ArticlePDF Available

Abstract and Figures

Studies evaluating bikeability usually compute spatial indicators shaping cycling conditions and conflate them in a quantitative index. Much research involves site visits or conventional geospatial approaches, and few studies have leveraged street view imagery (SVI) for conducting virtual audits. These have assessed a limited range of aspects, and not all have been automated using computer vision (CV). Furthermore, studies have not yet zeroed in on gauging the usability of these technologies thoroughly. We investigate, with experiments at a fine spatial scale and across multiple geographies (Singapore and Tokyo), whether we can use SVI and CV to assess bikeability comprehensively. Extending related work, we develop an exhaustive index of bikeability composed of 34 indicators. The results suggest that SVI and CV are adequate to evaluate bikeability in cities comprehensively. As they outperformed non-SVI counterparts by a wide margin, SVI indicators are also found to be superior in assessing urban bikeability and potentially can be used independently, replacing traditional techniques. However, the paper exposes some limitations, suggesting that the best way forward is combining both SVI and non-SVI approaches. The new bikeability index presents a contribution in transportation and urban analytics, and it is scalable to assess cycling appeal widely.
Content may be subject to copyright.
Assessing bikeability with street view imagery and computer vision
Koichi Itoa,Filip Biljeckia,b,
aDepartment of Architecture, National University of Singapore, Singapore
bDepartment of Real Estate, National University of Singapore, Singapore
Urban planning
Deep learning
Google Street View
Studies evaluating bikeability usually compute spatial indicators shaping cycling conditions and
conflate them in a quantitative index. Much research involves site visits or conventional geospa-
tial approaches, and few studies have leveraged street view imagery (SVI) for conducting virtual
audits. These have assessed a limited range of aspects, and not all have been automated using
computer vision (CV). Furthermore, studies have not yet zeroed in on gauging the usability of
these technologies thoroughly. We investigate, with experiments at a fine spatial scale and across
multiple geographies (Singapore and Tokyo), whether we can use SVI and CV to assess bike-
ability comprehensively. Extending related work, we develop an exhaustive index of bikeability
composed of 34 indicators. The results suggest that SVI and CV are adequate to evaluate bike-
ability in cities comprehensively. As they outperformed non-SVI counterparts by a wide margin,
SVI indicators are also found to be superior in assessing urban bikeability and potentially can
be used independently, replacing traditional techniques. However, the paper exposes some lim-
itations, suggesting that the best way forward is combining both SVI and non-SVI approaches.
The new bikeability index presents a contribution in transportation and urban analytics, and it is
scalable to assess cycling appeal widely.
This is the Accepted Manuscript version of an article published by Elsevier in the journal Transportation Research Part C: Emerging Technologies
in 2021, which is available at: Cite as: Ito K, Biljecki F (2021): Assessing bikeability
with street view imagery and computer vision. Transportation Research Part C, 132: 103371.
©2021, Elsevier. Licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International ( nd/4.0/)
1. Introduction
Bicycles play an important role in making cities environmentally sustainable, healthy, and economically vibrant
(Neves and Brand,2019;Cao and Shen,2019;Rojas-Rueda et al.,2011;Volker and Handy,2021;Wang et al.,2016;
Horacek et al.,2018;Wang et al.,2019a;Yeh et al.,2019;Chen et al.,2020b;Alaoui and Tekouabou,2021). To
evaluate the extent to which cycling is facilitated, the notion of bikeability was created by complementing the concept
of walkability. Subsequently, index systems, as instruments to quantify it, have been developed by many studies (Porter
et al.,2020;Cain et al.,2018;Manton et al.,2016;Gullón et al.,2015;Koh and Wong,2013;Horacek et al.,2012;
Wahlgren et al.,2010;Hoedl et al.,2010;Clifton et al.,2007;Titze et al.,2012;Arellana et al.,2020;Winters et al.,
2013;Guler and Yomralioglu,2021;Lin and Wei,2018;Gholamialam and Matisziw,2019;Resch et al.,2020;Kamel
et al.,2020;Grigore et al.,2019;Osama et al.,2020;Chevalier and Xu,2020;Kang et al.,2019;Schmid-Querg et al.,
2021;Galanis et al.,2018;Lowry et al.,2016;Faghih Imani et al.,2019;Boongaling et al.,2021). This topic became
even more relevant with the advent of bicycle-sharing systems (Du et al.,2019;Luo et al.,2020). Data collection
methodologies to calculate bikeability indexes are for a large part derived from field observations, entailing time-
consuming manual work and limiting the spatial extent that can be surveyed. As technologies have advanced, these
assessments have been supplemented by new data sources, such as crowdsourcing and virtual observations (Kalvelage
et al.,2018;Gullón et al.,2015;Abadi and Hurwitz,2018). Nevertheless, previous studies face various issues, such as
slow data collection process, the balance between subjectivity and objectivity of data, lack of street-level information,
and standardization of spatial granularity. The recent availability of street view imagery (SVI) has yielded opportunities
for new approaches to urban studies (Biljecki and Ito,2021), enabling a wealth of images from the pedestrian and cyclist
perspective that may be used to assess walkability and bikeability and that is available remotely (Figure 1). In parallel,
developments in computer vision (CV) have catalyzed the means to process the profusion of photos automatically and
efficiently. They have already been utilized for assessing walkability (Nagata et al.,2020). SVI and CV are not entirely
new to bikeability either. For example, studies by Tran et al. (2020) and Gu et al. (2018) have used CV and SVI to assess
particular aspects of bikeability. However, no comprehensive bikeability assessment study has been conducted yet, and
there has been no critical evaluation of the usability of such technologies in comparison to conventional methods that
have been dominating the field so far.
Corresponding author (F. Biljecki)
ORCID (s):
First Author et al.: Preprint submitted to Elsevier Page 1 of 27
Figure 1: Illustration of an urban setting together with one of the corresponding street-level views, highlighting several
aspects that may indicate bikeability. The method presented in this paper takes advantage of a substantial number of
visual features that may be extracted automatically from street view images and engage them to generate a composite
index that suggests cycling appeal at a fine spatial scale and across multiple cities.
Considering the developments in computer vision and proliferation of street view imagery (e.g. increased coverage
of commercial services such as Google Street View and introduction of crowdsourced alternatives), we believe that
such research is needed and timely. This study tests a hypothesis that CV and SVI coupled together have a strong
potential to overcome the issues faced by conventional methods in gauging how friendly urban streets are to cyclists.
Further, as the availability of SVI is now considered to be virtually omnipresent, a notable research gap is comparative
studies that involve more than one city. Thus, in this study we aim to answer the following research questions: can we
use CV techniques and SVI data to comprehensively assess bikeability within and among cities? If yes, can SVI and CV
alone be used to assess bikeability, replacing traditional techniques entirely? To answer these questions, we develop a
bikeability index with 34 indicators under five categories and implement it in Singapore and Tokyo. The contributions
of our study are the novel investigation of CV techniques’ and SVI indicators’ usefulness in comprehensive bikeability
assessment, construction of a new comprehensive bikeability index regarding an unprecedented amount of aspects,
and provision of a new data collection method that can extract subjective and objective indicators from street-level
information at a larger spatial scale and across multiple geographies, thereby overcoming the issues found by the
previous studies. We believe that the method and the index may be scaled around the world, including additional cities
in future work. It is important to note that our method includes also a survey with a large number of human participants
to investigate whether CV techniques may estimate human perception of cycling appeal automatically.
In Section 2, we conduct a comprehensive literature review to affirm the research gap. The comprehensive and
structured overview of the state of the art is another contribution of this paper. Section 3explains the data sources and
methodologies used in this study. Section 4describes the results of this study together with a discussion. Finally, we
conclude our study as well as discuss further directions for future research in Section 6.
2. Related work
2.1. Bikeability studies
Many studies have explored various aspects of the built environment that can influence people’s cycling behavior
(Bauman et al.,2012;Nielsen and Skov-Petersen,2018;Pritchard et al.,2019;Daraei et al.,2021;Kraus and Koch,
2021;Nazemi et al.,2021;McNeil,2011;Ma and Dill,2017;Cicchino et al.,2020;Nogal and Jiménez,2020;Sottile
et al.,2019;Berger and Dörrzapf,2018;Porter et al.,2018;Aldred et al.,2020;Long and Zhao,2020;Doubleday et al.,
2021;Brüchert et al.,2020;Martin et al.,2021;Attard et al.,2021). Studies on the association of the built environment
and cycling conditions became well-established, and many researchers developed indexes to assess specific aspects
of the built environment that can affect cycling behavior and comprehensively quantify bikeability, i.e. the extent to
which an environment is friendly for bicycling.
In the early days of related work between 2010 and 2015, methodologies to collect data are mostly conducted
through field surveys, thus tending to be time- and resource-intensive. Studies by Hoedl et al. (2010), Horacek et al.
(2012), and Koh and Wong (2013) involve field observation by experts, and such a methodological limitations precludes
bikeability assessment at large scales. Moreover, only a few studies (Koh and Wong,2013) include both objective and
First Author et al.: Preprint submitted to Elsevier Page 2 of 27
subjective indicators, and most of the studies (Wahlgren et al.,2010;Hoedl et al.,2010;Horacek et al.,2012) only
focused on either one of the indicators. The spatial granularity of sample points is loosely defined and not standardized
in some studies as well (Wahlgren et al.,2010;Horacek et al.,2012).
Recent bikeability studies have become more standardized and scalable. Although some studies still use field ob-
servation (Manton et al.,2016;Cain et al.,2018), more studies apply emerging technologies and data sources such
as remote sensing images, manual virtual auditing using SVI, and crowdsourcing, to collect data (Krenn et al.,2015;
Gullón et al.,2015;Winters et al.,2016;Kalvelage et al.,2018), enabling large-scale and comparative assessment of
bikeability. However, as data collection methods become more scalable, subjective indicators are excluded by many
studies (Krenn et al.,2015;Manton et al.,2016;Winters et al.,2016;Cain et al.,2018;Kalvelage et al.,2018). More-
over, remotely sensed imagery cannot capture street-level information, and manual virtual auditing and crowdsourcing
data collection require a large amount of time and resources. For example, a recent study by Arellana et al. (2020)
utilizes virtual auditing in SVI, but this data collection process was reported to be six months long for a city-scale study
area, suffering from the same issue of a time-intensive method as the aforementioned studies. To overcome the issue of
the balance between street-level information and scalability, recent studies couple SVI with CV techniques to automate
the indicator extraction (Gu et al.,2018;Tran et al.,2020), but the number of indicators extracted by such methods
remains still limited. Moreover, subjective indicators for bikeability assessment have not been extracted from SVI by
using CV yet. Therefore, there is a need for further studies to examine the possibility of extracting more indicators —
both objective and subjective — from SVI by using CV.
Structuring the rundown on related work, Table 1 summarizes indicators used in the reviewed studies. Most of
the indicators could be categorized into connectivity, environment, infrastructure, vehicle–cyclist interaction (V–C
interaction), and perception. In developing bikeability indexes, the previous studies have faced the issues of the time-
intensive data collection process, the balance between subjectivity and objectivity, extraction of street-level informa-
tion, and standardization of spatial granularity. Table 2 summarizes the issues mentioned above. Literature reviews
on bikeability by Kellstedt et al. (2021) and Castañon and Ribeiro (2021) indicate that the past development of bike-
ability assessment has been driven by innovative uses of advanced new technologies, thereby suggesting that newer
technologies may overcome these issues mentioned above. While there have been studies that have utilized SVI and
CV techniques to assess bikeability, and thus indicating the reliability of SVI as a data source that can be used in
a scalable manner, they suffer from shortcomings that we seek to mitigate in our work. Primarily, previous studies
used these technologies to assess very limited aspects of bikeability (i.e. only up to a few indicators), did not collect
and assess subjective indicators, and have been evaluated on limited areas. Such a gap necessitates further studies to
examine how much SVI and CV techniques are usable to assess bikeability comprehensively.
2.2. Street view imagery in urban studies
The growth of the spatial coverage of street view imagery and the development of computer vision techniques have
catalyzed the recent proliferation of studies that utilize them, both in transportation and urban planning and beyond
(Wang et al.,2019b;Song et al.,2020;Wu and Biljecki,2021;Ye et al.,2020;Fan et al.,2021;Chen et al.,2021a).
This section focuses on studies that use SVI to extract information of data used in previous bikeability studies under
four categories that have been delineated by this study (i.e. environment, infrastructure, vehicle-cyclist interaction, and
perception), and to examine subjects pertaining to cycling.
One of the more explored aspects in related research is urban greenery. Quantification of greenery by using image
segmentation, CV techniques that can classify categories of objects at a pixel-level, has enabled many various urban
studies ranging from simple assessment of the distribution of vegetation and interdisciplinary examinations of the
relationships between greenery and various aspects of cities, such as physical activities of residents and real estate (Ye
et al.,2019a;Lu,2019;Ye et al.,2019b). Other studies quantify greenery, sky view factor (i.e. openness), and buildings
(i.e. enclosure) with semantic segmentation to measure characteristics of cities in a scalable manner (Li et al.,2017;
Gong et al.,2018,2019;Li and Ratti,2019;Toikka et al.,2020;Wang and Vermeulen,2020;Ma et al.,2021;Zhou
et al.,2021). These features that are extracted from SVI, which have been leveraged for a variety of applications, can
be also used to evaluate bikeability.
The high scalability of CV techniques and SVI has also multiplied opportunities for infrastructure assessment at a
city scale. Hall et al. (2018) propose a methodology to detect and classify traffic signals, and Chacra and Zelek (2018)
develop a CV-based model to detect infrastructure anomalies from SVI. Assessment of urban accessibility is conducted
by Najafizadeh and Froehlich (2018), developing models to detect accessibility problems from SVI, such as missing
curb ramps and street surface issues. Further, Ding et al. (2021) use object detection and classification CV models to
First Author et al.: Preprint submitted to Elsevier Page 3 of 27
map bike lane networks from SVI. Mapillary, a crowdsourced SVI service, developed a dataset called Vistas, which
contains 25,000 images collected from around the world, and annotated according to 66 categories (Neuhold et al.,
2017). Cityscapes is another frequently used street-level dataset for segmentation (Cordts et al.,2016;Gong et al.,
2019;Nagata et al.,2020). Although Mapillary Vistas has many infrastructure-related categories (e.g. bike lane and
bike parking) that other segmentation training datasets do not regard, it has not been widely used in urban studies.
Thus, using this dataset might expand the prospects of SVI for bikeability assessment.
These techniques have been utilized in transport studies as well. Goel et al. (2018) find that the number of cyclists
manually counted in GSV images is strongly correlated with the cycling mode share reported by cities in the UK
(r = 0.92). Zhang et al. (2019a) and Chen et al. (2020a) also examine the relationship between visual features and
urban traffic volume and reveal that SVI can be used to explain more than 65 percent of the spatiotemporal mobility
pattern. These findings hint that SVI may be used to estimate traffic volume, which is an important aspect of bikeability
(Labetski and Chum,2020).
First Author et al.: Preprint submitted to Elsevier Page 4 of 27
Table 1
An overview of indicators used by previous studies grouped into five categories.
