Modeling the impact of street-level built environment on cyclists’ route choice using street view images and GPS data

Abstract

Understanding how cyclists select their routes is crucial for designing infrastructure that promotes sustainable urban mobility. This study investigates the impact of street-level built environment features on cyclists’ route choices by integrating GPS trajectory data with Google Street View imagery. Using data collected in Utrecht, the Netherlands, we develop a comprehensive cycling network and employ deep learning-based image segmentation to quantify built environment attributes, such as greenery, road, and buildings. A path size logit model is applied to assess the influence of these factors on route selection. Our findings reveal that cyclists prefer routes with lower traffic light density, more greenery, and fewer motorized vehicles, while they tend to avoid routes with wide roads and dense building facades. Interestingly, while an overall green environment encourages cycling, an uneven distribution of greenery can deter cyclists. These insights highlight the need for a well-balanced urban streetscape to enhance cycling appeal. The results provide recommendations for urban planners to improve cycling infrastructure, emphasizing the importance of integrating green and blue spaces while ensuring safe and efficient travel paths. Future research will explore the impact of additional factors, such as noise and air quality, and further refine our understanding of cyclist route preferences.

Keywords

built environment street view image route choice cyclist

Introduction

Bicycles are increasingly recognized as a critical component of a sustainable transport future, offering benefits, such as reduced congestion, zero emissions, and improved public health. However, realizing this potential requires promoting cycling and ensuring that infrastructure supports and encourages its adoption. High-quality, well-placed cycling infrastructure shapes cyclists’ route choices, improves safety, and enhances the overall experience (Beecham et al., 2023; Ito et al., 2023). Despite growing investments in cycling networks, there is still a gap in understanding how cyclists choose their routes in real-world conditions. This research seeks to bridge that gap by leveraging novel data sources and advanced analytical methods to gain deeper insights into cyclist route choice behavior. By identifying the key factors that influence route selection, our findings will help to inform evidence-based investments in cycling infrastructure, ultimately fostering a more bike-friendly urban environment.

Compared to walking, bicycles offer a much larger travel distance within the same travel time, which makes them a key component of the X-min City concept (Allam et al., 2022). Regarding the importance of bicycles in sustainable urban development, numerous studies have been conducted to uncover the impact factors on bicyclists’ travel behavior (Heinen et al., 2010; Pearson et al., 2023; Yang et al., 2010). As one of the essential aspects of cyclists’ travel behavior, route choice modeling can indicate the cyclists’ preferences, which helps local governments plan proper bicycle infrastructure and promote bicycle usage. Benefiting from the rapid development of sensing technology, bicycle route choice modeling has experienced a significant rise. For example, highly detailed GPS data has been collected through mobile applications, which can form precise trajectories of the users. This study investigates cyclists’ route choices in Utrecht, Netherlands, utilizing GPS data from the Snifferbike system and street-level built environment details obtained from Google Street View images. GPS trajectories were adjusted using an advanced algorithm to ensure precision. A path size logit model was employed to analyze route choice behavior while addressing the impact of overlapping paths. The findings provide essential insights for urban planners and policymakers aiming to improve cycling infrastructure and promote active mobility. Our results underscore the significance of a well-designed street-level built environment on cyclists’ route preferences and willingness-to-cycle, offering valuable insights to enhance cycling safety and comfort. Additionally, by quantifying cyclists’ willingness to deviate from the shortest route for environmental advantages, this research supports evidence-based investments in cycling infrastructure.

