong.
Email: gbsun@hku.hk
2 EPB: Urban Analytics and City Science 0(0)

p < 0.001). We suggest that walkability evaluation systems should be developed on pedestrian
networks instead of road networks, especially for high-density cities.

Keywords
Walkability, Walk Score, pedestrian network, Hong Kong, high-density city

Introduction
Walking is the most basic physical activity and assumes great importance in current urban
planning and design. A standardised measurement makes walkability easily studied analyt-
ically and across disciplines. Attempts to create walkability measures started in the 1990s,
with increasingly sophisticated composite walkability indices being proposed in the past
decade (Hall and Ram, 2018; Shashank and Schuurman, 2019). One gained wide attention
is Walk Score, which is developed by a group of planners and urbanists and provides
standardised walkability evaluations (Carr et al., 2011). Leveraging on web and amenities
data, Walk Score evaluates a great number of cities in the USA, Canada, Australia and
New Zealand. The ready-to-use scoring data and algorithms permit extensive evaluation by
practitioners and researchers without their own modelling capacity. A review of 42 Walk
Score-related studies shows that the score has been used as an explanatory variable to
predict various performance outcomes, including physical activity, housing value and
social cohesion (Hall and Ram, 2018).
However, Walk Score data are not available to Asian cities. Superimposing its default
algorithm to unsupported regions would yield inaccurate and unreliable measures, as Asian
cities tend to be denser, and pedestrian networks differ from road networks. Pedestrian
routes in high-density cities, especially hilly ones such as Hong Kong, have many non-
vehicular short cuts between two-dimensional (2D) road network links, some of which
deviate from the road network in three-dimension (3D). Pedestrians may use footbridges
to cross large transport infrastructure or use underground paths connected with metro
stations to access different destinations (Zacharias and He, 2018) as well as using public
and semi-public buildings such as shopping malls as thoroughfares (Shelton et al., 2011).
Such pedestrian pathways are not captured in road network data, but they are crucial in
walkability assessments. Moreover, using road networks as a surrogate for pedestrian net-
works overlooks the incompatibility of road segments to pedestrians use; for example, the
expressways and cross-harbour/mountain tunnels are car-only. Using a default Walk Score
algorithm based on a road network to study walkability would undermine the accuracy of
walkability evaluation and any policy, engineering or design interventions. It is particularly
the case in high-density cities.
In this study, we investigated why and how a pedestrian network-based walkability score
(WS) performs better than the road network-based WS. We constructed a walkability scor-
ing system using amenities data and a 3D digital pedestrian network for Hong Kong (Sun
et al., 2019). The same algorithm was applied to the regular road network in order to analyse
how the scoring was underestimated. The pedestrian network-based WS and road network-
based WS were assigned to their respective network junctions and were calculated using a
workflow that integrates multiple open-source programming languages. Since scores are
given to junctions, their dense distribution allows us to capture nuanced differences that
a 3D pedestrian network would make in an urban fabric. This study aims to fill the gap in
Zhao et al. 3

the existing literature by measuring the advantages of using pedestrian network as the basis
of walkability evaluation, and conversely, measuring the error incurred by using a road
network as a surrogate.