Publication Connectivity Environment Infrastructure V-C interaction Perception
Clifton et al. (2007) Cul-de-sac Land use Sidewalk Attractiveness
Continuity Slope Pavement Traffic speed Safety
Greenery Path obstruction Street parking Cleanliness
Enclosure Sidewalk buffer Traffic control
Building design Road condition
Setback Curb cuts
Street light
Street amenity
Directional sign
Power line
Bike lane
Transit facilities
Hoedl et al. (2010) N/A Greenery Bikelane N/A
Land use Sidewalk Traffic speed
Billboards Traffic volume
Open space
Wahlgren et al. (2010) Directness Air quality N/A Traffic speed Attractiveness
Intersection Noise Traffic volume Crowdedness
Greenery Cylist speed Safety
Slope Traffic separation Beauty
Horacek et al. (2012) N/A Slope Pavement Traffic speed Beauty
Street light Traffic volume
Potholes Traffic control
Path size
Sidewalk buffer
Curb cut
Bike lane
Koh and Wong (2013) Detour Slope Directional sign N/A Safety
Intersection POIs Pavement Crowdedness
Greenery Shelter
Gullón et al. (2015) Different routes Greenery Pavement Cleanliness
Land use mix Street amenity Traffic control Beauty
Street light Attractiveness
Krenn et al. (2015) N/A Greenery Bike lane Traffic separation N/A
Manton et al. (2016) Intersection N/A Road width Traffic volume N/A
Street parking
Traffic separation
Winters et al. (2016) Intersection Slope Bike lanes Mode share N/A
Distance to POIs
Hartanto et al. (2017) Intersection Water Road type Traffic speed N/A
Directness Greenery Pavement Traffic volume
Buildings Street light
Slope Bike parking
Cain et al. (2018) Intersection Land use Transit facilities Traffic control N/A
Informal path No. of pedestrian Roll-over curb
Cul-de-sac Water Street amenity
Landscape Bike parking
Hardscape Curb cuts
Graffiti Bike lane
Setback Crossing
Shade Road width
Building design Sidewalk buffer
Bike lane
Gu et al. (2018) Street density Shade Bike lane Traffic separation N/A
POIs Crossing
Arellana et al. (2020) N/A Slope Bike lane Traffic control N/A
Greenery Sidewalk Mode share
Building design Street obstruction Traffic speed
Crime presence Bike lane width
Street light
Police presence
Security camera
Tran et al. (2020) Directness POIs Bike lane N/A N/A
First Author et al.: Preprint submitted to Elsevier Page 5 of 27
Table 2
Characteristics of methods used by previous studies.
Publication Data Weighting system No. of indicators
Clifton et al. (2007) Objective & Individual assessment 0 Yes
Hoedl et al. (2010) Objective Individual assessment 0 Yes
Wahlgren et al. (2010) Subjective Individual assessment 0 Yes
Horacek et al. (2012) Objective Unequal arbitrary weight 0 Yes
Koh and Wong (2013) Objective & Survey-based weight 0 Yes
Gullón et al. (2015) Objective & Unequal arbitrary weight 0 Yes
Krenn et al. (2015) Objective Equal weight 0 No
Manton et al. (2016) Objective Regression modeling 0 Yes
Winters et al. (2016) Objective Equal weight 0 No
Hartanto et al. (2017) Objective Equal weight 0 Yes
Cain et al. (2018) Objective Unequal arbitrary weight 0 Yes
Gu et al. (2018) Objective Entropy weight 3 No
Arellana et al. (2020) Objective Survey-based weight 0 Yes
Tran et al. (2020) Objective Equal weight 3 Yes
Extraction of information from images has also enabled the prediction of urban perception based on SVI. Naik
et al. (2014) conduct surveys on safety, in which 7,872 unique participants from 91 countries ranked 4,109 images
using 208,738 pairwise comparisons. Dubey et al. (2016) develop a dataset called Place Pulse 2.0, which consist of
110,988 images from 56 cities and 1,170,000 pairwise comparisons answered by 81,630 online survey participants on
six perception scores, and predicted these scores with convolutional neural network models, such as VGGNet, and this
study’s dataset and methodology have been utilized and replicated by other studies (Kang et al.,2021;Qiu et al.,2021).
Yao et al. (2021) conduct a similar study to predict perception scores by surveying 20 volunteers on 1,000-2,000 images
and predicting perception scores using a random forest model and features extracted by FCN-8s as inputs. Verma et al.
(2020) also recruit only 79 participants to rate 200 images, extracting high- and low-level features from SVI by using
CV techniques (e.g. image segmentation, object detection, classification, and edge detection), and designing models
to predict fourteen perception scores, such as “pleasant”, “boring”, and “safe”. Moreover, the validity of conducting
perception surveys has been examined and found to be as reliable as surveys based on the real environment by Feng
et al. (2021).
Lu et al. (2019) and Wang et al. (2020) examine the association between urban greenery (extracted from SVI) and
cycling behaviors, and Hollander et al. (2020) study the correlation between transportation planning and the perceived
safety of the built environment. A study by Tran et al. (2020) is the only study that relies on SVI to create a bikeability
index, but this study uses SVI only to assess very limited aspects of bikeability, such as greenery and enclosure.
Therefore, there has not been any study that developed a comprehensive bikeability index mirroring the wide array of
aspects jointly covered by related work and one that has utilized SVI as a major data source to calculate it, a gap that
is bridged by our study.
2.3. Summary
The literature review elucidates the proliferation of urban studies that utilize SVI and CV techniques and, at the
same time, it exposes the absence of studies that take advantage of them for assessing cycling appeal and developing
a bikeability index. Such omission is possibly caused by obstacles of using CV techniques and the complexities
of developing a comprehensive bikeability index system. However, previous studies demonstrate the possibility of
replacing traditional data collection for bikeability indicators, as they collect information on similar aspects that may
indicate cycling appeal.
This paper aims to bridge this gap by designing a thorough bikeability index that is largely derived using SVI and
CV, and it also uses the opportunity to critically investigate their value and independence when doing so. Moreover,
First Author et al.: Preprint submitted to Elsevier Page 6 of 27
as SVI data has virtually global coverage, the study puts scalability under the spotlight; thus, the combination of
scalability of SVI and efficient extraction of indicators through CV can reduce the time and resource cost of bikeability
assessment compared to conventional data sources and methods. Finally, potential barriers of this approach for urban
planners are also examined to understand its application and prospects for adoption.
3. Methodology
This study uses six data sources (i.e. SVI, surveys, OpenStreetMap (OSM), Land Use (LU), Digital Elevation
Model (DEM), and Air Quality Index (AQI)) to evaluate 34 indicators under five categories (i.e. connectivity, envi-
ronment, infrastructure, perception, and vehicle-cyclist interaction). The comprehensive selection of these indicators
is a contribution on its own, as its scope is unprecedented, and may serve as a resource to develop indexes in other
domains. Testing the scalability of the method, Singapore and Tokyo, as two geographies with disparate characteris-
tics, are selected as study areas. Thus, data retrieval and indicator extractions are conducted to develop the composite
index by weighting each category and indicator equally. Figure 1 highlights some examples of phenomena in SVI that
are characterized in this method.
3.1. Selection of indicators
Conducting a review of studies that developed bikeability indexes, this study devised an exhaustive list of indicators
used by them (Hartanto et al.,2017;Winters et al.,2016;Gullón et al.,2015;Wahlgren et al.,2010;Horacek et al.,
2012;Hoedl et al.,2010;Clifton et al.,2007;Manton et al.,2016;Cain et al.,2018;Koh and Wong,2013). This
inventory was used to create an own index to expand related work and minimize the bias when selecting indicators.
Duplicates and those indicators that cannot be obtained and/or are unsuitable for this study have been excluded, such as
noise and the presence of informal paths and crimes. After filtering, out of 65 unique indicators found in the previous
studies in total, 34 indicators (i.e. about 52%) were kept, forming the most extensive instance to date. Future studies
can use our method while partially modifying this list to incorporate more or fewer indicators as long as they can be
extracted from the data sources we used.
These 34 unique indicators are categorized into five categories, namely, connectivity, environment, infrastructure,
perception, and vehicle-cyclist interaction (see Table 3). Regarding data sources, 21 indicators are to be extracted from
SVI, 10 from OSM, one from each of LU, DEM, and AQI.
Although this study does not examine if the selected indicators can predict bike usage, most of the reviewed previous
studies also selected them based on literature reviews, except for a few (Winters et al.,2016;Arellana et al.,2020) that
use regression analysis. Moreover, higher bikeability does not necessarily lead to higher counts of bike usage. For
example, Arellana et al. (2020) examine and find that other factors, such as population density and socio-economic
characteristics, also play a role (Munira et al.,2021). Therefore, investigating whether the bikeability index can predict
bike usage is out of the scope of this research.
Table 3 lists indicators with their data sources, extraction methods, and scaling methods. For scaling methods,
min-max scaling and negative min-max scaling were used (see Equation 1 and Equation 2). The sampling method is
further explained in the following sections.
𝑥Min-Max Scaled =𝑥− min(𝑥)
max(𝑥) − min(𝑥)(1)
𝑥Negative Min-Max Scaled = 1 − 𝑥− min(𝑥)
max(𝑥) − min(𝑥)(2)
3.2. OSM and SVI data retrieval
This study relies on OSM to retrieve information on indicators and also for the retrieval of locations on streets that
are used to fetch street view images. The completeness of OSM in the study area is deemed adequate (Barrington-Leigh
and Millard-Ball,2017;Biljecki,2020). To retrieve the OSM data, OSMnx (Boeing,2017) was used, obtaining about
250,000 points and 440,000 points for Singapore and Tokyo, respectively. After that, 7,142 points were randomly
selected. The number of points to be collected was set at 7,142 because this is the maximum number of SVI images
that can be retrieved within the free credit provided by GSV API every month. Future studies can also utilize the
initial trial credit as well, but if there are financial constraints, the rise of volunteered SVI, such as KartaView and
Mapillary, may ameliorate this issue in the future. For this study, the API of GSV was used to collect images because
of its extensive coverage and high quality in both metropolises.
First Author et al.: Preprint submitted to Elsevier Page 7 of 27
Table 3
Indicators forming the comprehensive bikeability index introduced by this study.
Indicators Data Extraction Scale
No. of intersection with lights OSM Aggregation (500m) 0-1 (Negative Min-Max)
No. of intersection without lights OSM Aggregation (500m) 0-1 (Negative Min-Max)
No. of cul-de-sac OSM Aggregation (500m) 0-1 (Negative Min-Max)
Slope DEM Calculation 0-1 (Negative Min-Max)
No. of POI OSM Aggregation (500m) 0-1 (Min-Max)
Shannon land use mix index Land use Aggregation (500m) 0-1 (Min-Max)
Air quality index AQI Spatial interpolation 0-1 (Negative Min-Max)
Scenery: greenery SVI Segmentation 0-1 (Min-Max)
Scenery: buildings SVI Segmentation 0-1 (Min-Max)
Scenery: water SVI Segmentation 0-1 (Min-Max)
Type of road OSM Aggregation (100m) service, track = 0.1
primary, primary_link = 0.2
secondary, secondary_link = 0.4
tertiary, tertiary_link = 0.5
unclassified = 0.6
pedestrian path = 0.8
cycleway = 1
Others = 0
Presence of potholes SVI Segmentation 1 if not present
0 if present
Presence of street light SVI Segmentation 1 if present
0 if not present
Presence of bike lanes SVI Segmentation 1 if present
0 if not present
No. of transit facilities OSM Aggregation (500m) 0-1 (Min-Max Scale)
Type of pavement OSM Aggregation (100m) unhewn_cobblestone, cobblestone = 0.2
sett, metal, wood = 0.4
paved = 0.5
concrete:lanes, plates, paving_stones = 0.6
asphalt, concrete = 1
Others = 0
Presence of street amenities SVI Segmentation 1 if present
0 if not present
Presence of utility pole SVI Segmentation 1 if not present
0 if present
Presence of bike parking SVI Segmentation 1 if present
0 if not present
Road width OSM Aggregation (100m) 0-1 (Divide by 10)
1 if width is larger than 10m
Presence of sidewalk SVI Segmentation 1 if present
0 if not present
Presense of crosswalk SVI Segmentation 1 if present
0 if not present
Presence of curb cuts SVI Segmentation 1 if present
0 if not present
Attractiveness for cycling SVI Surveys, Modeling 0-1 (Min-Max)
Spaciousness SVI Surveys, Modeling 0-1 (Min-Max)
Cleanliness SVI Surveys, Modeling 0-1 (Min-Max)
Building design attractiveness SVI Surveys, Modeling 0-1 (Min-Max)
Safety as a cyclist SVI Surveys, Modeling 0-1 (Min-Max)
Beauty SVI Surveys, Modeling 0-1 (Min-Max)
Attractiveness for living SVI Surveys, Modeling 0-1 (Min-Max)
Vehicle-Cyclist Interaction
No. of vehicles SVI Detection, Aggregation (500m) 0-1 (Negative Min-Max)
Presence of on-street parking OSM Aggregation (100m) 1 if not present
0 if present
Presence of traffic lights / stop signs SVI Segmentation 1 if present
0 if not present
No. of speed control devices OSM Aggregation (100m) 1 if present
0 if not present
First Author et al.: Preprint submitted to Elsevier Page 8 of 27
3.3. Extraction of indicators
The feature extraction process for indicators differs among different categories and data sources. Feature extractions
for land use, topography, and AQI are relatively straightforward. For land use, we use an entropy formula derived from
the Shannon index developed by Frank et al. (2005) within 500m buffers from sample points (see Equation 3).
𝐿𝑈 𝑀 = −1 𝑛
𝑝𝑖∗ ln(𝑝𝑖)∕ ln(𝑛)(3)
In this formula, 𝐿𝑈 𝑀 is the land-use mix score, 𝑝𝑖 is the proportion of the neighborhood covered by the land use
𝑖against the total area for all the land-use categories, and 𝑛is the number of land-use categories. A land-use mix score
of 1 indicates the highest mix possible while a score of 0 indicates the area contains a single land use. This results in
values between 0, the lowest mixed-use level, and 1, the highest mixed-use level.
For topography, the slope is calculated from DEM data created by Yamazaki et al. (2017) and sampled at each
sample point, and the value was scaled with min-max scaling.
AQI data is collected from the Air Quality Index1by automating the retrieval of stations and AQI data in both cities
with a Python library called Selenium. The annual average AQI of stations in the cities is calculated by taking the daily
maximum value of various pollutants (e.g. PM2.5) in the unit of µg/m3, and then spatial interpolation was conducted
to estimate the average AQI at the sample points in this study by using inverse distance weighted interpolation (see
Equation 4).
𝑖=1 𝑧𝑖
𝑖=1 1
In this formula, 𝑧𝑝denotes the interpolated values of the target points, 𝑛represents the number of points used to
interpolate from, 𝑧𝑖shows the value being interpolated from, and 𝑑𝑝
𝑖is the distance between the target point and the
point being used to interpolate from. The calculated AQI is then scaled into values between 0 and 1 by using min-max
Feature extractions from OSM are conducted by taking buffers from sample points. For the density of intersection
with/without lights, cul-de-sac, shops along the route, and transit facilities, we created 500m buffers from sample
points and aggregated the number of these indicators because we need to consider the surrounding contexts as well.
Other indicators collected from OSM, such as the type of roads and the type of pavement, are collected by creating
100m buffers from sample points and converting categorical values to numerical if necessary.
Finally, this study conducts feature extractions from SVI in two ways. For objective indicators, two CV techniques
— segmentation and object detection — were used to extract features. For segmentation, the In-Place Activated
BatchNorm model trained on Mapillary Vistas with WideResNet38 and DeepLab3 developed by Bulò et al. (2018) is
selected because of its high accuracy of the mean intersection over union of 53.42%. Indicators such as the scenery
along bike lanes (i.e. built-up area, greenery, sky view factor, water) are quantified by calculating the ratios of the
pixels categorized as them over the total number of pixels in the image, which are then scaled into between 0 and 1
with min-max scaling. Because it is not meaningful to calculate pixels of other objects such as street lights, bike lanes,
street amenities, and bike parking, they are quantified in a binary manner (i.e. score 1 if they are in the image and 0
if they are not in the image). Object detection is used for one indicator, that is, the number of vehicles, as a previous
study suggests that SVI can be used to estimate the traffic in the area (Zhang et al.,2019a). For this task, we opt for the
GluonCV’s model zoo developed by Guo et al. (2020) and select a YOLOv3 model pre-trained on Pascal VOC dataset
with Darknet53 as the base model, which can detect the number of bicycles, pedestrians, and vehicles with an average
precision of 58.2% at the intersection over a union of 0.5. We chose this model because of its reported relatively
high speed and accuracy compared to other models, such as Faster Region-based Convolutional Neural Networks and
Single Shot Detection (Srivastava et al.,2021;Li et al.,2020). Moreover, the core of our paper is separate from the
models, thus if necessary, new models can be used in any part of this methodology, as our approach is model-agnostic.