Literature review

As for the route choice modeling, the researchers collected data from two major channels—the Stated Preference (SP) survey (Majumdar and Mitra, 2018; Stinson and Bhat, 2003) and the Revealed Preference (RP) survey (Chung et al., 2024; Shah and Cherry, 2021). The stated preference method has faced criticism due to the surveys’ hypothetical nature, which may not uncover the actual behavior (De Corte et al., 2021; Morrison, 2000). In contrast, RP surveys collect actual route choices made by participants, which provides data that reflects preferences in real-world decision-making contexts. However, the time-consuming data collection process limits the sample size. Additionally, there exist differences between the revealed and the actual routes due to the imprecisely recalled routes by participants (Stinson and Bhat, 2003). With the introduction of GPS, it overcomes the limitations of traditional surveys by automatically recording participants’ trajectories, making data collection more accurate and efficient. Therefore, it inspires research work in the route choice modeling of active transport modes (Broach et al., 2012; Zimmermann et al., 2017). Based on e-scooter data collected from Virginia Tech’s campus, Zhang et al. (2021) applied a recursive logit model to reveal the route preference of e-scooter riders. The results show that the e-scooter riders travel longer distances on the segments with bike lanes and multi-use paths. Hsueh and Lin (2023) conducted research to explore the influential factors of the route choices of scooter riders. The results point out that scooter riders are willing to use safer routes with longer travel time, distance, or costs. De Jong et al. (2023) applied linear regression analysis to investigate the cycling demand and route choice in Oslo, Norway. The result shows cyclists substantially deviate from the shortest paths and prefer flatter, water-facing, and less crossing routes. Meister et al. (2024) collected 3700 cycling trajectories and compared two path-based and one link-based route choice models for cyclists. The result indicates that all three models allow for the same behavioral interpretation. However, the in-depth interpretation of the impact factors has significant differences.

To uncover cyclists’ route preferences, it is necessary to generate feasible alternative routes and construct a choice set of options in addition to the selected routes (Halldórsdóttir et al., 2014). Therefore, the method of choice set generation and its size play crucial roles in route choice modeling (Bekhor et al., 2006; Prato and Bekhor, 2007). One commonly used choice set generation method is to create the shortest path (Lu et al., 2018). Existing research indicates that the routes chosen by cyclists may differ significantly from the shortest path. Some studies introduce additional alternative routes based on various route characteristics. Casello and Usyukov (2014) generated alternative bike routes based on the travel path’s directness and safety. Khatri et al. (2016) applied a labeled route method to create alternative routes, considering maximum biking infrastructure, shortest paths, minimum signalized intersections, least proportion of one-way road segments, and fewest junctions separately. Hood et al. (2011) proposed a route choice model based on GPS data from cyclists in San Francisco. The result shows that cyclists prefer to select routes with bicycle lanes, less steep slopes, short distances, and fewer turns. However, they also point out that traffic volume, traffic speed, number of lanes, crime rates, and nightfall had no effects on bicycle route choice. Different from the work of Hood et al. (2011), more works indicated that traffic volume is a significantly impact factor of bike route choice (Lu et al., 2018; Sener et al., 2009; Winters et al., 2010). By taking into account all the unique routes selected by shared bike users as both observed and alternative routes, Scott et al. (2021) conclude that users are willing to detour for certain attributes, such as bicycle facilities, but generally avoid circuitous routes, excessive turns, steep slopes, and roads with heavy traffic volume.

Recently, some researchers have pointed out that the built environment significantly impacts active mobility decisions (Yap et al., 2023; Zhang et al., 2024). They conducted an online survey to investigate the influence of greenness in urban streetscapes on the attractiveness of streets and the route choices of cyclists. The results show that most respondents prefer to cycle on green streets and avoid gray streets, which suggest that greening streetscapes is a promising way to enhance cycling in urban areas. Researchers developed a combination of bottom-up and top-down approaches to understand the influence of streetscape design features on cyclists’ route choices. The results indicated that despite differences between usual and ideal routes regarding streetscape design features (such as the presence of vegetation, water bodies, and road intersections), factors like infrastructure provision and quality have a more significant impact when selecting a route.

The recent rapid development of computer vision technology together with the availability of street view images (SVI) has provided opportunities for new approaches to research in active transport (Gao and Fang, 2025; Hankey et al., 2021; Ito and Biljecki, 2021). Ki and Lee (2021) compared the conventional greenery variables and the Green View Index (GVI) derived from SVI. Their findings revealed that GVI had a stronger correlation with walking time than conventional greenery metrics because of the ability to better reflect the actual greenery exposure experienced by pedestrians. By quantifying street-level factors from SVI, Koo et al. (2022) try to uncover how neighborhood- and street-level built environments would affect pedestrians’ behavior. The findings indicate that streetscape factors are more effective in explaining walking mode choices, as they capture the pedestrian experience from an eye-level perspective. Basu and Sevtsuk (2022) explored walk preference for various route attributes, including route length, number of turns, highway exposure, number of amenities, sidewalk width, Sky View Factor (SVF), GVI, and traffic volume, in Boston, MA, by using a path size logit model. The results indicate that longer routes or those with many turns are less likely to be selected. Conversely, routes with amenities, storefronts, and wide sidewalks are perceived more favorably and are more frequently chosen by pedestrians. Furthermore, pedestrians in Boston tend to prefer routes with greater sky visibility (i.e., streets with lower building edges) and natural greenery. For cycling route modeling, a street view-based audit approach is developed to evaluate the impact of built environment characteristics on adolescents’ cycling behaviors (Vanwolleghem et al., 2014; Verhoeven et al., 2018). However, the conducted works are based on quite a small sample size and did not quantify street-level factors.