Previous work
The well-recognised urban design framework of 3Ds – density, diversity and design (Cervero
and Kockelman, 1997) or 5Ds – adding distance to transit and destination accessibility
(Ewing and Cervero, 2010), provides a widely adopted method to standardise the features
of a built environment conducive to walking behaviour. The fifth D, proximity to and
availability of amenities are the inherent reason and incentive to generate walking activity,
whereas other features ensure people’s confidence and preference in walking. The abun-
dance of amenities explains why streets with poor infrastructure in the Global South have
more walking activity than those well-established streets in upper-class towns in the US
(Dovey and Pafka, 2020). Moreover, Walk Score, a destination-based evaluation system,
yields a higher WS with a closer distance and/or a greater variety of amenities (Walk Score,
2019). Walk Score’s approach follows a distance decay function and codes amenities as
categorical variables (grocery stores, restaurants, hardware stores, etc.) with equal weight
for each item in the same category (Carr et al., 2011). A validation conducted in Japan
shows that Walk Score may also be applied to East Asian settings without detailed geo-
graphic data (Koohsari et al., 2018). This result lends credibility in using its framework
outside of its targeted region; however, more research is still needed to further support this
argument due to the diversity and density in Asian context.
Nevertheless, with more local knowledge, it is wise to adjust the framework to comply
with local people’s lifestyle, especially the weight and distance decay functions. For our case
city, Hong Kong, Walk Score does not suit her in two aspects. The first aspect is people’s
low tolerance to walking distance. Local urban designers revealed that Hong Kong people
generally are only willing to walk up to 400 m (Audi et al., 2010). Reasons are twofold.
Varieties of transport modes weaken the incentive to walk beyond 400 m. Besides walking,
tramway, bus, minibus and taxi are convenient alternatives. The confusing wayfinding, hilly
topography and humid weather are also a nuisance. Another reason is the abundance and
low differentiation in amenities that can satisfy people’s daily need without needing to travel
beyond 400 m. Taking grocery store as an example, two largest grocery store brands togeth-
er take up almost 70% of the market share, providing nearly 600 branches (Li, 2019).
Widely distributed convenient stores and pharmacies also carry similar products.
Restaurants are too omniscient. Within a street block, it is not unusual to have 10 dining
options that provide different styles of Chinese cuisines, Western and Southeast Asian
cuisines and fast food options. That is said, there would need adjustments on the distance
and weight decay functions when we revamp the walkability scoring system for Hong Kong.
The second aspect is the low accuracy of road network in describing pedestrian move-
ments. In high-density cities, the pedestrian network cannot be simplified as an offset from
the existing road network. Its network elements are richer and connections more
complicated. In Hong Kong, a newly created pedestrian network is 2.4 times longer in
length and 8.5 times larger in the number of links compared with the road network (Sun
et al., 2019). The difference is present in plane view and 3D, with a pedestrian network
model classifying network elements into multiple levels of connections (e.g. sidewalk,
footbridge, underground paths, crosswalk, ramp and rooftop paths). A recent study finds
out that the walking accessibility to neighbourhood open space from private housing differs
by 2–8% between using the 3D pedestrian network and conventional network (Tang et al.,
4 EPB: Urban Analytics and City Science 0(0)

2020). Several studies have started to research the importance of using pedestrian network in
walkability evaluation. Chin et al. (2008) studied the connectivity difference between road
and pedestrian networks. They conclude that when using pedestrian networks, there is an
increase of up to 120% in Pedshed, which is a connectivity measure given to the ratio of
straight line and street network distance. Tal and Handy (2012) studied the effect of
pedestrian network on walking accessibility in suburban neighbourhoods in Davis,
California. They find that pedestrian network length was 40% underestimated when
using road network data, and pedestrian network data greatly improve the assessment
accuracy of walking accessibility. A study of the footpath network in Belfast, UK shows
that pedestrian network metrics predicted physical activity behaviours more accurately than
road network metrics (Ellis et al., 2016). The differences are significant when using
pedestrian network rather than road network.
In summary, previous literature suggests that a walkability scoring system in high-density
cities can differ from those in low-density cities due to the 3D pedestrian network and
amenities usages. The WS evaluation needs to incorporate pedestrian network, local ame-
nities usage, in addition to the Walk Score framework to best capture the walkability in a
high-density Asian city.

Walkability scoring system


Pedestrian network and amenities data
Our Hong Kong 3D pedestrian network data (segments ¼ 269,242, junctions ¼ 225,091)
include 23 types of pedestrian paths, encompassing five height levels: surface level (Level
0, e.g. sidewalk), underground path (Level –2, e.g. metro stations), footbridge (Level 2),
underground linkage level that links surface and underground paths (Level –1) and
above ground linkage level that links surface and footbridges (Level 1). The linkage infra-
structure for all levels includes stairs, ramps, escalators and elevators (Sun et al., 2019). The
3D-pedestrian network data can thoroughly record various movement options available to
pedestrians in the high-density city of Hong Kong.
We obtained a road network model (segments ¼ 19,255, junctions ¼ 14,417) from the
Transport Department of Hong Kong to compare the two networks. In terms of segment
quantity, they are noticeably different. Figure 1 presents the differences. Blue lines are roads
for cars; red lines are paths for pedestrians, with different line widths for different height
levels: the thin lines are sidewalks and thick lines are elevated or underground paths. The
pedestrian network is much more densely distributed on a finer scale than the road network
in many places, and the road network tends to serve as the periphery for local walking
systems with many roads having no pedestrian network. The figure graphically emphasises
the intuitive assumption that walkability analyses using a road network can theoretically be
inaccurate in a 3D city.

We extracted geocoded amenities data from points of interest (POI) data sources. POIs
(N ¼ 329,644) were retrieved in 2017 from a location-based navigation company and vali-
dated by the government’s amenities data (Sun et al., 2018).

Data processing using integrated open-source programs


Our algorithm utilises PostgreSQL, an open-source database platform, and Python, an
open-source programming language. By using a Python–PostgreSQL connector, Python
Zhao et al. 5

Figure 1. Comparison between the pedestrian and road networks.

Figure 2. The workflow of the WS calculation.