First Author et al.: Preprint submitted to Elsevier Page 9 of 27
Because a street view image captures only the traffic volume of the road segment at that moment, it is not entirely
reliable to estimate the traffic based on just one photo at a point. Therefore, a buffer of 500m was generated for each
point, aggregating the number of vehicles found in the plentiful imagery in the buffer. After this processing, the mode
share was also scaled into values between 0 and 1 by min-max scaling. A previous study (Chen et al.,2020a) uses a
similar method to estimate the number of pedestrians and obtains Cronbach’s alpha above 0.8, indicating SVI’s high
reliability as a data source. However, it should be noted that the number of images within each buffer varies because
the sample points were randomly selected, which might produce some bias. Also, duplication of vehicles detected
in SVI can cause bias because some vehicles could have driven together with the vehicles collecting the street view
images. The aforementioned study (Chen et al.,2020a) did not consider this issue and still obtained high reliability;
thus, we leave improvement in methodology to detect the same vehicles with models such as Siamese-convolutional
neural network for future studies.
Subjective indicators are under the perception category, including attractiveness for cycling, spaciousness, cleanli-
ness, building design attractiveness, and perception of safety as a cyclist, and these indicators were predicted by using
features extracted from SVI. We follow the methodology used by Verma et al. (2020), where high- and low-level fea-
tures of images are used to predict perceptions of images (see Table 4). Low-level features can be extracted through
edge detection, blob detection, and Hue-Saturation-Lightness (HSL) extraction. Edges are quantified by detecting
edges and calculating the ratio of pixels that are categorized as edges over the total number of pixels. Blob detection
was used to calculate the number of blobs in SVI, and HSL extraction is conducted to calculate the average and stan-
dard deviation of each hue, saturation, and lightness. It should be noted that lightness might have some bias because of
variances in the time when the images were collected, although GSV has a standardized data collection procedure and
post-processing of images (Google,2018). As for high-level features, image classification (IC), object detection (OD),
and semantic segmentation (SS) are used to extract them. For OD and SS, this study uses the same models mentioned
above, and a ResNet50 model trained on Places365 data with an accuracy of 85.07% is used for IC (Zhou et al.,2018).
These extracted features are dependent variables to predict the perception indicators.
To collect the training data set and build models to predict indicators suggesting perception, a survey is conducted
on Amazon Mechanical Turk, for which this study has received an exemption from the Institutional Review Board
of the National University of Singapore, and in which the participants were compensated financially. For each city,
400 SVIs, i.e. 800 in total, were randomly selected for the survey, which was designed to have at least eight different
participants rate the images on the five indicators on a scale of 0 to 10. The large number of participants to rate each
image ensures reliability. The occasional and inevitable outliers among the responses were detected with the median
absolute deviation method and removed when the output is above three (see Equation 5).
MAD = median 𝑋𝑖̃
𝑋𝑖denotes each observation, and ˜
Xrepresents the median of all the observations.
The collected dataset is split into training and validation data sets in an 80:20 ratio to build predictive models with
LightGBM (Ke et al.,2017), which is gaining momentum in urban studies (Chen et al.,2021b) for its high accuracy
and low computational cost in training (Zhang et al.,2019b;Deng et al.,2018). In this study, we tuned the following
hyperparameters with 10-fold cross-validation: num_leaves, max_depth, min_child_samples, min_child_weight, sub-
sample, colsample_bytree. After conducting the prediction for the rest of the points, the predicted perception indicators
are scaled into values between 0 and 1 with min-max scaling.
Moreover, this study explores the relationships between all the features and perception scores by grouping obser-
vations into below- and above-average values of features and conducting Welch’s t-test (Delacre et al.,2017). We aim
to reveal the underlying visual effects of each feature on human perceptions.
3.4. Development of a new composite index to assess bikeability
The composite index of this study, one of its advancements and key contributions, is developed based on conflating
bikeability indexes developed by previous studies. There are three types of weighting systems developed by them,
which warrant a brief overview. One of them is the independent assessment of indicators. This method only evaluates
each indicator but does not give weights to them, and it is used by Wahlgren et al. (2010) and Hoedl et al. (2010).
Another type is the arbitrary weight. This method gives arbitrary weights to categories and indicators, which is adopted
by Cain et al. (2018); Horacek et al. (2012). The last type is equal weight. This method gives equal consideration to
all the categories and indicators, which actually — strictly speaking — belongs to the arbitrary weighting system, but
First Author et al.: Preprint submitted to Elsevier Page 10 of 27
Table 4
Features for predicting perception.
Visual features Definitions
tree_ss % of pixels classified as trees.
sky_ss % of pixels classified as the sky.
street_ss % of pixels classified as street and sidewalks.
built_ss % of pixels classified as buildings.
others_ss % of pixels classified as other remaining outdoor classes.
nature % of pixels classified as natural elements such as sky, tree, and water.
shannon Shannon entropy values calculated on SS task.
slum_ic Probability of being classified as Slum/Alley/Junkyard in IC task.
market_ic Probability of being classified as Bazaar/Flea market/Market in IC task.
built_other_ic Probability of being classified as Downtown/Embassy/Plaza in IC task.
green_other_ic Probability of being classified as Forest path/Forest road in IC task.
bicycle_od No. of bicycles detected in OD task.
bus_od No. of buses detected in OD task.
car_od No. of cars detected in OD task.
motorcycle_od No. of motorcycles/scooters detected in OD task.
person_od No. of persons detected in OD task.
traffic_light_od No. of traffic lights detected in OD task.
truck_od No. of trucks/auto rickshaws detected in OD task.
canny_edge_llf % of pixels detected as edges.
no_of_blobs_llf No. of blobs.
hue_mean_llf The mean value of the hue dimension in HSL color space.
hue_std_llf The standard deviation of the hue dimension in HSL color space.
lightness_mean_llf The mean value of the lightness dimension in HSL color space.
lightness_std_llf The standard deviation of the lightness dimension in HSL color space.
saturation_mean_llf The mean value of the saturation dimension in HSL color space.
saturation_std_llf The standard deviation of the saturation dimension in HSL color space.
this study differentiated them for clarity. This method is used by Hartanto et al. (2017), Winters et al. (2016), and
Tran et al. (2020), and it is a common system among the reviewed studies. This study adopts the equal weight system
for its simplicity and scaled all the indicators into values between 0 and 1 to prevent any indicators from excessively
influencing the composite index at the end (see Equation 6).
𝐼𝑛𝑑𝑒𝑥 =
𝑖=1 𝑥𝑖∗ (100∕(𝑁𝑐𝑁𝑐𝑖 ))(6)
In this equation, Index represents the bikeability index, 𝑥𝑖denotes the value of each indicator 𝑖,𝑁𝑐show the number
of categories, and 𝑁𝑐𝑖 stand for the number of indicators in the respective category.
Facilitating a critical analysis of the value of SVI over non-SVI counterparts, as one of the principal aims of this
study, the following indexes with different types of indicators are designed:
1. Index with SVI indicators and non-SVI indicators
2. Index with only SVI indicators
3. Index with only non-SVI indicators
These indexes and their sub-categories were compared with each other to examine how much SVI and non-SVI indi-
cators explain the overall variance. Finally, Figure 2 sums up the methodology.
First Author et al.: Preprint submitted to Elsevier Page 11 of 27
Figure 2: Illustration of the methodology of this study.
First Author et al.: Preprint submitted to Elsevier Page 12 of 27
Figure 3: Distribution of indicators describing connectivity across numerous locations in Singapore and Tokyo.
4. Results
4.1. Data collection
For OSM data, street network data and point data were retrieved. For street network data, 252,369 street segments
and 450,379 street segments were retrieved for Singapore and Tokyo, respectively. As for point data, 12,640 POIs
and 157 mass rail transit stations were collected from Singapore, 75,203 POIs, and 606 stations were collected from
Tokyo. For SVI, after removing indoor images and grey images from 7,142 images, 5,833 and 6,181 panorama images
remained for the two cities.
For LU, each city’s open data by the local government was used (Tokyo Metropolitan Government,2018;Singapore
Government,2020b). Because the categorizations of LU in each city were different, LU categories were harmonized
to residential, commercial, and industrial in this study, and other categories were excluded. AQI data in 2020 are
obtained for 5 and 126 stations in the city-state and the Japanese capital, respectively.
4.2. Extracted indicators and composite index
4.2.1. Connectivity
Connectivity is assessed based on the number of intersections with lights, intersections without lights, and cul-de-
sacs. Comparing Singapore and Tokyo, the latter achieves higher scores for connectivity (see Table 6 and Figure 8).
Due to a large number of indicators not all of them can be detailed in this paper, but as an example, Figure 3 illustrates
the distribution of values in this particular indicator. The score for the number of intersections without traffic lights is
much lower for Singapore while the other two indicators have a similar distribution.
4.2.2. Environment
The environment is evaluated based on slope, the number of POIs, land use mix, AQI, and pixel ratios of greenery,
buildings, and water, aiming to evaluate the natural and built environment. Compared to Singapore, Tokyo has a
higher mean score for this category (see Table 6 and Figure 8). The distributions of each indicator suggest that Tokyo
obtains noticeably higher scores for slope, land use, the pixel ratio of buildings. Such results reflect Tokyo’s relatively
First Author et al.: Preprint submitted to Elsevier Page 13 of 27
flat topography, organic distribution of land uses, and densely built buildings that create more enclosures for cyclists.
Singapore achieved higher scores in AQI and greenery, which also reflects its strategic environmental management as
a garden city (Han,2017;Palliwal et al.,2021).
4.2.3. Infrastructure
Infrastructure is evaluated based on the type of road and pavement, the width of the road, number of transit facil-
ities, and presence of potholes, street lights, bike lanes, street amenities, utility poles, bike parking spaces, sidewalks,
crosswalks, and curb cuts. This category aims to comprehensively evaluate various elements in the realm of infrastruc-
ture. Compared to Tokyo, Singapore achieves a much higher mean score for this category (see Table 6 and Figure 8).
Singapore obtains higher scores in the type of pavement (i.e. surface), the presence of street amenity, and, especially,
utility poles, which show the nearly opposite result from Tokyo. This result reflects Tokyo’s issue with many utility
poles, which not only deteriorates the beauty of the city but also can cause obstruction for pedestrians and cyclists
(Inajima and Urabe,2017).
4.2.4. Vehicle-cyclist interaction
Vehicle-cyclist interaction was evaluated based on the number of vehicles and speed control devices and the pres-
ence of on-street parking and traffic lights/stop signs. This category aims to assess how safely cyclists interact with
vehicles; thus, fewer and slower traffic leads to higher scores. The mean scores for both cities are also similar, while
Singapore has a slightly higher mean and standard deviation (see Table 6 and Figure 8). The analysis suggests very
small variances in the indicators and similar distributions for both cities. Such a result can potentially be explained:
Singapore and Tokyo have similar restrictions on vehicles; for example, on-street parking is strongly discouraged in
both cities (Barter,2010;Russo et al.,2019), and the numbers of vehicles per capita are similar: 0.22 and 0.17 (Barter,
2010;Singapore Government,2020a).
4.2.5. Perception
The perception is evaluated based on attractiveness for cycling, spaciousness, cleanliness, building design attrac-
tiveness, safety as cyclists, beauty, attractiveness for living. In this category, we will discuss the survey result, the
relationships between perception scores and high- and low-level features, the result of predictive modeling, and the
result of inferences from the models.
Surveys on 800 images, 400 images from each city, were conducted on Amazon Mechanical Turk to recruit eight
unique participants to rate each image’s perception scores mentioned above on a scale of 0 to 10. Before proceed-
ing further, it is important to assert that while our work strives for a high degree of automation, as in most studies
involving machine learning, a portion of the work is manual, i.e. the labeling of training data was done through crowd-
sourcing (Section 3.3). However, after the predictive modeling using the data obtained from the survey, the process of
determining the scores was automated.
The results of the survey indicate strong positive correlations among the perception scores, ranging from 0.58 to
0.79 in squared R, and exhibit no visible skewness in the data distribution (see Figure 4 and Figure 5). Figure 6 indicates
the results of the survey after excluding outliers, which reveal similar results for all the scores and illustrates that the
majority of responses are between three and eight. The figure also suggests that images from Singapore generally had
higher scores across all the measures.
Visual features from SVI are extracted through high-level feature extraction (i.e. semantic segmentation, classifi-
cation, and object detection) and low-level feature extraction (i.e. edge detection, blob detection, and HLS statistics).
After labeling each observation below or above the mean value of each feature, Welch’s t-test is conducted to find
features with statistically significant effects on perception scores based on the labeling. The total number of extracted
features is 519, and the number of scores is seven; therefore, there are 3,633 unique feature-score pairs.
For beauty, we found that features that contribute to better infrastructures such as curb and terrain contribute to
higher scores, while features that damage beauty such as utility poles and junkyards contribute to lower scores. Also,
for low-level features, larger standard deviations of HLS are associated with lower beauty scores.
As for building attractiveness, we realize that a higher pixel ratio of buildings and classification as slums leads to
lower scores, implying that participants did not prefer dense buildings. On the other hand, classification as residen-
tial neighborhood and campus leads to higher scores, suggesting residential and education as potential land uses and
typologies that make people feel attracted to buildings.
The result of cleanliness suggests that classification as campus and higher pixel ratio of curb leads to higher clean-
First Author et al.: Preprint submitted to Elsevier Page 14 of 27
Figure 4: A scatter plot matrix of the perception scores
obtained from the survey.
Figure 5: A correlation matrix of the perception scores
obtained from the survey.
liness scores, which might imply that these places and infrastructures are relatively better maintained. On the other
hand, classification as junkyard and landfill led to lower cleanliness scores, which is intuitively understandable because
of their strong associations with garbage. As for cycling attractiveness, natural elements such as terrain and vegetation
as well as classification as residential neighborhoods leads to higher cycling attractiveness scores, while utility poles
and construction sites are associated with lower scores. Attractiveness for living obtains a similar pattern to other
scores. A higher pixel ratio of terrain and curb entails higher scores, and classification as slum and landfill causes
lower scores. The result of safety shows that terrain and campus achieve higher scores and that other features such as
gas stations, junkyards, and slums were associated with lower safety scores. Lastly, spaciousness gains higher scores
when terrain, curb, and campus are present, while other features, such as utility poles, junkyards, and larger standard
deviations of HLS statistics are detrimental.
In this exploratory analysis, some parts of the study by Verma et al. (2020) are replicated such as low associations
among perception scores and edge detection and blob detection. However, other parts of the study such as the strong
influence of cars in object detection could not be reproduced, suggesting that the geography of the study and/or training
data may play a role. Moreover, more features from image classification are found to have an influence on perception
scores than other features.
Based on the study by Verma et al. (2020), predictive modeling is conducted for each perception score between 0
and 10 with high- and low-level features (see Table 4). Table 5 underscores the results of the modeling with different
performance metrics, where one can see low mean absolute error (MAE) around 0.65, mean absolute percentage error
(MAPE) around 0.1, and root mean squared error (RMSE) around 0.8; however, all the 𝑅2values were below 0. This
means that the models were worse than predicting the constant values regardless of input data.