In summary, most of the existing works on bike route choice did not provide insight into the impact of street-level built environments on route choice. In this study, the route choice of cyclists is explored based on collected GPS data and Google Street View collected from Utrecht, Netherlands; a city where biking is common and where cycling infrastructure is accessible throughout the city. In this work, we established a comprehensive cycling network containing cycling infrastructure, sidewalks, and local roads. Moreover, the google street view and GPS datasets are matched with the cycling network to extract street-level built environment information by using computer vision techniques. Therefore, we generated a choice route dataset that included the actual route, shortest route, green view route, blue view route, and route with the least traffic lights. Finally, the path size logit model is applied to explore the street-level built environment factors that affect cyclist route choice.

Hypothesis and analysis framework

Considering the empirical findings and limitations in existing works, we hypothesize that street view images can capture distinct features of the street-level built environment that are not evident in traditional spatial datasets, providing valuable insights into cyclists’ route choice preferences. Based on this hypothesis, our study pursues two main research objectives: (1) to quantify the influence of street-level built environment characteristics on cyclists’ route choices, and (2) to assess how attractive the current built environment is for cycling to urban amenities.

In order to achieve the objectives, an analysis framework is employed to analyze cyclist route choice behavior, as shown in Figure 1. It begins with data collected from Snifferbike, OpenStreetMap (OSM), and Google Street View. The Snifferbike data is applied to derive the actual cyclist with a map-matching algorithm and the road network from OSM. At the same time, origin-destination (OD) information is extracted from the Snifferbike data to generate a set of alternative routes. To characterize the visual attributes along these routes, a segmentation algorithm is applied to the Google Street View images to extract street-level built environment features. These features, combined with data from OSM, form a comprehensive set of built environment attributes for each route. By considering all the aforementioned information, a dataset with both actual and generated routes is delivered to feed the path size logit (PSL) model. This model estimates the likelihood of a route being chosen, correcting for the influence of overlapping paths through a path size adjustment factor. Ultimately, this framework enables a nuanced analysis of how specific elements of the built environment affect cyclists’ route choices, providing insights that can inform urban planning and cycling infrastructure design.

Figure 1.

Research design: Overall workflow.

Data

Study area and Snifferbike data

Utrecht, Netherlands, as shown in Figure 2, is renowned for its exceptional cycling infrastructure, making it one of the most bike-friendly cities in the world. This study area occupies an area of 99.32 km² with roughly 359,370 inhabitants.¹ The city’s dedication to cycling dates back to 1885 when it built the Netherlands’ first official cycle path on the Maliebaan. As of 2021, Utrecht features an extensive 420 km network of bicycle paths, ensuring safe and convenient travel for cyclists. It is also home to the world’s largest bicycle parking facility, accommodating 12,500 bicycles, with a total of over 30,000 bicycle parking spaces near Utrecht Central Station. The city provides 21 bicycle parking facilities, including 5 pop-ups, with 18 of them staffed and monitored. Cycling plays a major role in daily life, with 33,000 cyclists using the busiest route each day and 56% of residents choosing bicycles as their preferred way to reach the city center.² Because of the high bike ownership and advanced cycling infrastructure, the city of Utrecht is an ideal place to conduct cycling-related research (Knap et al., 2023; Van Duppen and Spierings, 2013).

Figure 2.

Cycling network of the study area.

In this study, data from the Stifferbike system³ is used to explore the impact of street-level built environments on cyclists’ route choices. The data contains 10,349 trips between June 2019 and April 2021, with GPS location collected every 10 sec. Each trip provides timestamps and GPS traces. Due to the error of GPS information, cyclist trajectories usually need to be map-matched to existing road networks. The map-matching algorithm developed by Yang and Gidófalvi (2018) is applied to convert cyclists’ trajectories into polyline trip routes.