6 EPB: Urban Analytics and City Science 0(0)

code can directly communicate with PostgreSQL code and thus create a smooth workflow.
We also used PostGIS and pgRouting, two spatial extensions that extend the capabilities of
PostgreSQL to conduct spatial query and network routing. The workflow is illustrated in
Figure 2.
First, we eliminated several types of amenities due to their limited use in daily life such as
car repair companies and funeral venues. The cleaned data has 100,786 amenities, for which
we classified into 10 categories in a PostgreSQL database. The classification covers different
aspects of people’s daily needs, including shopping, dining, transportation, education, activ-
ities, healthcare and more. When categorising amenities, we incorporated amenities that
may not be regularly consumed in western cities but are daily use for Asian culture, includ-
ing tea house, fresh vegetable booth and dessert place (Sun et al., 2018). Details are shown in
Table 1.
Second, we converted the 3D pedestrian network into topology tables using the Create
Topology function in pgRouting extension. This function connects pedestrian network seg-
ments with junctions into a structured graph and navigable network (pgRouting
Community, 2016). The topologically structured network sets rules for connecting different
levels (e.g. footbridges and sidewalks), such that levels can only be connected if they are
linked by linkage elements such as stairs or escalators. We use the topology graph to identify
the start and endpoints of line segments in the pedestrian network and structure them into
‘junction’ tables. The junction tables store all the connecting points of the network, within
and between levels. Junctions also serve as our units of analysis for walkability scoring. We
attached amenities to line segment using the Locate Point to Line function in PostGIS
extension, which calculates each amenity’s relative position to its closest line segment
based on distance fraction (Figure 2).
Third, we applied a network-distance routing function to extract all amenities within a
400-m network distance from each junction. The benchmark of 400 m is equivalent to
5 minutes of walk and hence is often used in active travel and physical research in Hong
Kong and elsewhere (Cerin et al., 2007; Duncan et al., 2013; Ker and Ginn, 2003; Tang
et al., 2020; Villanueva et al., 2014). The routing function, named withPointsDD, belongs to
the pgRouting extension and is derived from the Dijkstra algorithm, a classic routing algo-
rithm that is used to find the shortest path. The Dijkstra algorithm enumerates the junctions
that are connected to an origin and assigns a value based on distance or the difficulty to
travel. Junctions further away are assigned values following the similar logic, but the value
only updates if the new value is smaller than the previously given one. Only the lowest value
junctions are kept, which leads to the shortest route (pgRouting Community, 2016).
The routing function in pgRouting allows users to simultaneously consider nearby ame-
nities for all line segments within defined geographical boundaries. It considerably expedited
the calculation process from several days without the efficient routine to within one hour for
our dataset of 269,242 segments, 225,091 junctions and 100,786 amenities.

WS algorithm
We used a Python algorithm to calculate WS for all junctions and used a Python–
PostgreSQL connector to send results back to the PostgreSQL database. Our walkability
scoring system takes into account distance, quantity of amenities and diversity within 400 m
from each network junction. Distance is represented in a distance decay function (D), quan-
tity of amenities is weighted following a weight decay function (W). The functions vary for
different amenities as shown in Table 1. Diversity is measured by a mixture index (M)using
an entropy formula derived from the Shannon Index. This entropy formula (Formula (1))
Zhao et al.

Table 1. Walkability scoring system.

Distance step Weight step decay


decay function f(D) function f(W)

Amenities categories Distance (m) Score Number Weight

Shop (includes comprehensive 0–100 10 1 0.5


market, convenience store, 101–200 8 2 0.3
shopping-related site and 201–300 5 3 0.1
supermarket) 301–400 3 4 0.1
Metro station 0–200 10 1 0.9
Bus stop 201–400 7 2 0.1
Restaurant (includes fast-food 0–200 10 1 0.5
restaurant, cafe, foreign res- 201–400 5 2 0.3
taurant, food-related site, 3 0.1
casual dining option, bakery, 4 0.1
tea house and cold store)
Entertainment (includes bar,
club, pub, internet bar and
theatre)
Park
Sports venue
School
Personal care (includes beauty
salon, nail salon and spa)
Healthcare (includes general
hospital and clinic, specialist
hospital)
7
8 EPB: Urban Analytics and City Science 0(0)

has been adopted in previous studies (Mavoa et al., 2018; Sun et al., 2018). These functions
jointly determine the score for each junction (Formula 2). The overall score is the sum of 10
amenity categories and a mixture index. These 11 items are fitted into a 0–100 number range
by multiplying a fraction term 1011
Mixture Index
!
X
n
M¼ 1 pi  lnðpi Þ lnðnÞ (1)
i¼1

where pi ¼ (number of accessible* amenities in each category)/(total number of accessible


amenities), n is the number of amenities categories, i is one amenities category (i ¼ 1. . .10),
and 1 is the maximum mix and 0 is the least mix; *‘accessible’ means being within 400 m
network distance from a junction.