Although Verma et al. (2020) achieves 𝑅2as high as 0.66 by using the same visual features, our study could not
achieve comparable results. This discrepancy might be due to the disparate level of data quality in the SVI training set.
After the modeling, we randomly selected SVI images with higher and lower scores and found that some partially grey
images and indoor images were still in the set despite the efforts to clean the data prior to the modeling. Such noise in
SVI data is not often discussed in previous studies using SVI, and this issue remains to be solved in the future (Biljecki
and Ito,2021). To further improve the prediction, other visual feature extractions and selection methods need to be
explored as well.
4.2.6. Bikeability
After incorporating scores across all categories, bikeability scores were calculated at a fine spatial scale (Figure 7).
In Singapore, bikeability scores are generally distributed homogeneously barring a few outliers. The results for Tokyo
First Author et al.: Preprint submitted to Elsevier Page 15 of 27
Figure 6: Distribution of seven perception scores obtained from the survey by each city.
Table 5
Results of predictive modeling of perception indicators.
Target_variable MAE MAPE RMSE R2
beauty 0.63 0.10 0.81 -0.20
building_attractiveness 0.63 0.10 0.78 -0.14
cleanliness 0.63 0.10 0.79 -0.12
cycling_attractiveness 0.65 0.11 0.83 -0.18
living_attractiveness 0.69 0.12 0.87 -0.10
safety 0.72 0.12 0.89 -0.29
spaciousness 0.67 0.11 0.83 -0.19
seem to be more heterogeneous, with a lower score in the central area and peripheral areas and higher scores in between.
The distribution of data indicates no skewness in the data, and a comparison between Singapore and Tokyo in Table 6
hints that Singapore achieved a slightly higher mean value and lower standard deviation than Tokyo.
This bikeability index, however, has some issues that need to be exposed. Firstly, the perception prediction’s result
is not entirely reliable. Although MAE, MAPE, and RMSE of models turned out to be moderately acceptable, 𝑅2
values resulted in negative values, indicating the models are more inaccurate than simply predicting the mean values
First Author et al.: Preprint submitted to Elsevier Page 16 of 27
Figure 7: One of the key outputs of this study: maps of bikeability across Singapore and Tokyo, generated from the scores
at numerous locations in the two cities.
of the scores. Low variances among observations in some categories also need to be considered. Summary statistics
shown in Table 6 and data distribution shown in Figure 8 suggest some categories such as vehicle-cyclist interaction
had low standard deviation and skewed data distribution with a few outliers. This poses a question regarding the
appropriate indicators to be used that can differentiate bikeability scores within and between cities because most of the
sample points obtain highly similar scores for such categories. This question needs to be examined by expanding study
areas in different contexts to observe whether this phenomenon is simply caused by similar characteristics of streets in
the Asian context.
4.3. Comparison between SVI and Non-SVI indicators
This study compares three types of indexes to answer one of the research questions gauging the value of SVI in
bikeability studies. Besides the comprehensive index developed with both SVI and non-SVI indicators, indexes with
only SVI and non-SVI indicators are developed and compared with each other. Each category’s score was recalculated
based on the indicator types, and, for bikeability, we compare the indexes by plotting scores developed with different
indicators on the same sample points. Connectivity and perception were excluded from this analysis because these
categories only had one type of indicator, only non-SVI indicators, and only SVI indicators, respectively.
Indexes of the environment developed with SVI and non-SVI indicators are not correlated with each other and
that non-SVI indicators are more strongly correlated with the index developed with both of the indicators. On the
other hand, SVI indicators had more influence on infrastructure and vehicle-cyclist interaction. After combining all
the categories, bikeability scores were compared among different indicator types, and SVI indicators turned out to
have a stronger correlation with the index with both indicator types and lower kurtosis than non-SVI indicators (see
Figure 10 and Figure 9). Although this result is not a surprise given the larger number of SVI indicators than non-SVI
indicators, this result indicates the potential of SVI indicators to be used alone to evaluate bikeability because of its
high correlation, 𝑅2of 0.85, with the index with both indicators and capability to explain most of the variance.
Although it was found possible to estimate bikeability only with SVI indicators, some relevant non-SVI indicators
First Author et al.: Preprint submitted to Elsevier Page 17 of 27
Figure 8: Distribution of scores across categories in Singapore and Tokyo.
are either difficult or impossible to replace with SVI. For example, transit facilities, POIs, and land uses are not impossi-
ble but complicated to obtain from SVI and are much more straightforward to collect from OSM and other data sources,
which are anyway frequently available. Also, the slope and air quality are unnecessarily challenging to estimate from
SVI, and other data sources are much less involved to collect and much more accurate than it would be if estimating
them from SVI. Therefore, with these outcomes in mind, it is more beneficial for urban planners and researchers to
combine both SVI and non-SVI indicators to assess bikeability, complementing the best of the two worlds.
5. Challenges and future directions
5.1. Data quality
This study faced several challenges. One of such challenges was the quality of data. This issue surfaced in data
sources, such as SVI, OSM, and AQI. GSV was used to collect SVI in this study, and the sample points of collected
GSV panorama images were randomly selected to cover the entire study areas with a limited number of sample points
of GSV that can be obtained for free. A study by Kim et al. (2021) reported that larger sampling intervals lead to larger
variances of elements that can be extracted from SVI. Therefore, one of the limitations of this study is possible biases
and larger variances of data introduced by random sampling intervals when conducting random sampling. Future
studies can define shorter sampling intervals to reduce the possible bias. Another possible bias is the perspective of
SVI. Because SVI is collected from vehicles, it might not always represent the typical view of bicyclists, especially for
wide roads. This limitation is also faced by other previous studies on walkability that as well use 360-degree panoramas
for assessment, as they aim to gauge what pedestrians perceive (Wakamiya et al.,2019;Yencha,2019;Nagata et al.,
First Author et al.: Preprint submitted to Elsevier Page 18 of 27
Table 6
Summary statistics of the bikeability measures calculated for the study area.
Category City Mean Standard deviation
Bikeability Singapore 54.26 3.66
Tokyo 53.98 3.73
Connectivity Singapore 16.63 1.59
Tokyo 18.72 1.70
Environment Singapore 5.83 0.94
Tokyo 6.58 0.94
Infrastructure Singapore 9.88 1.90
Tokyo 7.86 2.36
Perception Singapore 11.42 1.57
Tokyo 10.40 1.74
Vehicle–cyclist interaction Singapore 10.50 2.12
Tokyo 10.42 1.61
Figure 9: A scatter plot matrix of bikeability score
with only SVI and non-SVI indicators.
Figure 10: A correlation matrix of bikeability score
with only SVI and non-SVI indicators.
2020;Villeneuve et al.,2018;Li et al.,2018;Zhang et al.,2018;Weld et al.,2019). In this study, the result showed
that cycling paths are not highly prevalent in both Singapore and Tokyo; thus, the panorama views possibly present
most of the actual cycling perspectives. Although SVI on cycling paths is not widely available currently, future studies
might be able to work on SVI on cycling paths with the rise of crowdsourced SVI services, such as Mapillary.
As for OSM data, while in general of high quality for both locations, its data completeness became a bottleneck
when evaluating some particular indicators (e.g. traffic speed and the number of traffic lanes) and resulted in their
eliminations, so future studies should explore different data sources to collect them.
AQI data are available in both cities, but the number of AQI stations in Singapore, five, was far fewer than Tokyo,
First Author et al.: Preprint submitted to Elsevier Page 19 of 27
126, and such low spatial granularity of data might have affected the result. Therefore, future studies can incorporate
predictive modeling to estimate AQI for each sample point by using traffic data, meteorological data, and land use data
(Chen et al.,2010).
Low variances among sample points for some indicators were found as well. This is potentially problematic because
indicators with extremely low variances cannot differentiate sample points; thus, this needs to be examined more by
investigating more diverse sets of cities.
Perception modeling was another challenge faced in this study due to the nature of the approach. The compre-
hensive survey was conducted on a service on which the majority of participants are not residents of the study areas,
which may be advantageous but also detrimental. On the one hand, survey results might have been different if they
were conducted with residents in the study areas (Difallah et al.,2018). To mitigate the potential bias, future studies
can utilize different survey services that can specify the residence of the participants. On the other hand, having par-
ticipants of a cross-city study residing in neither study area may mitigate the bias of residents living in one city but not
being familiar with the other one.
5.2. Required skills
Because the data collection and processing was conducted using Python, this method requires users to have a
moderately advanced understanding of programming. Computational power is also a challenge because this study’s
method uses CV techniques that require graphics processing capabilities, which are not available widely. To mitigate
these limitations, future studies may consider developing a GUI software and an API to allow users to input data and
assess bikeability.
5.3. Development of the index
This study selected the indicators based on the systematic review. However, purposes of cycling and socio-
economic characteristics might affect the indicator selection and weight assignment, as examined by Arellana et al.
(2020). Moreover, a bikeability assessment based on route simulations between origin and destinations called capability-
wise walkability score has been proposed by a previous study (Blečić et al.,2021). Thus, the incorporation of such
new designs of bikeability index might be able to enhance our method as well.
6. Conclusion
In previous studies, it has been a challenge to make bikeability assessment scalable while keeping a good balance of
objectivity and subjectivity, and minimizing the work required. The few studies that have used SVI and CV techniques
to automate the process and increase the geographic coverage rather assessed limited aspects of bikeability and have
not done so critically, nor have they investigated the inclusion of multiple cities.
We advanced the comprehensive assessment of bikeability using street view imagery and computer vision. The
contributions of our study are the creation of an exhaustive bikeability index inspired by previous studies with SVI indi-
cators extracted with CV techniques as the major data source, novel exploration of automatable subjective assessment
of bikeability, comparison of SVI and non-SVI indicators for the first time, and a novel investigation of the potential
of using SVI indicators independently for bikeability assessment.
Non-perception SVI indicators were extracted using semantic segmentation and object detection, and perception
SVI indicators were predicted by training models with survey results as target variables and visual features extracted
from SVI as independent variables. Bikeability indexes that range from 0 to 100 were developed in Singapore and
Tokyo and compared with each other, which resulted in slightly higher bikeability scores in Singapore, 54.26 on
average, than in Tokyo, 53.98 on average. A thorough comparison between SVI and non-SVI indicators was made
to examine which has more influence when predicting conditions and the appeal of cycling. SVI indicators turned
out to have a much stronger correlation with the estimated bikeability index, an 𝑅2of 0.85, outperforming non-SVI
indicators, which had an 𝑅2of 0.4. However, the usefulness of the latter should not be discounted.
In summary, the takeaways of this research are:
This study has demonstrated that we can use CV techniques and SVI to comprehensively assess bikeability
within and among cities. The paper details the design and implementation of a bikeability index that relies on
SVI and CV, which is calculated at both a fine spatial scale and aggregated at the level of a city. The study asserts
that this index at least supplements traditional instruments used in this research domain.
First Author et al.: Preprint submitted to Elsevier Page 20 of 27
Indicators that are derived from SVI dwarf those that are computed from non-SVI counterparts. A large portion
of the variance in the overall index was explained by SVI indicators, overshadowing those prominent in orthodox
mechanisms used hitherto.
SVI may potentially be used on its own to assess bikeability in the built environment. However, as elaborated
in the previous sections, this study faced several challenges that, despite the convincing usability of SVI, may
not make taking this independent route always viable. Thus, future studies need to ameliorate several practical
issues. Further, the comparative advantage comes at a price — it is more difficult to obtain SVI indicators in
comparison to the non-SVI counterparts. Therefore, despite the relative usability and independence of either, it
might be beneficial to use both to assess bikeability, complementing their pros and cons.
Based on the findings from this study, future research should focus not only on overcoming the challenges discussed
in the previous section but also on further enhancing the index, which advances the state of the art but may nevertheless
benefit from further work. The index can be improved by expanding the scope of SVI indicators to be included and
also improving the indicator weights according to cyclists’ preferences. The enlargement of the indicators can be done
by building CV models to extract more indicators, such as road type, land use, and type of pavement. Some studies
have explored creating weights based on cyclists’ preferences through surveys, creating different weights for different
types of cyclists (Arellana et al.,2020). Such improvement of the weights can benefit better decision-making in urban
planning according to the demographics in target areas, thus, future studies should also incorporate such methods.
Another direction for future work would be coupling the developed index with instances introduced to assess other
urban aspects such as walkability and livability (Zhao et al.,2020;Benita et al.,2020), to investigate relationships or
complement them.
List of acronyms
AQI Air Quality Index
CV Computer Vision
DEM Digital Elevation Model
HSL Hue-Saturation-Lightness
IC Image Classification
LU Land Use
MAE Mean Absolute Error
MAPE Mean Absolute Percentage Error
GSV Google Street View
OD Object Detection
OSM OpenStreetMap
POI Point of Interest
RMSE Root Mean Squared Error
SS Semantic Segmentation
SVI Street View Imagery
V–C Interaction Vehicle–Cyclist Interaction
We appreciate the comments by Jeffrey Ho and the design contribution by April Zhu (National University of
Singapore). We thank the members of the NUS Urban Analytics Lab for the discussions and the reviewers for their
suggestions. The data sources used in this study are gratefully acknowledged. This research is part of the project Large-
scale 3D Geospatial Data for Urban Analytics, which is supported by the National University of Singapore under the
Start Up Grant R-295-000-171-133. This study has received an exemption from the Institutional Review Board (IRB)
of the National University of Singapore under the reference code NUS-IRB-2021-29.
Abadi, M.G., Hurwitz, D.S., 2018. Bicyclist’s perceived level of comfort in dense urban environments: How do ambient traffic, engineering
treatments, and bicyclist characteristics relate? Sustainable Cities and Society 40, 101–109. doi:10.1016/j.scs.2018.04.003.
First Author et al.: Preprint submitted to Elsevier Page 21 of 27
Alaoui, E.A.A., Tekouabou, S.C.K., 2021. Intelligent management of bike sharing in smart cities using machine learning and internet of things.
Sustainable Cities and Society 67, 102702. doi:10.1016/j.scs.2020.102702.
Aldred, R., García-Herrero, S., Anaya, E., Herrera, S., Mariscal, M., 2020. Cyclist injury severity in spain: A bayesian analysis of police road injury
data focusing on involved vehicles and route environment. International Journal of Environmental Research and Public Health 17. doi:10.3390/
Arellana, J., Saltarín, M., Larrañaga, A.M., González, V.I., Henao, C.A., 2020. Developing an urban bikeability index for different types of cyclists
as a tool to prioritise bicycle infrastructure investments. Transportation Research Part A: Policy and Practice 139, 310–334. doi:10.1016/j.
Attard, M., Cañas, C., Maas, S., 2021. Determinants for walking and cycling to a university campus: Insights from a participatory Active Travel
workshop in Malta. Transportation Research Procedia 52, 501–508. doi:10.1016/j.trpro.2021.01.059.
Barrington-Leigh, C., Millard-Ball, A., 2017. The world’s user-generated road map is more than 80% complete. PLOS ONE 12, e0180698 – 20.
Barter, P.A., 2010. Parking Policy in Asian Cities. SSRN Electronic Journal URL:, doi:10.2139/
Bauman, A.E., Reis, R.S., Sallis, J.F., Wells, J.C., Loos, R.J., Martin, B.W., 2012. Correlates of physical activity: why are some people physically
active and others not? The Lancet 380, 258–271. doi:10.1016/s0140- 6736(12)60735-1.
Benita, F., Kalashnikov, V., Tunçer, B., 2020. A Spatial Livability Index for dense urban centers. Environment and Planning B: Urban Analytics
and City Science , 239980832096015doi:10.1177/2399808320960151.