Built environment information

In this work, street-level built environment information is extracted from the street view images downloaded from Google Street View Static API. First, we sampled the road network with 100-meter intervals in order to properly cover the network and reduce duplicated information. By considering the azimuth of the road, four street view images (SVIs) of each data point have been downloaded from Google. Each image covers a 90-degree perspective. Therefore, four images cover a 360-degree street view perspective which covers the observation range of cyclists. In total, we downloaded 55,122 street view images for 13,781 locations. Then, we utilized the MobileNetV3-Large image segmentation model, trained and validated on the Cityscapes Dataset, to extract detailed street-level built environment information (Howard et al., 2019). From street view images (SVIs), we identified 18 key elements: road, non-motorized transport (NMT) facilities, building, wall, fence, pole, traffic light, traffic sign, vegetation, terrain, sky, person, rider, car, truck, bus, train, motorcycle, and bicycle. These elements comprehensively characterize urban built environments and traffic conditions. We subsequently calculated the proportion of several elements (such as, roads, NMT facilities, buildings, walls, fences, vegetation, terrain, and sky), represented as the average pixel ratio across the dataset. To examine the influence of the street-level built environment on cycle route choice, we identified several candidate variables representing route attributes (see Table 1). A detailed description of these variables, along with their distributions across the chosen route and three alternative paths, is provided in Table 1. These variables were quantified using deep convolutional neural networks to process a large set of Google Street View images. All variables were measured at the segment level and aggregated to the route level using weighted averages, with segment lengths as weights before being scaled to a 0–1 range.

Table 1.

Descriptive statistics.

	Description	Chosen route			Alternatives
		Chosen route			Blue	Green	Shortest	Min. traffic lights
		Mean	min	Max	Mean	Mean	Mean	Mean
Distance	Route length in km	2.082	0.008	23.210	3.470	3.476	1.987	3.948
Max. road	Max. % of road visibility per km	13.403	0.021	159.193	63.709	62.037	61.039	60.497
Avg. road	Avg. % of road visibility per km	4.532	0.021	59.430	9.306	9.157	9.058	8.221
Min. road	Min. % of road visibility per km	1.789	0.000	26.203	0.712	0.709	0.710	0.652
Max. NMT facilities	Max. % of NMT facilities visibility per km	4.618	0.006	177.027	22.845	22.756	21.730	22.785
Avg. NMT facilities	Avg. % of NMT facilities visibility per km	1.448	0.004	16.146	2.851	2.827	2.724	2.680
Min. NMT facilities	Min. % of NMT facilities visibility per km	0.604	0.000	14.288	0.180	0.179	0.179	0.169
Max. building	Max. % of building visibility per km	9.005	0.004	242.040	39.169	38.055	37.739	49.238
Avg. building	Avg. % of building visibility per km	2.061	0.004	76.455	5.433	5.363	5.362	6.027
Min. building	Min. % of building visibility per km	0.502	0.002	44.735	0.352	0.351	0.353	0.336
Max. wall	Max. % of wall visibility per km	1.018	0.000	101.927	4.259	4.220	4.112	4.652
Avg. wall	Avg. % of wall visibility per km	0.209	0.000	12.762	0.393	0.392	0.384	0.406
Min. wall	Min. % of wall visibility per km	0.064	0.000	3.108	0.022	0.022	0.022	0.021
Max. fence	Max. % of fence visibility per km	1.794	0.000	239.266	7.972	8.014	7.879	8.806
Avg. fence	Avg. % of fence visibility per km	0.393	0.000	29.963	0.830	0.824	0.807	0.804
Min. fence	Min. % of fence visibility per km	0.108	0.000	3.496	0.039	0.039	0.039	0.036
Traffic sign	% of the route covered by traffic signs	0.860	0.000	1.000	0.902	0.901	0.902	0.890
Pole	% of the route has poles	0.996	0.000	1.000	0.995	0.995	0.995	0.992
Traffic light	% of the route has traffic lights	0.629	0.000	1.000	0.632	0.630	0.630	0.406
Max. green view	Max. % of green view visibility per km	20.874	0.115	598.059	58.793	60.579	58.052	65.007
Avg. green view	Avg. % of green view visibility per km	7.434	0.115	201.877	8.785	9.007	8.620	9.231
Min. green view	Min. % of green view visibility per km	3.170	0.000	57.738	1.254	1.259	1.253	1.193
Max. blue view	Max. % of blue view visibility per km	16.988	0.000	151.143	70.727	69.288	68.492	66.898
Avg. blue view	Avg. % of blue view visibility per km	6.115	0.000	107.518	10.312	10.094	9.969	8.956
Min. blue view	Min. % of blue view visibility per km	2.846	0.000	107.518	0.835	0.832	0.830	0.781
Person	Avg. % of person visibility per km	0.014	0.000	3.347	0.026	0.022	0.024	0.020
Car	Avg. % of car visibility per km	0.186	0.000	5.687	0.364	0.363	0.357	0.435
Truck	Avg. % of truck visibility per km	0.010	0.000	0.670	0.044	0.044	0.044	0.034
Bus	Avg. % of bus visibility per km	0.011	0.000	2.076	0.049	0.049	0.049	0.043
Train	Avg. % of train visibility per km	0.010	0.000	1.976	0.039	0.040	0.040	0.044
Motorcycle	Avg. % of motorcycle visibility per km	0.003	0.000	0.182	0.009	0.009	0.009	0.009
Bicycle	Avg. % of bicycle visibility per km	0.030	0.000	4.522	0.070	0.068	0.068	0.066