Walkability Score
!
n X
X m
10
WS ¼ fðDÞ  fðWÞ þ M  10  (2)
i j
11

where D is the distance step decay function, W is weight step decay function, M is the
mixture index, n is the number of amenities categories, i is one amenities category
(i ¼ 1. . .10), m is the number of valid amenities in each of the n categories and j is the
ranking of the valid amenities.
People tend to make different uses of amenities and with various thresholds of walking
distance. For these and supply-side reasons, density and layout of amenities are also very
different in high-density cities (Shelton et al., 2011). High concentrations of similar desti-
nations, close substitutes or even identical products such as multiple branches of a chain are
common in a neighbourhood. Taking the example of quantity and density of accessible
destinations, the benefit of destination count from the same category plateaus after a certain
threshold, following the law of diminishing marginal utility, and hence using count is better
for reflecting a location’s vitality than destination density. For daily life uses, residents may
perceive no difference between having 10–20 convenient stores and another 30–40 in a
100-m walking-distance neighbourhood. Similar rules may be applicable for some other
amenities categories; for instance, on average, within 400-m walking distance, there are
162 restaurants in Hong Kong (Sun et al., 2018). It is unlikely that residents use (or
weight) each restaurant in their neighbourhood equally.
Therefore, in our algorithm, the decay functions were set separately for shops, metro
stations, bus stops and other categories (Table 1). For shops, we assumed that people would
prefer the closest shops (D0–100m ¼ 10, D101–200m ¼ 8, D201–300m ¼ 5, D301–400m ¼ 3), and their
daily needs would be satisfied by the nearest four shops (W1 ¼ 0.5, W2 ¼ 0.3, W3,4 ¼ 0.1);
hence the fifth one onward would not be considered. Public transit system in Hong Kong is
highly accessible; 90% of people use public transits on a daily basis (Wang, 2014). Transit
stops are separated into metro stations and bus stops, with the latter being more abundant.
They are not substitutes for one another in Hong Kong because most bus lines are
intentionally set to feed metro ridership rather than running in parallel (Wang, 2014).
For metro stations and bus stops, we assumed those within 200 m to be preferable
(D0–200m ¼ 10, D201–400m ¼ 7) and the closest one to be more meaningful than the second
Zhao et al. 9

closest (W1 ¼ 0.9, W2 ¼ 0.1). For all other categories, the distance decay was set to a default
function, with places within 200 m being more preferred, between 200m and 400 m being less
preferred, and beyond 400 m being not desirable (D0–200m ¼ 10, D201–400m ¼ 5). The weight
decay for these categories considers the nearest four candidates, similar to the shop
category.
The setting of 100 m and 200 m step distance decay thresholds is in reference to the
average block size in Hong Kong, which is 111 m (SD ¼ 36), and the average interval
between surface pedestrian crossings, around 200 m (Wang, 2014). Therefore, the most
convenient areas for walking are the home block and adjacent blocks. A route that does
not need to cross a vehicular road is likely to be the least walking effort route choice.
Since the main objective of this study is to investigate how much the WSs would differ
when using the pedestrian versus road networks, we applied the same algorithm to the road
network data, following the same workflow (Figure 2).
Furthermore, a sensitivity analysis was conducted to explore the impact brought by the
weight settings in order to rule out the possible inaccuracy caused by weight settings. In this
analysis, we recalculated the pedestrian-network WS with equal weights. For instance, the
weights for shops, restaurants and other categories were set to be 0.25, instead of 0.5, 0.3,
0.1, 0.1; the weights for metro station and bus stops were set to 0.5. This setting equalises the
weights while maintaining a maximum score of 10 for each category. The average score was
compared with that from the original WS.

Scoring difference: How underestimated is walkability scoring when


using a road network?
We used building footprints from the Hong Kong Lands Department’s topographic map
(iB1000, geographic scale: 1:1000) to delineate the built-up area for the walkability scoring
comparisons. This approach effectively narrows down our focus to 25% of the land that is
meaningful to study, as around 75% of Hong Kong is composed of mountains and other
types of non-built-up areas.
The network junction-based WS indices were interpolated into raster layers for score
difference comparisons. An Inverse Distance Weighting Interpolation in GIS (QGIS, v3.10)
was used to create two WS raster surfaces from the network junction-based WS results. This
tool interpolates the WS for unmeasured areas based on scores from nearby junctions. Since
the average length of the network segment in the pedestrian network in Hong Kong is 31 m
(SD ¼ 56) (Sun et al., 2019), we set the raster cell size to 30 m by 30 m, optimising compu-
tation workload and analysis grain size. The score differences are calculated by subtracting
road network-based WS raster from pedestrian network-based WS raster data. The sub-
tracted raster surface can be understood as a measure of error when using the road network
to measure walkability.
We first analysed scoring differences at town levels by aggregating scores using district
administrative boundaries. We calculated the differences in both total scores and each ame-
nities category and compared the gap between old and new towns. The old towns in Hong
Kong are historic districts that are naturally developed over time through the evolution of
land-use under market competition and government planning regulations and restrictions.
The new towns were developed mainly due to the housing needs of the increased population,
started in the early 1970s.
Second, we compared the score underestimations on a finer street scale. We selected the
top 20% (first quantile) WS underestimated by the road network, for which the 20%
10 EPB: Urban Analytics and City Science 0(0)