Berger, M., Dörrzapf, L., 2018. Sensing comfort in bicycling in addition to travel data, pp. 524–534. doi:10.1016/j.trpro.2018.10.034.
Biljecki, F., 2020. Exploration of open data in Southeast Asia to generate 3D building models. ISPRS Annals of Photogrammetry, Remote Sensing
and Spatial Information Sciences VI-4/W1-2020, 37–44. doi:10.5194/isprs-annals- vi-4- w1-2020-37- 2020.
Biljecki, F., Ito, K., 2021. Street view imagery in urban analytics and GIS: A review. Landscape and Urban Planning 215, 104217. doi:10.1016/
Blečić, I., Cecchini, A., Congiu, T., Fancello, G., Talu, V., Trunfio, G.A., 2021. Capability-wise walkability evaluation as an indicator of ur-
ban peripherality. Environment and Planning B: Urban Analytics and City Science 48, 895–911. URL:
2399808320908294, doi:10.1177/2399808320908294. publisher: SAGE Publications Ltd STM.
Boeing, G., 2017. OSMnx: New methods for acquiring, constructing, analyzing, and visualizing complex street networks. Computers, Environment
and Urban Systems 65, 126–139. doi:10.1016/j.compenvurbsys.2017.05.004.
Boongaling, C.G.K., Luna, D.A., Samantela, S.S., 2021. Developing a street level walkability index in the Philippines using 3D pho-
togrammetry modeling from drone surveys. GeoJournal URL:, doi:10.1007/
s10708-021- 10441-2.
Brüchert, T., Hasselder, P., Quentin, P., Bolte, G., 2020. Walking for transport among older adults: A cross-sectional study on the role of the built
environment in less densely populated areas in northern germany. International Journal of Environmental Research and Public Health 17, 1–22.
Bulò, S.R., Porzi, L., Kontschieder, P., 2018. In-Place Activated BatchNorm for Memory-Optimized Training of DNNs. arXiv:1712.02616 URL:
Cain, K.L., Geremia, C.M., Conway, T.L., Frank, L.D., Chapman, J.E., Fox, E.H., Timperio, A., Veitch, J., Van Dyck, D., Verhoeven, H., Reis,
R., Augusto, A., Cerin, E., Mellecker, R.R., Queralt, A., Molina-García, J., Sallis, J.F., 2018. Development and reliability of a streetscape
observation instrument for international use: MAPS-global. The International Journal of Behavioral Nutrition and Physical Activity 15, 19.
doi:10.1186/s12966-018- 0650-z.
Cao, Y., Shen, D., 2019. Contribution of shared bikes to carbon dioxide emission reduction and the economy in Beijing. Sustainable Cities and
Society 51, 101749. doi:10.1016/j.scs.2019.101749.
Castañon, U.N., Ribeiro, P.J.G., 2021. Bikeability and Emerging Phenomena in Cycling: Exploratory Analysis and Review. Sustainability 13,
2394. doi:10.3390/su13042394.
Chacra, D.A., Zelek, J., 2018. Municipal Infrastructure Anomaly and Defect Detection, in: 26th European Signal Processing Conference (EU-
SIPCO), pp. 2125–2129. doi:10.23919/EUSIPCO.2018.8553322.
Chen, J., Stouffs, R., Biljecki, F., 2021a. Hierarchical (Multi-Label) Architectural Image Recognition and Classification, in: Proceedings of the 26th
International Conference of the Association for Computer-Aided Architectural Design Research in Asia (CAADRIA) 2021, pp. 161–170.
Chen, L., Bai, Z., Kong, S., Han, B., You, Y., Ding, X., Du, S., Liu, A., 2010. A land use regression for predicting NO2 and PM10 concen-
trations in different seasons in Tianjin region, China. Journal of Environmental Sciences 22, 1364–1373. doi:
S1001-0742(09)60263- 1.
Chen, L., Lu, Y., Sheng, Q., Ye, Y., Wang, R., Liu, Y., 2020a. Estimating pedestrian volume using Street View images: A large-scale validation
test. Computers, Environment and Urban Systems 81, 101481. doi:10.1016/j.compenvurbsys.2020.101481.
Chen, W., Liu, Q., Zhang, C., Mi, Z., Zhu, D., Liu, G., 2020b. Characterizing the stocks, flows, and carbon impact of dockless sharing bikes in
china. Resources, Conservation and Recycling 162. doi:10.1016/j.resconrec.2020.105038.
Chen, W., Wu, A.N., Biljecki, F., 2021b. Classification of urban morphology with deep learning: Application on urban vitality. Computers,
Environment and Urban Systems 90, 101706. doi:10.1016/j.compenvurbsys.2021.101706.
Chevalier, A., Xu, L., 2020. On the applicability of a western bikeability index in the chinese context. International Review for Spatial Planning
and Sustainable Development 8, 59–93. doi:10.14246/IRSPSD.8.1_59.
Cicchino, J.B., McCarthy, M.L., Newgard, C.D., Wall, S.P., DiMaggio, C.J., Kulie, P.E., Arnold, B.N., Zuby, D.S., 2020. Not all protected bike
lanes are the same: Infrastructure and risk of cyclist collisions and falls leading to emergency department visits in three U.S. cities. Accident
Analysis & Prevention 141, 105490. doi:10.1016/j.aap.2020.105490.
Clifton, K.J., Livi Smith, A.D., Rodriguez, D., 2007. The development and testing of an audit for the pedestrian environment. Landscape and Urban
First Author et al.: Preprint submitted to Elsevier Page 22 of 27
Planning 80, 95–110. doi:10.1016/j.landurbplan.2006.06.008.
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 Understanding. arXiv:1604.01685 URL:
Daraei, S., Pelechrinis, K., Quercia, D., 2021. A data-driven approach for assessing biking safety in cities. EPJ Data Science 10, 1–16. doi:10.
1140/epjds/s13688-021- 00265-y.
Delacre, M., Lakens, D., Leys, C., 2017. Why Psychologists Should by Default Use Welch’s t-test Instead of Student’s t-test. International Review
of Social Psychology 30, 92–101. doi:10.5334/irsp.82.
Deng, L., Pan, J., Xu, X., Yang, W., Liu, C., Liu, H., 2018. PDRLGB: Precise DNA-binding residue prediction using a light gradient boosting
machine. BMC Bioinformatics 19. doi:10.1186/s12859-018-2527-1.
Difallah, D., Filatova, E., Ipeirotis, P., 2018. Demographics and Dynamics of Mechanical Turk Workers, in: Proceedings of the Eleventh ACM
International Conference on Web Search and Data Mining, ACM, Marina Del Rey CA USA. pp. 135–143. doi:10.1145/3159652.3159661.
Ding, X., Fan, H., Gong, J., 2021. Towards generating network of bikeways from Mapillary data. Computers, Environment and Urban Systems 88,
101632. doi:10.1016/j.compenvurbsys.2021.101632.
Doubleday, A., Choe, Y., Isaksen, T., Miles, S., Errett, N., 2021. How did outdoor biking and walking change during COVID-19?: A case study of
three U.S. cities. PLoS ONE 16. doi:10.1371/journal.pone.0245514.
Du, Y., Deng, F., Liao, F., 2019. A model framework for discovering the spatio-temporal usage patterns of public free-floating bike-sharing system.
Transportation Research Part C: Emerging Technologies 103, 39–55. doi:10.1016/j.trc.2019.04.006.
Dubey, A., Naik, N., Parikh, D., Raskar, R., Hidalgo, C.A., 2016. Deep Learning the City: Quantifying Urban Perception at a Global Scale, in:
Leibe, B., Matas, J., Sebe, N., Welling, M. (Eds.), Computer Vision – ECCV 2016, Springer International Publishing, Cham. pp. 196–212.
doi:10.1007/978-3- 319-46448- 0_12.
Faghih Imani, A., Miller, E.J., Saxe, S., 2019. Cycle accessibility and level of traffic stress: A case study of Toronto. Journal of Transport Geogra-
phy 80, 102496. URL:, doi:10.1016/j.jtrangeo.
Fan, H., Kong, G., Zhang, C., 2021. An Interactive platform for low-cost 3D building modeling from VGI data using convolutional neural network.
Big Earth Data 5, 49–65. doi:10.1080/20964471.2021.1886391.
Feng, G., Zou, G., Piga, B.E.A., Hu, H., 2021. The Validity of Street View Service Applied to Ambiance Perception of Street: A Comparison of
Assessment in Real Site and Baidu Street View, in: Shin, C.S., Di Bucchianico, G., Fukuda, S., Ghim, Y.G., Montagna, G., Carvalho, C. (Eds.),
Advances in Industrial Design, Springer International Publishing, Cham. pp. 740–748. doi:10.1007/978-3- 030-80829-7_91.
Frank, L.D., Schmid, T.L., Sallis, J.F., Chapman, J., Saelens, B.E., 2005. Linking objectively measured physical activity with objectively measured
urban form: findings from SMARTRAQ. American Journal of Preventive Medicine 28, 117–125. doi:10.1016/j.amepre.2004.11.001.
Galanis, A., Papanikolaou, A., Eliou, N., 2018. Bikeability audit in urban road environment: Case study in the city of Volos, Greece. doi:10.4018/
978-1- 5225-5210- 9.ch021.
Gholamialam, A., Matisziw, T., 2019. Modeling bikeability of urban systems. Geographical Analysis 51, 73–89. doi:10.1111/gean.12159.
Goel, R., Garcia, L.M.T., Goodman, A., Johnson, R., Aldred, R., Murugesan, M., Brage, S., Bhalla, K., Woodcock, J., 2018. Estimating city-level
travel patterns using street imagery: A case study of using Google Street View in Britain. PLOS ONE 13, e0196521. doi:10.1371/journal.
Gong, F.Y., Zeng, Z.C., Zhang, F., Li, X., Ng, E., Norford, L.K., 2018. Mapping sky, tree, and building view factors of street canyons in a
high-density urban environment. Building and Environment 134, 155–167. doi:10.1016/j.buildenv.2018.02.042.
Gong, Z., Ma, Q., Kan, C., Qi, Q., 2019. Classifying Street Spaces with Street View Images for a Spatial Indicator of Urban Functions. Sustainability
11, 6424. doi:10.3390/su11226424.
Google, 2018. "Street View ready (pro grade)" specifications. URL:
Grigore, E., Garrick, N., Fuhrer, R., Axhausen, I., 2019. Bikeability in Basel. Transportation Research Record 2673, 607–617. doi:10.1177/
Gu, P., Han, Z., Cao, Z., Chen, Y., Jiang, Y., 2018. Using Open Source Data to Measure Street Walkability and Bikeability in China: A Case of
Four Cities. Transportation Research Record 2672, 63–75. doi:10.1177/0361198118758652.
Guler, D., Yomralioglu, T., 2021. Location Evaluation of Bicycle Sharing System Stations and Cycling Infrastructures with Best Worst Method
Using GIS. The Professional Geographer , 1–18doi:10.1080/00330124.2021.1883446.
Gullón, P., Badland, H.M., Alfayate, S., Bilal, U., Escobar, F., Cebrecos, A., Diez, J., Franco, M., 2015. Assessing Walking and Cycling Envi-
ronments in the Streets of Madrid: Comparing On-Field and Virtual Audits. Journal of Urban Health: Bulletin of the New York Academy of
Medicine 92, 923–939. doi:10.1007/s11524-015- 9982-z.
Guo, J., He, H., He, T., Lausen, L., Li, M., Lin, H., Shi, X., Wang, C., Xie, J., Zha, S., Zhang, A., Zhang, H., Zhang, Z., Zhang, Z., Zheng,
S., Zhu, Y., 2020. GluonCV and GluonNLP: Deep Learning in Computer Vision and Natural Language Processing. arXiv:1907.04433 URL:
Hall, P., Lu, Y., Lu, J., Zhang, S., 2018. Traffic signal detection and classification in street views using an attention model. Computational Visual
Media 4, 253–266. doi:10.1007/s41095-018- 0116-x.
Han, H., 2017. Singapore, a Garden City: Authoritarian Environmentalism in a Developmental State. The Journal of Environment & Development
26, 3–24. doi:10.1177/1070496516677365.
Hartanto, K., Grigolon, A.B., Maarseveen, M., Brussel, M., 2017. Developing a bikeability index in the context of transit-oriented development
(TOD)., in: 15th International Conference on Computers in Urban Planning and Urban Management (CUPUM), Adelaide, Australia.
Hoedl, S., Titze, S., Oja, P., 2010. The bikeability and walkability evaluation table reliability and application. American Journal of Preventive
Medicine 39, 457–459. doi:10.1016/j.amepre.2010.07.005.
Hollander, J.B., Nikolaishvili, G., Adu-Bredu, A.A., Situ, M., Bista, S., 2020. Using deep learning to examine the correlation between
First Author et al.: Preprint submitted to Elsevier Page 23 of 27
transportation planning and perceived safety of the built environment. Environment and Planning B: Urban Analytics and City Science ,
Horacek, T., Dede Yildirim, E., Kattelmann, K., Brown, O., Byrd-Bredbenner, C., Colby, S., Greene, G., Hoerr, S., Kidd, T., Koenings, M., Morrell,
J., Olfert, M., Phillips, B., Shelnutt, K., White, A., 2018. Path analysis of campus walkability/bikeability and college students’ physical activity
attitudes, behaviors, and body mass index. American Journal of Health Promotion 32, 578–586. doi:10.1177/0890117116666357.
Horacek, T.M., White, A.A., Greene, G.W., Reznar, M.M., Quick, V.M., Morrell, J.S., Colby, S.M., Kattelmann, K.K., Herrick, M.S., Shelnutt,
K.P., Mathews, A., Phillips, B.W., Byrd-Bredbenner, C., 2012. Sneakers and spokes: An assessment of the walkability and bikeability of U.S.
postsecondary institutions. Journal of Environmental Health 74, 8–15.
Inajima, T., Urabe, E., 2017. Koike’s plan for Tepco to remove utility poles in Tokyo an Olympian task. URL:
jp/news/2017/04/03/business/koikes-plan- tepco-remove- utility-poles-tokyo-olympian-task/.
Kalvelage, K., Kalvelage, K., Dorneich, M.C., Dorneich, M.C., Seeger, C.J., Seeger, C.J., Welk, G.J., Welk, G.J., Gilbert, S., Gilbert, S., Moon, J.,
Moon, J., Jafir, I., Jafir, I., Brown, P., Brown, P., 2018. Assessing the validity of facilitated-volunteered geographic information: comparisons of
expert and novice ratings. GeoJournal 83, 477–488. doi:10.1007/s10708-017- 9781-z.
Kamel, M., Sayed, T., Bigazzi, A., 2020. A composite zonal index for biking attractiveness and safety. Accident Analysis and Prevention 137.
Kang, H., Kim, D., Yoo, S., 2019. Attributes of perceived bikeability in a compact urban neighborhood based on qualitative multi-methods.
International Journal of Environmental Research and Public Health 16. doi:10.3390/ijerph16193738.
Kang, Y., Zhang, F., Gao, S., Peng, W., Ratti, C., 2021. Human settlement value assessment from a place perspective: Considering human
dynamics and perceptions in house price modeling. Cities 118, 103333. URL:
pii/S026427512100233X, doi:10.1016/j.cities.2021.103333.
Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., Liu, T.Y., 2017. LightGBM: a highly efficient gradient boosting decision tree,
in: Proceedings of the 31st International Conference on Neural Information Processing Systems, Curran Associates Inc., Red Hook, NY, USA.
pp. 3149–3157.
Kellstedt, D.K., Spengler, J.O., Foster, M., Lee, C., Maddock, J.E., 2021. A Scoping Review of Bikeability Assessment Methods. Journal of
Community Health 46, 211–224. doi:10.1007/s10900-020- 00846-4.