Methodology

Choice set generation

The cyclist route choice set is generated with the labeled route method proposed by Ben-Akiva et al. (1984). The method has been widely applied in cyclist route choice studies (Broach et al., 2010; Khatri et al., 2016). Four alternative routes were generated alongside one selected route. Using the labeled route method, Dijkstra’s algorithm was employed to identify alternative paths between the origin and destination, optimizing specific built environment attributes (Dijkstra, 1959). Apart from the shortest path, alternative routes were determined based on route-level attributes such as green view, blue view, and minimal traffic lights. These routes were integrated into the street networks of the study area to obtain attributes along each route.

Path size logit

The path size logit (PSL) model is an extension of the traditional multinomial logit model used in transportation and route choice analysis (Ben-Akiva and Bierlaire, 1999). It addresses the issue of overlapping routes by incorporating a path size factor into the utility function, which adjusts the probability of selecting a route based on its uniqueness compared to other alternatives. In transportation networks, many routes share common segments, leading to correlation among alternatives. This violates the Independence of Irrelevant Alternatives (IIA) property of the standard logit model. The PSL model corrects this by introducing a path size adjustment term. The path size factor ( ${P S}_{k}$ ) quantifies the uniqueness of a route k by considering the proportion of each link’s length to the total route length, weighted by the number of routes sharing that link. It is defined as follows:

{P S}_{k} = \sum_{a \in A_{k}} \frac{l_{a}}{L_{k}} \cdot \frac{1}{\sum_{m \in C} δ_{a m}}

(1)

where

A_{k}

is the set of links in route k,

l_{a}

is the length of link a,

L_{k}

is the total length of route k,

δ_{a m}

is an indicator (1 if link a is part of route k and 0 otherwise), and C is the set of all routes.

The utility of route k is modified to include the logarithm of the path size factor:

U_{k} = \sum {β_{k i} X}_{k i} + β_{p s} \cdot \ln ({P S}_{k}) + ε_{k}

(2)

where

X_{k i}

is the observed road attribute

X_{i}

on route k, such as road, NMT facilities, building, wall, fence, pole, traffic light, traffic sign, vegetation, terrain, sky, person, rider, car, truck, bus, train, motorcycle, and bicycle;

\ln ({P S}_{k})

is the logarithm of the path size factor;

β_{k i}

is the weight of different road attribute

X_{k i}

;

β_{p s}

is a coefficient that determines the weight of the path size adjustment; and

ε_{k}

is the error term.

The probability of choosing route k from the choice set C is expressed as follows:

P (k | C) = \frac{e^{U_{k}}}{\sum_{m \in C} e^{U_{m}}}

(3)

Including the path size factor ensures that routes with significant overlap receive lower adjusted utilities, reducing the likelihood of selecting similar alternatives. This makes the model more realistic and better reflects observed route choice behavior in networks with overlapping paths.