Figure 3. Walkability scoring difference using pedestrian network versus road network.

benchmark was chosen with reference to a previous study on this 3D pedestrian network
(Sun et al., 2019). We spatially overlapped them with 3D pedestrian network segments to
analyse how the different degrees of underestimations may depend on the different levels of
the 3D pedestrian network. A univariate linear model was applied, with network levels being
the independent variables and the top 20% underestimated scores being the dependent
variable. It was used to study how 3D levels (Level –2, –1, 1 and 2) helped explain the
walkability scoring underestimations. The statistical analyses were conducted in R language.

Results
The pedestrian network-based WS (N ¼ 225,091, Score ¼ 44.1) was 92% higher than the
road network-based WS (N ¼ 14,417, Score ¼ 23.0) in terms of the average score across
the city. The walkability difference is visualised in Figure 3. The areas in red are those
where walkability was underestimated when using the road network. These are mainly
located in town centres. Places with blue colour show an overestimated walkability when
using the road network for the analysis. These are mainly in the suburbs and tend to have
sparse pedestrian networks.
In the subcategories, we found the greatest underestimations happened in parks (279%),
school (183%) and sports facilities (179%) when using pedestrian network rather than road
network. On the other hand, the underestimations of the walkability for shops (38%) and
Zhao et al.

Table 2. WS comparison between old and new towns.


No. of Overall Sports Personal Mixture
Network Town junctions score Shop Metro Bus Restaurant Entertainment Park facility School care Healthcare index

Pedestrian New 132,380 36.04 5.65 1.22 7.4 5.33 1.03 1.59 2.43 3.41 2.34 3.65 0.56
Old 92,711 55.47 (54%a) 7.71 (36%) 3.27 (167%) 9.01 (22%) 7.73 (45%) 3.02 (194%) 2.98 (88%) 4.59 (89%) 5.51 (61%) 4.56 (95%) 6.25 (71%) 0.64 (14%)
Road New 5636 18.03 3.98 0.55 3.77 3.33 0.33 0.43 0.8 1.04 0.88 1.54 0.32
Old 8781 26.13 (45%) 5.2 (31%) 1.21 (120%) 4.44 (18%) 4.72 (42%) 1.12 (236%) 0.66 (54%) 1.44 (80%) 1.82 (75%) 1.8 (103%) 2.49 (62%) 0.38 (20%)
a
Percentage equals the percent of old town WS average that is greater than the new town’s.
11
12 EPB: Urban Analytics and City Science 0(0)

Figure 4. The underestimated areas: (a) and (b) two highly rated old towns and (c) the lowest-rated
new town.

restaurants (51%) were less. This may be due to the amenities’ layout that made them easily
accessed by either road or pedestrian network.
The sensitivity analysis result shows that the WS was not impacted by the unequally
assigned weights. The average WS for equal weights was 45.3, which is 2.7% higher than the
original WS. This result indicates that changes in weight settings have little impact on score
evaluation.

Walkability scoring underestimation at town scale


We observed a consistent greater underestimation in the scores for old towns than in those
for new towns across all categories. The average overall score of the pedestrian network-
based WS for old towns was 54% higher than for the new town (Table 2).
Two highly rated old towns and the lowest-rated new town in terms of average overall
scores were compared to visualise the determining factors in the networks. As Figure 4
shows, the highest rating town in terms of score underestimation has a much denser pedes-
trian network distribution than road network (difference ¼ 35.9, underestimation ¼ 184%).
The second highest-rated old town has a grid layout and dense distribution of amenities
(35.3, 99%). The town with the least scoring difference is a new town (9.3, 50%). It had
sparser pedestrian networks compared to old towns, and amenities were clustered in a small
area but very scattered elsewhere. This may be because the new town development was
designed with an architectural form that a few residential high-rises share one podium,
which is usually a mall where amenities are concentrated (Shelton et al., 2011).