Kim, J.H., Lee, S., Hipp, J.R., Ki, D., 2021. Decoding urban landscapes: Google street view and measurement sensitivity. Computers, Environment
and Urban Systems 88, 101626. doi:10.1016/j.compenvurbsys.2021.101626.
Koh, P.P., Wong, Y., 2013. Influence of infrastructural compatibility factors on walking and cycling route choices. Journal of Environmental
Psychology 36, 202–213. doi:10.1016/j.jenvp.2013.08.001.
Kraus, S., Koch, N., 2021. Provisional COVID-19 infrastructure induces large, rapid increases in cycling. Proceedings of the National Academy of
Sciences 118. doi:10.1073/pnas.2024399118.
Krenn, P.J., Oja, P., Titze, S., 2015. Development of a Bikeability Index to Assess the Bicycle-Friendliness of Urban Environments. Open Journal
of Civil Engineering 05, 451. doi:10.4236/ojce.2015.54045.
Labetski, A., Chum, A., 2020. Built environmental correlates of cycling accidents involving fatalities and serious injuries in london, uk. Frontiers
in Sustainable Cities 2, 59. doi:10.3389/frsc.2020.599635.
Li, M., Zhang, Z., Lei, L., Wang, X., Guo, X., 2020. Agricultural Greenhouses Detection in High-Resolution Satellite Images Based on Convolu-
tional Neural Networks: Comparison of Faster R-CNN, YOLO v3 and SSD. Sensors 20, 4938. URL: 8220/
20/17/4938, doi:10.3390/s20174938.
Li, X., Ratti, C., 2019. Using Google Street View for Street-Level Urban Form Analysis, a Case Study in Cambridge, Massachusetts, in: D’Acci,
L. (Ed.), The Mathematics of Urban Morphology. Springer International Publishing, pp. 457–470. doi:10.1007/978- 3-030- 12381-9_20.
Li, X., Santi, P., Courtney, T.K., Verma, S.K., Ratti, C., 2018. Investigating the association between streetscapes and human walking activities using
Google Street View and human trajectory data. Transactions in GIS 22, 1029–1044. URL:
10.1111/tgis.12472, doi:10.1111/tgis.12472. _eprint:
Li, X.j., Ratti, C., Seiferling, I., 2017. Quantifying the shade provision of street trees in urban landscape: A case study in Boston, USA, using
Google Street View. Landscape and Planning 169. doi:10.1016/j.landurbplan.2017.08.011.
Lin, J.J., Wei, Y.H., 2018. Assessing area-wide bikeability: A grey analytic network process. Transportation Research Part A: Policy and Practice
113, 381–396. doi:10.1016/j.tra.2018.04.022.
Long, Y., Zhao, J., 2020. What makes a city bikeable? a study of intercity and intracity patterns of bicycle ridership using mobike big data records.
Built Environment 46, 55–75. doi:10.2148/benv.46.1.55.
Lowry, M.B., Furth, P., Hadden-Loh, T., 2016. Low-Stress Neighborhood Bikeability Assessment to Prioritize Bicycle Infrastructure. URL: number: 16-1115.
Lu, Y., 2019. Using Google Street View to investigate the association between street greenery and physical activity. Landscape and Urban Planning
191, 103435. doi:10.1016/j.landurbplan.2018.08.029.
Lu, Y., Yang, Y., Sun, G., Gou, Z., 2019. Associations between overhead-view and eye-level urban greenness and cycling behaviors. Cities 88,
10–18. doi:10.1016/j.cities.2019.01.003.
Luo, H., Zhao, F., Chen, W.Q., Cai, H., 2020. Optimizing bike sharing systems from the life cycle greenhouse gas emissions perspective. Trans-
portation Research Part C: Emerging Technologies 117, 102705. doi:10.1016/j.trc.2020.102705.
Ma, L., Dill, J., 2017. Do people’s perceptions of neighborhood bikeability match "reality"? Journal of Transport and Land Use 10, 291–308.
Ma, X., Ma, C., Wu, C., Xi, Y., Yang, R., Peng, N., Zhang, C., Ren, F., 2021. Measuring human perceptions of streetscapes to better inform urban
renewal: A perspective of scene semantic parsing. Cities 110, 103086. doi:10.1016/j.cities.2020.103086.
Manton, R., Rau, H., Fahy, F., Sheahan, J., Clifford, E., 2016. Using mental mapping to unpack perceived cycling risk. Accident Analysis &
Prevention 88, 138–149. doi:10.1016/j.aap.2015.12.017.
Martin, A., Morciano, M., Suhrcke, M., 2021. Determinants of bicycle commuting and the effect of bicycle infrastructure investment in london:
First Author et al.: Preprint submitted to Elsevier Page 24 of 27
Evidence from uk census microdata. Economics and Human Biology 41. doi:10.1016/j.ehb.2020.100945.
McNeil, N., 2011. Bikeability and the 20-min Neighborhood: How Infrastructure and Destinations Influence Bicycle Accessibility. Transportation
Research Record 2247, 53–63. doi:10.3141/2247- 07.
Munira, S., Sener, I.N., Zhang, Y., 2021. Estimating Bicycle Demand in the Austin, Texas Area: Role of a Bikeability Index. Journal of Urban
Planning and Development 147, 04021036. URL:,
doi:10.1061/(ASCE)UP.1943-5444.0000725. publisher: American Society of Civil Engineers.
Nagata, S., Nakaya, T., Hanibuchi, T., Amagasa, S., Kikuchi, H., Inoue, S., 2020. Objective scoring of streetscape walkability related to leisure
walking: Statistical modeling approach with semantic segmentation of Google Street View images. Health & Place 66, 102428. doi:10.1016/
Naik, N., Philipoom, J., Raskar, R., Hidalgo, C., 2014. Streetscore – Predicting the Perceived Safety of One Million Streetscapes. MIT web domain
Najafizadeh, L., Froehlich, J.E., 2018. A Feasibility Study of Using Google Street View and Computer Vision to Track the Evolution of Urban Ac-
cessibility, in: Proceedings of the 20th International ACM SIGACCESS Conference on Computers and Accessibility, Association for Computing
Machinery, New York, NY, USA. pp. 340–342. doi:10.1145/3234695.3240999.
Nazemi, M., van Eggermond, M.A.B., Erath, A., Schaffner, D., Joos, M., Axhausen, K.W., 2021. Studying bicyclists’ perceived level of safety using
a bicycle simulator combined with immersive virtual reality. Accident Analysis & Prevention 151, 105943. doi:10.1016/j.aap.2020.105943.
Neuhold, G., Ollmann, T., Bulo, S.R., Kontschieder, P., 2017. The Mapillary Vistas Dataset for Semantic Understanding of Street Scenes, in: 2017
IEEE International Conference on Computer Vision (ICCV), IEEE, Venice. pp. 5000–5009. doi:10.1109/ICCV.2017.534.
Neves, A., Brand, C., 2019. Assessing the potential for carbon emissions savings from replacing short car trips with walking and cycling using a
mixed GPS-travel diary approach. Transportation Research Part A: Policy and Practice 123, 130–146. doi:10.1016/j.tra.2018.08.022.
Nielsen, T., Skov-Petersen, H., 2018. Bikeability – urban structures suppor ting cycling. effects of local, urban and regional scale urban form factors on
cycling from home and workplace locations in denmark. Journal of Transport Geography 69, 36–44. doi:10.1016/j.jtrangeo.2018.04.015.
Nogal, M., Jiménez, P., 2020. Attractiveness of bike-sharing stations from a multi-modal perspective: The role of objective and subjective features.
Sustainability 12, 1–26. doi:10.3390/su12219062.
Osama, A., Albitar, M., Sayed, T., Bigazzi, A., 2020. Determining if walkability and bikeability indices reflect pedestrian and cyclist safety.
Transportation Research Record 2674, 767–775. doi:10.1177/0361198120931844.
Palliwal, A., Song, S., Tan, H.T.W., Biljecki, F., 2021. 3D city models for urban farming site identification in buildings. Computers, Environment
and Urban Systems 86, 101584. doi:10.1016/j.compenvurbsys.2020.101584.
Porter, A., Kohl, H., Pérez, A., Reininger, B., Gabriel, K., Salvo, D., 2018. Perceived social and built environment correlates of transportation and
recreation- only bicycling among adults. Preventing Chronic Disease 15. doi:10.5888/pcd15.180060.
Porter, A.K., Kohl, H.W., Pérez, A., Reininger, B., PetteeGabr iel, K., Salvo, D., 2020. Bikeability: Assessing the Objectively Measured Environment
in Relation to Recreation and Transportation Bicycling. Environment and Behavior 52, 861–894. doi:10.1177/0013916518825289.
Pritchard, R., Frøyen, Y., Snizek, B., 2019. Bicycle level of service for route choice—A GIS evaluation of four existing indicators with empirical
data. ISPRS International Journal of Geo-Information 8. doi:10.3390/ijgi8050214.
Qiu, W., Li, W., Liu, X., Huang, X., 2021. Subjectively Measured Streetscape Perceptions to Inform Urban Design Strategies for Shanghai. ISPRS
International Journal of Geo-Information 10, 493. URL: 9964/10/8/493, doi:10.3390/ijgi10080493.
number: 8 Publisher: Multidisciplinary Digital Publishing Institute.
Resch, B., Puetz, I., Bluemke, M., Kyriakou, K., Miksch, J., 2020. An interdisciplinary mixed-methods approach to analyzing urban spaces:
The case of urban walkability and bikeability. International Journal of Environmental Research and Public Health 17, 1–20. doi:10.3390/
Rojas-Rueda, D., Nazelle, A.d., Tainio, M., Nieuwenhuijsen, M.J., 2011. The health risks and benefits of cycling in urban environments compared
with car use: health impact assessment study. BMJ 343, d4521. doi:10.1136/bmj.d4521.
Russo, A., Ommeren, J.v., Dimitropoulos, A., 2019. The environmental and welfare implications of parking policies. doi:10.1787/16d610cc-en.
Schmid-Querg, J., Keler, A., Grigoropoulos, G., 2021. The Munich Bikeability Index: A Practical Approach for Measuring Urban Bikeability.
Sustainability 13, 1–14. doi:10.3390/su13010428.
Singapore Government, 2020a. Annual Motor Vehicle Population by Vehicle Type. URL:
annual-motor- vehicle-population- by-vehicle-type.
Singapore Government, 2020b. Master Plan 2019 Land Use layer. URL:
master-plan- 2019-land- use-layer.
Song, X.P., Richards, D.R., Tan, P.Y., 2020. Using social media user attributes to understand human–environment interactions at urban parks.
Scientific Reports 10, 808. doi:10.1038/s41598-020- 57864-4.
Sottile, E., Sanjust di Teulada, B., Meloni, I., Cherchi, E., 2019. Estimation and validation of hybrid choice models to identify the role of perception
in the choice to cycle. International Journal of Sustainable Transportation 13, 543–552. doi:10.1080/15568318.2018.1490465.
Srivastava, S., Divekar, A.V., Anilkumar, C., Naik, I., Kulkarni, V., Pattabiraman, V., 2021. Comparative analysis of deep learning image detection
algorithms. Journal of Big Data 8, 66. URL:, doi:10.1186/s40537- 021-00434- w.
Titze, S., Krenn, P., Oja, P., 2012. Developing a bikeability index to score the biking-friendliness of urban environments. Journal of Science and
Medicine in Sport 15, S29–S30. doi:10.1016/j.jsams.2012.11.071.
Toikka, A., Willberg, E., Mäkinen, V., Toivonen, T., Oksanen, J., 2020. The green view dataset for the capital of Finland, Helsinki. Data in Brief
30, 105601. doi:10.1016/j.dib.2020.105601.
Tokyo Metropolitan Government, 2018. Maps of Cities in Tokyo. URL:
Tran, P.T.M., Zhao, M., Yamamoto, K., Minet, L., Nguyen, T., Balasubramanian, R., 2020. Cyclists’personal exposure to traffic-related air pollution
and its influence on bikeability. Transportation Research Part D: Transport and Environment 88, 102563. doi:10.1016/j.trd.2020.102563.
First Author et al.: Preprint submitted to Elsevier Page 25 of 27
Verma, D., Jana, A., Ramamritham, K., 2020. Predicting human perception of the urban environment in a spatiotemporal urban setting using locally
acquired street view images and audio clips. Building and Environment 186, 107340. doi:10.1016/j.buildenv.2020.107340.
Villeneuve, P.J., Ysseldyk, R.L., Root, A., Ambrose, S., DiMuzio, J., Kumar, N., Shehata, M., Xi, M., Seed, E., Li, X., Shooshtari, M., Rainham,
D., 2018. Comparing the Normalized Difference Vegetation Index with the Google Street View Measure of Vegetation to Assess Associations
between Greenness, Walkability, Recreational Physical Activity, and Health in Ottawa, Canada. International Journal of Environmental Research
and Public Health 15, 1719. URL: 4601/15/8/1719, doi:10.3390/ijerph15081719. number: 8 Publisher:
Multidisciplinary Digital Publishing Institute.
Volker, J.M.B., Handy, S., 2021. Economic impacts on local businesses of investments in bicycle and pedestrian infrastructure: a review of the
evidence. Transport Reviews , 1–31doi:10.1080/01441647.2021.1912849.
Wahlgren, L., Stigell, E., Schantz, P., 2010. The active commuting route environment scale (ACRES): development and evaluation. International
Journal of Behavioral Nutrition and Physical Activity 7, 58. doi:10.1186/1479- 5868-7- 58.
Wakamiya, S., Siriaraya, P., Zhang, Y., Kawai, Y., Aramaki, E., Jatowt, A., 2019. Pleasant Route Suggestion based on Color and Object Rates, in:
Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, Association for Computing Machinery, New York,
NY, USA. pp. 786–789. URL:, doi:10.1145/3289600.3290611.
Wang, M., Vermeulen, F., 2020. Life between buildings from a street view image: What do big data analytics reveal about neighbourhood organi-
sational vitality? Urban Studies , 0042098020957198doi:10.1177/0042098020957198.
Wang, R., Lu, Y., Wu, X., Liu, Y., Yao, Y., 2020. Relationship between eye-level greenness and cycling frequency around metro stations in Shenzhen,
China: A big data approach. Sustainable Cities and Society 59, 102201. doi:10.1016/j.scs.2020.102201.
Wang, X., Lindsey, G., Schoner, J.E., Harrison, A., 2016. Modeling Bike Share Station Activity: Effects of Nearby Businesses and Jobs on Trips
to and from Stations. Journal of Urban Planning and Development 142, 04015001. doi:10.1061/(ASCE)UP.1943-5444.0000273. publisher:
American Society of Civil Engineers.
Wang, X., Rodríguez, D.A., Sarmiento, O.L., Guaje, O., 2019a. Commute patterns and depression: Evidence from eleven Latin American cities.
Journal of Transport & Health 14, 100607. doi:10.1016/j.jth.2019.100607.
Wang, Y., Zhang, D., Liu, Y., Dai, B., Lee, L.H., 2019b. Enhancing transportation systems via deep learning: A survey. Transportation Research
Part C: Emerging Technologies 99, 144–163. doi:10.1016/j.trc.2018.12.004.
Weld, G., Jang, E., Li, A., Zeng, A., Heimerl, K., Froehlich, J.E., 2019. Deep Learning for Automatically Detecting Sidewalk Accessibility
Problems Using Streetscape Imagery, in: The 21st International ACM SIGACCESS Conference on Computers and Accessibility, Association for
Computing Machinery, New York, NY, USA. pp. 196–209. URL:, doi:10.1145/3308561.
Winters, M., Brauer, M., Setton, E.M., Teschke, K., 2013. Mapping Bikeability: A Spatial Tool to Support Sustainable Travel. Environment and
Planning B: Planning and Design 40, 865–883. doi:10.1068/b38185.