Results

Route choice preference

This study employed Path Size Logit models to explore the built environment factors that influenced route choice preferences. To effectively detect multicollinearity, the Variance Inflation Factor (VIF) is applied to select variables (Salmerón et al., 2018; Thompson et al., 2017). García et al. (2024) indicated that using the variable with VIF values lower than 10 is able to mitigate the collinearity. After applying the criteria above, the highly correlated variables are removed, and the rest of the variables are applied to develop the models. Built Environment indices are normalized with MinMaxScaler. The study primarily focused on one-way trips and excluded circular trips because it is challenging to generate alternative routes for round trips that have the same origin and destination points. The estimated results are shown in Table 2.

Table 2.

Model results (n = 29,762).

	Model 1				Model 2				Model 3
Variable	Beta		p-value	VIF	Beta		p-value	VIF	Beta		p-value
Distance	−5.937	***	0	3.5	−5.927	***	0	3.4	−0.001	†	0.812
Avg. road	−20.895	***	0	8.6	−20.358	***	0	7.6	−19.207	***	0
Min. road	19.777	***	0	6.4	19.652	***	0	6.3	13.540	***	0
Avg. NMT facilities	−39.242	***	0	7.1	−38.812	***	0	6.9	−24.525	***	0
Min. NMT facilities	21.012	***	0	5.9	20.555	***	0	5.9	12.849	***	0
Avg. building	−14.189	***	0	4.7	−14.420	***	0	4.7	−10.187	***	0
Avg. wall	7.717	***	0	2.4	6.741	***	0	2.1	−6.631	***	0
Avg. fence	−1.662		0.418	4.1
Pole	−6.187	***	0	8.9	−6.182	***	0	8.8	1.599	***	0
Traffic lights	−15.467	***	0	3.7	−16.819	***	0	2.6	−37.267	***	0
Max. green view	−25.117	***	0	4.0	−24.267	***	0	3.6	−27.453	***	0
Avg. green view	37.769	***	0	10.0	35.428	***	0	7.2	31.471	***	0
Min. green view	−1.919		0.22	6.6
Min. blue view	16.831	*	0.03	8.1	6.950	*	0.046	6.4	29.038	***	0
Avg. person	6.539	*	0.008	1.7	−9.781	**	0.004	1.7	9.979	***	0
Avg. car	−9.852	***	0	2.6	−4.255	***	0	2.5	−1.297	*	0.045
Avg. truck	1.695		0.567	1.6
Avg. bus	7.740		0.019	1.3
Avg. train	−5.106	*	0.003	1.4	−4.255	*	0.017	1.3	−31.610	***	0
Avg. motorcycle	−12.584	***	0	1.7	−12.463	*	0.015	1.7	−14.740	***	0
Avg. bicycle	9.683	***	0	2.6	9.773	***	0	2.6	7.246	***	0
Ln(Path size)	12.963	***	0	4.2	12.965	***	0	4.1
Rho-squared	0.8753				0.8751				0.4215
Log-likelihood	−2155.3				−2159.4				−9998.8
AIC	4354.58				4354.72				20,031.57

Significance codes: ***p < 0.001; **p < 0.01; *p < 0.05; †p < 0.1.

The results from Table 2 reveal several important insights into the impact of the built environment on cyclists’ route preferences. Both Model 1 and Model 2, demonstrate strong fits to the data with similar Rho-squared values of 0.8753 and 0.8751. Model 2 offers an improved approach by excluding insignificant variables, which do not strongly impact route choice in this dataset. Model 2 shows a slightly lower AIC value by removing insignificant variables. This simplification improves model interpretability and fit while maintaining robustness.