Walkability scoring underestimations at street scale


We chose three areas to visualise the underestimations and their relationships with 3D street
layouts. Each highly rated area was accompanied by a Google Street View to show the
actual built environment. Figure 5(a) shows that footbridges (Level 2) overlapped with areas
Zhao et al. 13

Figure 5. Walkability scoring underestimations at street scale.

containing significant underestimations. Figure 5(b) displays an area where the end of the
footbridge and the traditional street gained the highest score underestimations, implying
that the traditional streets and linkage level helped improve the estimation. Figure 5(c)
shows that metro underground paths and entrances were underestimated greatly (Levels
–2 and –1).
There was a statistically significant difference between 3D network-level groups on WS
underestimations as determined by the univariate linear model (F ¼ 338.71, p < 0.001).
The mean score underestimations for Level 0 is the reference category (Table 3), showing
the lowest underestimation of all levels (33.75, p < 0.001). Level –2 underground spaces
expanded the WS underestimations to 36.72 (p < 0.001). Level 2 footbridge paths increased
WS underestimations by 3.85 compared to Level 0 (37.6, p < 0.001). Level –1’s increase in
the score underestimations was the most from 33.75 to 38.67 (p < 0.001), and Level 1’s
increase was the least by 2.68 (36.43, p < 0.001).

Discussion and conclusion


In this study, we designed a walkability scoring system for high-density cities, which used a
3D pedestrian network model of Hong Kong for the scoring calculation. We found that
using road network widely underestimated the scoring, and the pedestrian networks could
help approach a more accurate walkability assessment.
14 EPB: Urban Analytics and City Science 0(0)

Table 3. Univariate linear modelling walkability scoring underestimations


with pedestrian network levels.

WS underestimations (std. err.)

Constant (level 0) 33.75* (0.03)


Level –2 2.97* (0.34)
Level –1 4.92* (0.50)
Level 1 2.68* (0.14)
Level 2 3.85* (0.13)

Observations 159,232
R2 0.01
Adjusted R2 0.01
Residual Std. Error 11.89 (df ¼ 159227)
F Statistic 338.71* (df ¼ 4; 159227)
*p < 0.01.

Specifically, we found that WS based on the pedestrian network is 92% higher than those
based on road network. This result echoes previous studies in low-density cities (Chin et al.,
2008; Tal and Handy, 2012). Walkability scoring enables researchers to use one standardised
measure to assess built environment performance, for example, to test the association
between urban morphology and physical activity, housing price and healthy behaviours
(Duncan et al., 2013; Pivo and Fisher, 2011). Except for small experimental studies, previous
WSs were based on road networks. The results might be accurate if the road and pedestrian
networks in low-density cities are similar; however, if we used the road network in a high-
density city, the WS result would be unreliable, due to the striking network discrepancies.
We found that WS underestimations were by no means random. For example, they were
larger for Hong Kong’s old towns than for new towns. The scoring difference revealed that
old town networks are better connected and might be more pedestrian-friendly, resulting
from dense street layouts, plentiful network choices and abundant amenities.
Our study of street-level WS underestimations shows that footbridges and underground
paths cause greater walkability scoring differences. These extensions of walking spaces into
the air (footbridges) and underground have profoundly changed the pedestrian circulations
in Hong Kong and have made the walking system into a unique 3D space. A previous study
of Hong Kong’s Central-Mid-Escalator system, a typical 3D case in Hong Kong, shows that
the 3D walking system is an urban regenerator, changing the pedestrian volume as well as
surrounding amenities (Zacharias, 2013).
We examined the areas with the top 20% scoring underestimation and quantified the
magnitudes that each 3D network level can make compared to a scenario if there was only
surface sidewalk. The surface-level WS was underestimated by 33.75 WS on average. This
means that highly scored areas could score above 50 when the average road network-based
WS is 23. Footbridges and underground linkages could enlarge the underestimations by
11% and 15% compared to the surface level. The greatest increase from underground
linkage (4.92, 15%) indicates the importance of these linkages in the 3D pedestrian
system. There is a large underground walking system, but the links for that are not as
abundant as links for the upper level. Establishing more underground linkages should be
encouraged in urban construction. For upper-level walkways, the large increase (3.85, 11%)
in scores for footbridges indicates that they are an important above-ground component to
the 3D pedestrian network. It would be beneficial to establish more footbridge connections
Zhao et al. 15