Winters, M., Teschke, K., Brauer, M., Fuller, D., 2016. Bike Score®: Associations between urban bikeability and cycling behavior in 24 cities.
International Journal of Behavioral Nutrition and Physical Activity 13, 18. doi:10.1186/s12966- 016-0339-0.
Wu, A.N., Biljecki, F., 2021. Roofpedia: Automatic mapping of green and solar roofs for an open roofscape registry and evaluation of urban
sustainability. Landscape and Urban Planning 214, 104167. doi:10.1016/j.landurbplan.2021.104167.
Yamazaki, D., Ikeshima, D., Tawatari, R., Yamaguchi, T., O’Loughlin, F., Neal, J.C., Sampson, C.C., Kanae, S., Bates, P.D., 2017. A high-
accuracy map of global terrain elevations: Accurate Global Terrain Elevation map. Geophysical Research Letters 44, 5844–5853. doi:10.
Yao, Y., Wang, J., Hong, Y., Qian, C., Guan, Q., Liang, X., Dai, L., Zhang, J., 2021. Discovering the homogeneous geographic domain of human
perceptions from street view images. Landscape and Urban Planning 212, 104125. URL:
article/pii/S0169204621000888, doi:10.1016/j.landurbplan.2021.104125.
Ye, C., Zhang, F., Mu, L., Gao, Y., Liu, Y., 2020. Urban function recognition by integrating social media and street-level imagery. Environment
and Planning B: Urban Analytics and City Science , 239980832093546doi:10.1177/2399808320935467.
Ye, Y., Richards, D., Lu, Y., Song, X., Zhuang, Y., Zeng, W., Zhong, T., 2019a. Measuring daily accessed street greenery: A human-scale approach
for informing better urban planning practices. Landscape and Urban Planning 191, 103434. doi:10.1016/j.landurbplan.2018.08.028.
Ye, Y., Xie, H., Fang, J., Jiang, H., Wang, D., 2019b. Daily Accessed Street Greenery and Housing Price: Measuring Economic Performance of
Human-Scale Streetscapes via New Urban Data. Sustainability 11, 1741. doi:10.3390/su11061741.
Yeh, C.C., Lin, C.Y., Hsiao, J.H., Huang, C.H., 2019. The effect of improving cycleway environment on the recreational benefits of bicycle tourism.
International Journal of Environmental Research and Public Health 16. doi:10.3390/ijerph16183460.
Yencha, C., 2019. Valuing walkability: New evidence fromcomputer vision methods. Transportation Research Part A: Policy and Practice 130, 689–
709. URL:, doi:10.1016/j.tra.2019.09.053.
Zhang, F., Wu, L., Zhu, D., Liu, Y., 2019a. Social sensing from street-level imagery: A case study in learning spatio-temporal urban mobility
patterns. ISPRS Journal of Photogrammetry and Remote Sensing 153, 48–58. doi:10.1016/j.isprsjprs.2019.04.017.
Zhang, J., Mucs, D., Norinder, U., Svensson, F., 2019b. LightGBM: An Effective and Scalable Algorithm for Prediction of Chemical Toxic-
ity–Application to the Tox21 and Mutagenicity Data Sets. Journal of Chemical Information and Modeling 59, 4150–4158. URL: https:
//, doi:10.1021/acs.jcim.9b00633. publisher: American Chemical Society.
Zhang, Y., Siriaraya, P., Wang, Y., Wakamiya, S., Kawai, Y., Jatowt, A., 2018. Walking down a Different Path: Route Recommendation based on
Visual and Facility based Diversity, in: Companion Proceedings of the The Web Conference 2018, International World Wide Web Conferences
Steering Committee, Republic and Canton of Geneva, CHE. pp. 171–174. URL:, doi:10.
Zhao, J., Sun, G., Webster, C., 2020. Walkability scoring: Why and how does a three-dimensional pedestrian network matter? Environment and
Planning B: Urban Analytics and City Science , 239980832097787doi:10.1177/2399808320977871.
Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A., 2018. Places: A 10 Million Image Database for Scene Recognition. IEEE transactions
First Author et al.: Preprint submitted to Elsevier Page 26 of 27
on pattern analysis and machine intelligence 40, 1452–1464. doi:10.1109/TPAMI.2017.2723009.
Zhou, H., Liu, L., Lan, M., Zhu, W., Song, G., Jing, F., Zhong, Y., Su, Z., Gu, X., 2021. Using Google Street View imagery to capture micro built
environment characteristics in drug places, compared with street robbery. Computers, Environment and Urban Systems 88, 101631. doi:10.
First Author et al.: Preprint submitted to Elsevier Page 27 of 27
... The improvement in obtaining street view imagery (SVI) methods enables easy and remote access to abundant eye-level images from the perspectives of pedestrians and cyclists (Ito & Biljecki, 2021). Also, the developing technologies in computer vision (CV) result in the emergence of multiple efficient ways to process a large number of SVIs automatically (Ito & Biljecki, 2021;Jian Kang et al., 2018;Qiu et al., 2022;Xu et al., 2022). ...
... The improvement in obtaining street view imagery (SVI) methods enables easy and remote access to abundant eye-level images from the perspectives of pedestrians and cyclists (Ito & Biljecki, 2021). Also, the developing technologies in computer vision (CV) result in the emergence of multiple efficient ways to process a large number of SVIs automatically (Ito & Biljecki, 2021;Jian Kang et al., 2018;Qiu et al., 2022;Xu et al., 2022). The multidisciplinary implementation of remote sensing, computer vision, and machine learning is effective to make the best use of geographic data available from Earth observation sensors and ground imagery (Lefèvre et al., 2017). ...
... The mentioned technologies have been widely implemented to assess walkability (Nagata et al., 2020) and bikeability (Gu et al., 2018;Ito & Biljecki, 2021;Tran et al., 2020). However, compared to the studies of walking and cycling, less is known about the association between eye-level streetscapes and running behaviors (Huang, Jiang & Yuan, 2022). ...
Full-text available
The running environment is found to relate to running behaviors. Prior studies focused more on the mental benefits of runners' perceived quality of the natural environment. However, how built environment factors, especially the micro-level streetscapes and corresponding perceptions affect running intensity is less addressed. This might be due to the limited availability of large-scale per-ception and running data. Additionally, the spatial interaction effects on running are also less known. We hypothesize that the physical features and the perceptions of streetscapes both can af-fect runners’ route choices, which in turn can be reflected via the running intensity. To test our hy-pothesis, we explored the associations between the number of runners on streets and six groups of environmental factors (i.e., sociodemographic factors, safety attributes, land use and built environ-ment factors, traffic-related factors, objective street environment features, and subjective street envi-ronment perceptions) in Boston. To tackle the gap of lacking runner data, we took advantage of the semi-open-source Strava Heatmap to extract running intensity information. We then applied geo-graphic information system (GIS), deep learning, and computer vision to measure the key indices for streetscape quality and perceptions. The associations between running intensity and streetscape quality and perceptions were tested, while other variables are controlled. In addition to the ordinary least squares model (OLS), the spatial autoregressive combined model (SAC) and geographically weighted regression model (GWR) were also conducted to account for the spatial dependence and heterogeneity effects. The results indicated that: 1) the objective and subjective eye-level streetscape were both significant for runners’ route choices; 2) the proximity of natural elements including nat-ural land, tree density, green open space area, vegetation, and terrains alongside the streets could encourage running; 3) other built environment factors such as the points of interests (POIs), street light density and length of the street segment had positive impacts on running amount; 4) accessi-bility to running environment and transportation facilities could increase running amount; 5) safer street conditions (perceived safer streets and less crime) had positive impacts on running and; 6) the indices hindering running are crime density, population density, motorcycles and traffic lights on streets. The proposed framework based on semi-open data sources like Strava maps is effective and can be applied to larger geospatial contexts and even comparison studies between various cities, which significantly fill the gap of missing data on running and outdoor sports. Our findings also enrich the literature on the interaction between micro-level street environment and human behavior and can be used for urban environment improvement to facilitate running and promote public health. Keywords: Street View Image; Deep Learning; Street Measures; Running; Route Choice; Boston
... In parallel, coupling the UAV oblique photography and computer vision (CV) has become an important method for quantifying vast urban landscapes (Lyu et al., 2020). Thanks to the fast development of CV, such as semantic segmentation and object detection, studies on visual quality evaluation based on such trending techniques are proliferating Garg et al., 2021;Wilkins et al., 2022;Wu and Biljecki, 2021;Ito and Biljecki, 2021). CV can process the profusion of images automatically, objectively and efficiently, and it is not entirely new to riverscapes either (Sharma et al., 2021). ...
... These three types have their own characteristics, and each plays an instrumental role in spatial information sciences, producing significant volumes of data contributing to a wide range of domains and use cases ( Figure 1). The increasing production of imagery can be partly explained by the democratization of UAVs and SVI due to the decreasing cost of exploitation (Sun and Scanlon, 2019), the increase in the number of deployed satellites (Ghamisi et al., 2019), and the growing coverage of commercial services such as Google Street View, Baidu Maps, and volunteered geographic information (Yan et al., 2020;Ito and Biljecki, 2021). ...
Full-text available
Traditional approaches for visual perception and evaluation of river landscapes adopt on-site surveys or assessments through photographs. The former is expensive, hindering large-scale analyses, and it is conducted only on street-level or top-down imagery. The latter only reflects the subjective perception and also entails a laborious process. Addressing these challenges, this study proposes an alternative: a novel workflow for visual analysis of urban river landscapes by combining unmanned aerial vehicle (UAV) oblique photography with computer vision (CV) and virtual reality (VR). The approach is demonstrated with an experiment on a section of the Grand Canal in China where UAV oblique panoramic imagery has been processed using semantic segmentation for visual evaluation with an index system we designed. Concurrent surveys, immersive and non-immersive VR, are used to evaluate these photos, with a total of 111 participants expressing their perceptions across multiple dimensions. Then, the relationship between the people's subjective visual perception and the river landscape environment as seen by computers has been established. The results suggest that using this approach, rivers and surrounding landscapes can be analyzed automatically and efficiently, and the mean pixel accuracy (MPA) of the developed model is 90%, which advances state of the art. The results of this study can benefit urban planners in formulating riverside development policies, analyzing the perception of plans for a future scenario before an area is redeveloped, and the method can also aid relevant parties in having a macro understanding of the overall situation of the river as a basis for follow-up research. Due to simplicity, accuracy and effectiveness, this workflow is transferable and cost-effective for large-scale investigations of riverscapes and linear heritage. We openly release Semantic Riverscapes-the dataset we collected and processed, bridging another gap in the field.
... Open and conditional access street view images, such as Google Street View (GSV), Baidu Street View (BSV), and Tencent Street View, provide new data sources for quantifying ecological landscape characteristics of urban streets [21][22][23][24]. These street images were collected mainly by car and, consequently, stitched as VR photographs, then finally displayed as interactively panoramas, which were primarily designed for providing a geographic information service. ...
Full-text available
With the unprecedented urbanization processes around the world, cities have become the main areas of political, cultural, and economic creation, but these regions have also caused environmental degradation and even affected public health. Ecological landscape is considered as an important way to mitigate the impact of environmental exposure on urban residents. Therefore, quantifying the quality of urban road landscape and exploring its spatial heterogeneity to obtain basic data on the urban environment and provide ideas for urban residents to improve the environment will be a meaningful preparation for further urban planning. In this study, we proposed a framework to achieve automatic quantifying urban street quality by integrating a mass of street view images based on deep learning and landscape ecology. We conducted a case study in Xia-men Island and mapped a series of spatial distribution for ecological indicators including PLAND, LPI, AI, DIVISION, FRAC_MN, LSI and SHDI. Additionally, we quantified street quality by the entropy weight method. Our results showed the streetscape quality of the roundabout in Xiamen was relatively lower, while the central urban area presented a belt-shaped area with excellent landscape quality. We suggested that managers could build vertical greening on some streets around the Xiamen Island to improve the street quality in order to provide greater well-being for urban residents. In this study, it was found that there were still large uncertainties in the mechanism of environmental impact on human beings. We proposed to strengthen the in-depth understanding of the mechanism of environmental impact on human beings in the process of interaction between environment and human beings, and continue to form general models to enhance the ability of insight into the urban ecosystem.
... Mapillary 2 and KartaView 3 , now supply an enormous amount of georeferenced images around the world, often at a dense geographical resolution (Ma et al., 2019;Zhang et al., 2020;Ding et al., 2021). This source of data has been thoroughly exploited for a range of urban studies, some of which involve extracting information of buildings (Kruse et al., 2021;Ito and Biljecki, 2021;Rosenfelder et al., 2021;Yohannes et al., 2021;Kang et al., 2021;Helbich et al., 2021;Szcześniak et al., 2021;Cinnamon and Gaffney, 2021;Zhang et al., 2021a;Zhang et al., 2021b;Yin et al., 2021). ...
Full-text available
3D building models are an established instance of geospatial information in the built environment, but their acquisition remains complex and topical. Approaches to reconstruct 3D building models often require existing building information (e.g. their footprints) and data such as point clouds, which are scarce and laborious to acquire, limiting their expansion. In parallel, street view imagery (SVI) has been gaining currency, driven by the rapid expansion in coverage and advances in computer vision (CV), but it has not been used much for generating 3D city models. Traditional approaches that can use SVI for reconstruction require multiple images, while in practice, often only few street-level images provide an unobstructed view of a building. We develop the reconstruction of 3D building models from a single street view image using image-to-mesh reconstruction techniques modified from the CV domain. We regard three scenarios: (1) standalone single-view reconstruction; (2) reconstruction aided by a top view delineating the footprint; and (3) refinement of existing 3D models, i.e. we examine the use of SVI to enhance the level of detail of block (LoD1) models, which are common. The results suggest that trained models supporting (2) and (3) are able to reconstruct the overall geometry of a building, while the first scenario may derive the approximate mass of the building, useful to infer the urban form of cities. We evaluate the results by demonstrating their usefulness for volume estimation, with mean errors of less than 10% for the last two scenarios. As SVI is now available in most countries worldwide, including many regions that do not have existing footprint and/or 3D building data, our method can derive rapidly and cost-effectively the 3D urban form from SVI without requiring any existing building information. Obtaining 3D building models in regions that hitherto did not have any, may enable a number of 3D geospatial analyses locally for the first time.
... In addition, with the application of CV segmentation models, features can be effortlessly extracted as view indexes by computer for further analysis. For example, to examine housing prices , to assess bikeability (Ito & Biljecki, 2021), to measure pedestrian volume (Yin et al., 2015), and predict sun glare (Li et al., 2019). ...
Full-text available
Safe mobility and stable metro ridership are indispensable attributes of a healthy urban society. The Manhattan subway system serves 39% of its commuters as an essential public transit option; however, its annual ridership dropped by 3.48% from 2015 to 2018. Anecdotal evidence suggested that surging public fear of crime and assaults in and around subways might be a reason. At the same time, empirical studies indicated that transit crime and perceived safety could tremendously impact ridership. We hypothesize that ground-level urban-design quality would also relate to passengers' perceived safety and actual crime rates, subsequently affecting metro ridership. Nevertheless, how the juxtaposition of physical features and subjective perceptions of micro-scale street environments around subway stations correlates with crime frequencies is still unknown. We set out to quantify the correlations between crime reports and urban design quality within the ¼ mile buffer zone of Manhattan subway entrances. Our study extended the conventional crime model with objectively measured streetscapes and predicted human perceptions using the trending method of Street View Imagery (SVI) and the artificial intelligence of computer vision (CV) and machine learning (ML). We found that view indexes ("Person", "Sky", "Tree", "Wall", "Fence", "Signboard", etc.) along with subjective qualities (Safety, Enclosure and Complexity) are as performable as demographic attributes in explaining crime frequencies. Second, higher perceived safety does not necessarily link with lower crime risks. Lastly, parks as a point of interest (POI) serve as a crime deterrent. This study has great implications for urban design and transportation policies and provides references for other urban areas on how to facilitate safer public transit services and systems through enhancing built environments.