to enhance walking experience. These confirm that footbridges and underground linkages
are crucial components of the 3D walking system (Sun et al., 2019; Zacharias and He, 2018).
Practical implications of these findings include making these critical linkages a priority when
enhancing walkability.
Our findings suggest that designing accurate walkability assessment needs pedestrian
network data, especially for high-density cities. The pedestrian network data should ideally
be a 3D model that captures footbridges and underground paths. However, building such a
spatial infrastructure for a city is very challenging. Automatically generating the network
data using computer algorithms is infeasible because of the complexity and intensity in the
3D physical connections (Sun et al., 2019). However, our quantitative results show that it is
worth doing. As a crucial part of a city’s infrastructure, pedestrian network is the key to
analysing walking behaviours and building walkable cities. Urban planning in the 20th
century relied heavily on land-use transport models. They were built upon networks and
zonal models that assumed motor-based interactions. The goal of 21st-century urban plan-
ning is more complex and multi-dimensional. Optimising plans for walkability are now
prominent goals. It is inconsiderable that land use and transport optimisation can proceed
as a practical idea, without pedestrian network model. Our study presents a challenge to
urban planners and designers, showing why and how urban accessibility planning needs to
change for the 21st century.
We used open-source programs in the walkability system development, which aligns with
the current trend in urban analytics. It holds several merits over proprietary software pack-
ages. Besides the cost benefits, these tools’ openness and easiness to share are traits resonate
the core value of academia. The documentation and code are publicly available, with active
communities that enable frequent updates and error corrections. Moreover, the open-source
language, database and data visualisation software can work seamlessly together. In this
study, we benefitted from the flexibility of using plug-ins and extensions in open-source
tools. For instance, the pgRouting extension enabled us to efficiently search for amenities
based on network distance, and the database manager in QGIS allowed us to visualise tables
from PostgreSQL database directly.
There are limitations to this study. The focus has been to unravel the underestimations
caused by using road instead of pedestrian networks. We could have refined the decay
functions and 400 m walking threshold had we known the actual uses of each amenities
category (Sun et al., 2018). Amenity sizes could be incorporated to develop a more precise
weight function. Additionally, we did not consider the travel time difference between
upstairs and downstairs. An assumption was made that escalators or elevators are preferred
over vertical walking in Hong Kong, and thus, the travel time difference can be ignored.
Moreover, we have only applied the algorithms in one city, Hong Kong, and the general-
isation of results to other cities may need caution. Last, the emergence of new data opens up
opportunities for generating up-to-date and widely applicable walkability measurements,
but it is not the study focus. Despite the limitations, our finding of the striking under-
estimations made by a traditional road network calls for attention be given to using pedes-
trian network data in walkability assessment in other cities.
The strength of this study is that it is the first study to systematically profile the walk-
ability scoring underestimation, detailing the assessment differences between using a pedes-
trian network and a road network in a high-density city. This would be fundamental to
research and practice for promoting walkable cities. Future research focuses include explor-
ing the association between WS and related outcomes (e.g. physical activity, housing prices
and healthy behaviours), disentangling the impact of accurately measured WSs on those
outcomes through rigorous measurements, and verifying and optimising the algorithm
16 EPB: Urban Analytics and City Science 0(0)

based on user responses. The verification could be an expansion of the street view–WS
comparison as demonstrated in Figure 5 to study the coherence between calculated score
and the actual walking environment. Accurate walkability scoring will assist urban profes-
sionals in planning and designing effective walkable cities. Additionally, we are working on
building a web platform to support public use in assessing their neighbourhoods.
Our study sets the groundwork for future studies to emphasise the importance of the
pedestrian network in walkability scoring. It provides a systematic analysis which shows
that 3D network elements such as footbridges and underground paths are crucial compo-
nents of Hong Kong pedestrian network. The methodologies, including data processing
protocols, measures and modelling, provide a benchmark for walkability studies in other
high-density cities under similar contexts.

Acknowledgements
The authors thank Sebastian Stansfield and Dugald Wallace for their technical support including
constructing and optimising the walkability algorithm.

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 the research, authorship, and/or
publication of this article: GRF 17600818, HKU seed fund 104004782, and HKU PTF 102009741.

ORCID iDs
Jianting Zhao https://orcid.org/0000-0002-4251-5462
Guibo Sun https://orcid.org/0000-0001-8493-2953 Christ Webster https://orcid.org/0000-0002-
2171-7495

References
Audi M, Byorkman K, Couture A, et al. (2010) Measurement and Analysis of Walkability in Hong
Kong. Available at: http://web.cs.wpi.edu/rek/Projects/DesignHK_C11.pdf (accessed 1
September 2020).
Carr LJ, Dunsiger SI and Marcus BH (2011) Validation of walk score for estimating access to walk-
able amenities. British Journal of Sports Medicine 45(14): 1144–1148.
Cerin E, Macfarlane DJ, Ko H-H, et al. (2007) Measuring perceived neighbourhood walkability in
Hong Kong. Cities 24(3): 209–217.
Cervero R and Kockelman K (1997) Travel ridership and the 3Ds: Density, diversity and design.
Transportation Research Part D: Transport and Environment 2(3): 199–219.
Chin GKW, Van Niel KP, Giles-Corti B, et al. (2008) Accessibility and connectivity in physical
activity studies: The impact of missing pedestrian data. Preventive Medicine 46(1): 41–45.
Dovey K and Pafka E (2020) What is walkability? The urban DMA. Urban Studies 57(1): 93–108.
Duncan DT, Aldstadt J, Whalen J, et al. (2013) Validation of walk scores and transit scores for
estimating neighborhood walkability and transit availability: A small-area analysis. GeoJournal
78(2): 407–416.
Ellis G, Hunter R, Tully MA, et al. (2016) Connectivity and physical activity: Using footpath networks
to measure the walkability of built environments. Environment and Planning B: Planning and Design
43(1): 130–151.
Zhao et al. 17

Ewing R and Cervero R (2010) Travel and the built environment: A meta-analysis. Journal of the
American Planning Association 76(3): 265–294.
Hall CM and Ram Y (2018) Walk scoreVR and its potential contribution to the study of active transport
and walkability: A critical and systematic review. Transportation Research Part D: Transport and
Environment 61: 310–324.
Ker I and Ginn S (2003) Myths and realities in walkable catchments: The case of walking and transit.
Road & Transport Research 12(2): 69–80.
Koohsari MJ, Sugiyama T, Hanibuchi T, et al. (2018) Validity of Walk ScoreVR as a measure of
neighborhood walkability in Japan. Preventive Medicine Reports 9: 114–117.
Li C (2019) USDA GAIN report.
Mavoa S, Boulange C, Eagleson S, et al. (2018) Identifying appropriate land-use mix measures for use
in a national walkability index. Journal of Transport and Land Use 11(1): 681–700.
pgRouting Community (2016) pgRouting Project – Open Source Routing Library. pgRouting.
Available at: https://pgrouting.org/ (accessed 15 August 2020).
Pivo G and Fisher JD (2011) The walkability premium in commercial real estate investments. Real
Estate Economics 39(2): 185–219.
Shashank A and Schuurman N (2019) Unpacking walkability indices and their inherent assumptions.
Health & Place 55: 145–154.
Shelton B, Karakiewicz J and Kvan T (2011) The Making of Hong Kong: From Vertical to Volumetric.
London, UK: Routledge.
Sun G, Webster C, Ni MY, et al. (2018) Measuring high-density built environment for public health
research: Uncertainty with respect to data, indicator design and spatial scale. Geospatial Health
13(1): 35–47.
Sun G, Webster C and Zhang X (2019) Connecting the city: A three-dimensional pedestrian network
of Hong Kong. Environment and Planning B: Urban Analytics and City Science. DOI: 10.1177/
2399808319847204.
Tal G and Handy S (2012) Measuring nonmotorized accessibility and connectivity in a robust pedes-
trian network. Transportation Research Record: Journal of the Transportation Research Board.
DOI: 10.3141/2299-06.
Tang B-S, Wong KK, Tang KS, et al. (2020) Walking accessibility to neighbourhood open space in a
multi-level urban environment of Hong Kong. Environment and Planning B: Urban Analytics and
City Science. DOI: 10.1177/2399808320932575.
Villanueva K, Knuiman M, Nathan A, et al. (2014) The impact of neighborhood walkability on
walking: Does it differ across adult life stage and does neighborhood buffer size matter? Health
& Place 25: 43–46.
Walk Score (2019) Walk Score Methodology. Available at: https://www.walkscore.com/methodology.
shtml (accessed 11 November 2019).
Wang L (2014) An Analysis of Hong Kong Transit City. Beijing: China Architecture & Building Press.
Zacharias J (2013) The central-mid-levels escalator as urban regenerator in Hong Kong. Journal of
Urban Design 18(4): 583–593.
Zacharias J and He J (2018) Hong Kong’s urban planning experiment in enhancing pedestrian move-
ment from underground space to the surface. Tunnelling and Underground Space Technology 82:
1–8.

Jianting Zhao is a PhD student within the Department of Urban Planning and Design at the
University of Hong Kong. Her research interest lies in urban liveability including walkabil-
ity, bikeability and public space. She thrives on using quantitative measures and innovative
techniques to study cities.
18 EPB: Urban Analytics and City Science 0(0)

Guibo Sun is an Assistant Professor within the Department of Urban Planning and Design
at the University of Hong Kong. His research focuses on urban planning and design, trans-
port and health and applied GIS.

Chris Webster is a Professor within the Faculty of Architecture at the University of Hong
Kong. His research broadly examines the dynamics and processes relating to the built
environment, property rights and behaviours.

You might also like