Full-text available
Unsupervised learning (UL) has a long and successful history in untangling the complexity of cities. As the counterpart of supervised learning, it discovers patterns from intrinsic data structures without crafted labels, which is believed to be the key to real AI-generated decisions. This paper provides a systematic review of the use of UL in urban studies based on 140 publications. Firstly, the topic, technique, application, data type, and evaluation method of each paper are recorded, deriving statistical insights into the evolution and trends. Clustering is the most prominent method, followed by topic modeling. With the strong momentum of deep learning, a growing application field of UL methods is representing the complex real-world urban systems at multiple scales through multi-source data integration. Subsequently, a detailed review discusses how UL is applied in a broad range of urban topics, which are concluded by four dominant themes: urbanization and regional studies, built environment, urban sustainability, and urban dynamics. Finally, the review addresses common limitations regarding data quality, subjective interpretation, and validation difficulty of the results, which increasingly require interdisciplinary knowledge. Research opportunities are found in the rapidly evolving technological landscape of UL and in certain domains where supervised learning dominates.
Scholarly interest in the accessibility of ridesharing services stems from debates within the transportation and planning communities on the inequality of access to transit and the growing digital divide embedded within novel forms of transit services. Contributing to such discussions, this paper considers the city of Atlanta as a case study and explores the links between the spatial disparity of accessibility to different Uber ridesharing products and features of the built environment extracted from Google Street View (GSV) imagery. The variability of wait time for an Uber service is used as a proxy of accessibility, while semantic image segmentation is performed on GSV imagery using a deep learning model DeepLabv3+ to identify notable spatial features captured at the eye-level perspective around service pick-up points. Results from spatial models show that proportions of built environment features such as buildings, vegetation, and terrains are associated with longer waiting times. In contrast, larger salient regions with foreground features are associated with shorter waiting times for several Uber service products.
Understanding perceptions of urban scenes help planners to alleviate social inequalities. While many studies objectively extracted pixels of physical features from street-view image to proxy perceived qualities, few studies utilize subjectively-collected perceptions from visual surveys to reveal social inequality. We argue that large divergence can exist between the two measures over a same perceptual concept, which can lead to different and even opposite spatial implications. Little is done to investigate their spatial divergences relating to socioeconomic indicators. To fill the gap, we related five pairwise perceptions extracted from street-view images at the neighborhood level to three socioeconomic indicators. Results show that implications on social inequality diverge greatly across Shanghai. First, subjective perceptions have a more robust correlation with socioeconomic indicators than objective counterparts. Second, lower income households reside in neighborhoods with lower perceptual qualities regardless which measure is utilized. Third, two measures converge in perceived walkability but diverge in greenness, while poorer subjective greenness and better subjective walkability are related to the elderly. We enrich literature on the urban-scale mapping of street scene perceptions and provide valuable guidance on the indicator selection for assessing, designing, and managing urban environment that alleviate social inequalities.
Full-text available
Urban morphology is important in a broad range of investigations across the fields of city planning, transportation, climate, energy, and urban data science. Characterising buildings with a set of numerical metrics is fundamental to studying the urban form. Despite the rapid developments in 3D geoinformation science, and the growing 3D data availability, most studies simplify buildings to their 2D footprint, and when taking their height into account, they at most assume one height value per building, i.e. simple 3D. We take the first step in elevating building metrics into full/true 3D, uncovering the use of higher levels of detail, and taking into account the detailed shape of a building. We set the foundation of the new research line on 3D urban morphology by providing a comprehensive set of 3D metrics, implementing them in openly released software, generating an open dataset containing 2D and 3D metrics for 823,000 buildings in the Netherlands, and demonstrating a use case where clusters and architectural patterns are analysed through time. Our experiments suggest the added value of 3D metrics to complement existing counterparts, reducing ambiguity, and providing advanced insights. Furthermore, we provide a comparative analysis using different levels of detail of 3D building models. ARTICLE HISTORY
Full-text available
Semantic 3D building models are provided by public authorities and can be used in applications, such as urban planning, simulations, navigation, and many others. Since large-scale 3D models are typically derived from top-view digital surface models (DSM), they can have detailed roof structures but render planes for façade elements. Furthermore, buildings' underpasses are often unmodeled, which impacts road space modeling and the building's volume score. For refining semantic 3D building models, point clouds obtained from mobile laser scanning (MLS) seem to be suitable. In this paper, we present a method of underpass reconstruction by comparing building models' façades with co-registered MLS measurements. As an alternative approach to from-scratch reconstruction, it exploits existing semantic 3D building models and street-level MLS point clouds to enhance models where required. The method considers the uncertainties of 3D models and measurements in a Bayesian network. Analyzed conflicts between the two representations resulting from ray tracing are used to delineate the underpass's contours on a façade. Generalized contours are extruded to 3D solid geometries and subtracted from a raw 3D building model, while the semantics is mapped to form an updated semantic 3D building model. The experiments show that the method reaches an accuracy of 12 cm while testing on CityGML LoD2 building models and the open point cloud datasets TUM-MLS-2016 and TUM-FAÇ ADE representing the Technical University of Munich (TUM) city campus. The validation reveals differences between the reconstructed and updated models in both volumes (up to 18%) and surfaces (up to 20%). Such an extension of road corridors can improve 3D map usage for vehicle navigation and urban simulations.
Full-text available
There is a prevailing trend to study urban morphology quantitatively thanks to the growing accessibility to various forms of spatial big data, increasing computing power, and use cases benefiting from such information. The methods developed up to now measure urban morphology with numerical indices describing density, proportion, and mixture, but they do not directly represent morphological features from the human's visual and intuitive perspective. We take the first step to bridge the gap by proposing a deep learning-based technique to automatically classify road networks into four classes on a visual basis. The method is implemented by generating an image of the street network (Colored Road Hierarchy Diagram), which we introduce in this paper, and classifying it using a deep convolutional neural network (ResNet-34). The model achieves an overall classification accuracy of 0.875. Nine cities around the world are selected as the study areas with their road networks acquired from OpenStreetMap. Latent subgroups among the cities are uncovered through clustering on the percentage of each road network category. In the subsequent part of the paper, we focus on the usability of such classification: we apply our method in a case study of urban vitality prediction. An advanced tree-based regression model (LightGBM) is for the first time designated to establish the relationship between morphological indices and vitality indicators. The effect of road network classification is found to be small but positively associated with urban vitality. This work expands the toolkit of quantitative urban morphology study with new techniques, supporting further studies in the future.
Full-text available
Street view imagery has rapidly ascended as an important data source for geospatial data collection and urban analytics, deriving insights and supporting informed decisions. Such surge has been mainly catalysed by the proliferation of large-scale imagery platforms, advances in computer vision and machine learning, and availability of computing resources. We screened more than 600 recent papers to provide a comprehensive systematic review of the state of the art of how street-level imagery is currently used in studies pertaining to the built environment. The main findings are that: (i) street view imagery is now clearly an entrenched component of urban analytics and GIScience; (ii) most of the research relies on data from Google Street View; and (iii) it is used across myriads of domains with numerous applications – ranging from analysing vegetation and transportation to health and socio-economic studies. A notable trend is crowdsourced street view imagery, facilitated by services such as Mapillary and KartaView, in some cases furthering geographical coverage and temporal granularity, at a permissive licence.
Full-text available
Recently, many new studies applying computer vision (CV) to street view imagery (SVI) datasets to objectively extract the view indices of various streetscape features such as trees to proxy urban scene qualities have emerged. However, human perception (e.g., imageability) have a subtle relationship to visual elements that cannot be fully captured using view indices. Conversely, subjective measures using survey and interview data explain human behaviors more. However, the effectiveness of integrating subjective measures with SVI datasets has been less discussed. To address this, we integrated crowdsourcing, CV, and machine learning (ML) to subjectively measure four important perceptions suggested by classical urban design theory. We first collected ratings from experts on sample SVIs regarding these four qualities, which became the training labels. CV segmentation was applied to SVI samples extracting streetscape view indices as the explanatory variables. We then trained ML models and achieved high accuracy in predicting scores. We found a strong correlation between the predicted complexity score and the density of urban amenities and services points of interest (POI), which validates the effectiveness of subjective measures. In addition, to test the generalizability of the proposed framework as well as to inform urban renewal strategies, we compared the measured qualities in Pudong to other five urban cores that are renowned worldwide. Rather than predicting perceptual scores directly from generic image features using a convolution neural network, our approach follows what urban design theory has suggested and confirmed as various streetscape features affecting multi-dimensional human perceptions. Therefore, the results provide more interpretable and actionable implications for policymakers and city planners.
Full-text available
Sustainable roofs, such as those with greenery and photovoltaic panels, contribute to the roadmap for reducing the carbon footprint of cities. However, research on sustainable urban roofscapes is rather focused on their potential and it is hindered by the scarcity of data, limiting our understanding of their current content, spatial distribution, and temporal evolution. To tackle this issue, we introduce Roofpedia, a set of three contributions: (i) automatic mapping of relevant urban roof typology from satellite imagery; (ii) an open roof registry mapping the spatial distribution and area of solar and green roofs of more than one million buildings across 17 cities; and (iii) the Roofpedia Index, a derivative of the registry, to benchmark the cities by the extent of sustainable roofscape in term of solar and green roof penetration. This project, partly inspired by its street greenery counterpart ‘Treepedia’, is made possible by a multi-step pipeline that combines deep learning and geospatial techniques, demonstrating the feasibility of an automated methodology that generalises successfully across cities with an accuracy of detecting sustainable roofs of up to 100% in some cities. We offer our results as an interactive map and open dataset so that our work could aid researchers, local governments, and the public to uncover the pattern of sustainable rooftops across cities, track and monitor the current use of rooftops, complement studies on their potential, evaluate the effectiveness of existing incentives, verify the use of subsidies and fulfilment of climate pledges, estimate carbon offset capacities of cities, and ultimately support better policies and strategies to increase the adoption of instruments contributing to the sustainable development of cities.
Full-text available
Walking behavior is influenced by both objective and subjective aspects of the built environment at the macro and micro scales. Most walkability studies focused on objective macro or mesoscale variables. The few studies that included microlevel indicators used various methods and sources to quantify street level urban design features, each with its own limitations. This study used drone photogrammetry to capture street features in a rapidly urbanizing area in the Philippines and showed that observational, distance, and view-related types of measurement can be done using a single 3D model. An inter-rater reliability test was conducted for observational indicators and showed good to excellent reliability. Using the quantified street features, we tested its correlation with scores generated from a walker perception survey to develop a composite walkability index that can be used for urban design and planning. Results showed that 13 walkability sub-models are statistically significant, wherein models pertaining to safety assumed the highest weights while complexity and imageability models ranked lowest. This study validated many of the street level indicators previously reported, while also suggested new ones. For some indicators, model effects were opposite of what was previously reported such as number of people, buildings with non-rectangular silhouettes and view of sky across, which reflect the unique characteristics of the study area. Findings provide new insights on walkability which may lead to improvements in the pedestrian environment, especially in the context of developing countries.
Full-text available
A computer views all kinds of visual media as an array of numerical values. As a consequence of this approach, they require image processing algorithms to inspect contents of images. This project compares 3 major image processing algorithms: Single Shot Detection (SSD), Faster Region based Convolutional Neural Networks (Faster R-CNN), and You Only Look Once (YOLO) to find the fastest and most efficient of three. In this comparative analysis, using the Microsoft COCO (Common Object in Context) dataset, the performance of these three algorithms is evaluated and their strengths and limitations are analysed based on parameters such as accuracy, precision and F1 score. From the results of the analysis, it can be concluded that the suitability of any of the algorithms over the other two is dictated to a great extent by the use cases they are applied in. In an identical testing environment, YOLO-v3 outperforms SSD and Faster R-CNN, making it the best of the three algorithms.
This study contributes to research and practice by demonstrating the use of a composite measure, a bikeability index, to facilitate the use of and improve the performance of direct demand models for bicycle traffic, especially when only limited observation is available. The city of Austin was selected as a case study to develop the model using bicycle volume from 44 intersections. Existing knowledge and data were leveraged to develop the bikeability index that encompasses multiple built environment features (bicycle route length, comfort, connectivity, destination density, and transit coverage) to quantify the bike-friendliness of the network. In addition to the index, the demand model contained five demographic and land use variables. Some of the variables provided unique insights into bike travel behavior within the city, such as the significant and positive influence of the presence of bike signals and bike-accessible bridges. Along with the improved scalability and transferability of the modeling approach, the results and discussion are expected to facilitate and/or guide informed strategies and educational programs to increase nonmotorized activity in Austin as well as other regions.
A better formalization of place - where people live, perceive, and interact with others - is crucial for understanding socioeconomic environment and human settlement. The widely used hedonic pricing model for houses was proposed from the perspective of space, focusing mostly on static house structural information and objective built environment factors. However, the value of house settlement is not only determined by its spatial settings, but also varies from one place to another with different cultures, human dynamics, human perceptions and social interactions. In this work, we introduce a place-oriented hedonic pricing model (P-HPM) that incorporates human dynamics and human perceptions of places to understand human settlement. As an empirical study, we employ a large volume of house price data in Boston and Los Angeles, including detailed house and locational amenity information. Besides, we take the hourly number of visits to places as a proxy of human mobility patterns, and obtain human perceptions of places extracted from large-scale street-view images using deep learning. The results show that the P-HPM outperformed the traditional HPM significantly in these two cities. Moreover, through a geographically weighted regression analysis and the Monte Carlo test, we find that the impacts of the proposed place-related variables on house prices are stable across space. Our results provide new insights into the assessment of human settlement values by incorporating the role of place using multi-source big geo-data.
The street view service, like Google Street View and Baidu Street View, has emerged as a research tool to capture the visual perception data of observers. This research explores the validity of Street View service tool using in visual perception assessment on historical block—can Street View service provide perceptual results consistent with reality. We conducted a survey based on the real environment and Baidu Street View pictures on two typical historical streets in Harbin, China. User’s subjective perception of the quality of streets and ambiance were compared. The findings show Street View has good validity on the subjective perception part of street quality, but on the ambiance perception, there is a significant difference between Street View and real site audit. Meanwhile, the validity of Street View on the ambiance differs for different types of streets.
The distribution of human perceptions in urban area was obtained. • This study first focuses on the spatial homogeneity of human perceptions. • A method is proposed to discover the homogeneous geographic domain of human perceptions. • This study explored the role of urban function in shaping human perceptions. Human perception of place refers to residents' psychological feelings about urban areas. Many studies of human perceptions have focused on a specific geographic location. Whether the distribution of human perceptions in continuous city space shows specific characteristics and how to disclose these phenomena remains a direction worth exploring. Due to cities' heterogeneity, quantitatively identifying the homogeneous perception regions at a fine scale within large urban regions is challenging. This study proposed a novel method to discover the homogeneous geographic domain of human perception using massive street view images. First, human perceptions of the urban visual environment were evaluated using street view images. Next, perception network models were constructed based on the road network and perception assessment results. Then, the Infomap community detection algorithm was used to identify homogeneous human perception communities. The qualitative and quantitative results verified our approach's effectiveness for capturing human perceptions' homogeneous geographic domain. Moreover, driving factor analysis was conducted to determine the urban function that may cause a community to be perceived differently based on point-of-interest (POI) data. In general, our method for combining human perceptions and the topology of urban roads could identify the homogeneous perception domain, which is valuable for urban structure studies and human perception assessment.