Several variables significantly influence cyclists’ route choices. Notably, the distance negatively impacts route utility which indicates the cyclists’ preference for shorter routes. Similarly, routes with wider roads will be less likely to be selected due to their association with less cycling-friendly environments. However, the positive effect of the minimum observed road indicates that cyclists will prefer the route to have a certain width of the road. This may be because the certain lane width of the road facility shared by cyclists and vehicles makes cyclists feel safer. At the same time, the result shows that the wider the NMT facilities the less likely cyclists will select the route. The reason behind that can be the wider the NMT facilities the more pedestrians are going to use the facility which may lead to the inconvenience or illegal of cycling on it. Meanwhile, cyclists would prefer to have NMT facilities throughout their cycling routes. Cyclists are less likely to select the routes that go through the residence area because of the significant negative impacts of the average building index throughout the route. However, the route with a higher wall index is preferred by cyclists which is possible because the wall separates the NMT facilities from the motorways. Another reason is that diverse facades would increase the cyclist’s preference for the route (Ramirez Juarez et al., 2023). The average green view index is positively associated with route preference, highlighting the preference for consistent greenery along routes. This finding is consistent with existing work (Ramirez Juarez et al., 2023). Interestingly, the higher maximum green view index has a negative influence, suggesting that cyclists would prefer routes with more evenly distributed greenery. Additionally, higher densities of visible people (average person) and bicycles (average bicycle) increase route attractiveness, possibly reflecting social and safety considerations. In contrast, the presence of higher densities of cars, trains, and motorcycles negatively impact route preferences, indicating a desire to avoid noisy and motorized environments. The result is consistent with the works conducted by Lu et al. (2018) and Sener et al. (2009). The minimum blue view index, representing the visibility of blue elements such as the sky or water, also positively influences route choice, underscoring the importance of scenic features. This is also consistent with the work of Ramirez Juarez et al. (2023) which pointed out that cyclists wish to avoid the routes with tall buildings. The results indicate that cyclists chose the routes with lower traffic light coverage. The result is consistent with the work conducted by Krenn et al. (2014) and Menghini et al.(2010).

We estimated an additional model (Model 3) with the same specification as Model 2, except for the omission of the path size correction. This adjustment was made to assess the impact of the path size correction on model goodness-of-fit and the coefficient values, along with their equivalent cycling distance effects (discussed further in the next subsection). The results indicate a significant decrease in rho-squared, from 0.8751 to 0.4219. While most coefficient estimates retain the same signs, their values change significantly. These findings highlight that failing to account for overlap between alternatives can substantially distort attribute effects, resulting in biased and inaccurate estimates that compromise the reliability of policy analysis.

Operationalizing coefficients as willingness-to-cycle measures

Using coefficient ratios to present a marginal rate of substitution between variables has been applied to measure the willingness-to-pay and willingness-to-walk (Basu and Sevtsuk, 2022; Olszewski and Wibowo, 2005). By capturing the tradeoffs between two variables, such coefficient ratios serve as valuable tools for policy-making. In this study, we use route distance as the reference variable, allowing the effects of built environment factors to be expressed in terms of “willingness-to-cycle.” For example, $β_{D i s t a n c e} = - 5.927$ and $β_{A v g . r o a d} = - 20.358$ in Model 2. We calculate the willingness-to-cycle for a 10% increase of avg. road as follows:

W T C (a v g . r o a d) = - \frac{{10 \times β}_{A v g . r o a d}}{β_{D i s t a n c e}} = - \frac{- 20.358 \times 10}{- 5.927} = - 34.35

(4)

This suggests that every additional 10% of average road observed along the route is perceived as a 34.35% increase in route distance, highlighting the significant cognitive burden that avg. road imposes on cyclists. In terms of willingness-to-cycle, this can be interpreted as: “each additional 10% of average road observed on the route reduces the willingness-to-cycle by 34.35%.” We present the willingness-to-cycle values for all built environment factors that were statistically significant in our models in Table 3. The values are derived from a PSL model with the path size correction (Model 2).

Table 3.

Interpretation of route choice model coefficients as “willingness-to-cycle.”

Variable	Willingness-to-cycle (WTC) in %	Model 2
Avg. road	An average increase in avg. road index by 10% changes WTC by:	−34.35
Min. road	An average increase in min. road index by 10% changes WTC by:	33.16
Avg. NMT facilities	An average increase in avg NMT facility index by 10% changes WTC by:	−65.49
Min. NMT facilities	An average increase in min. NMT facility index by 10% changes WTC by:	34.68
Avg. building	An average increase in avg. wall index by 10% changes WTC by:	−24.33
Avg. wall	An average increase in avg. building index by 10% changes WTC by:	11.37
Pole	An average increase in Pole index by 10% changes WTC by:	−10.43
Traffic lights	An average increase in Traffic lights index by 10% changes WTC by:	−28.38
Max. green view	An average increase in max. green view index by 10% changes WTC by:	−40.95
Avg. green view	An average increase in avg. green view index by 10% changes WTC by:	59.78
Min. blue view	An average increase in min. blue view index by 10% changes WTC by:	11.73

Figure 3 illustrates the willingness-to-cycle to various urban amenities. In this study, we classified these amenities into three categories: Educational facilities (e.g., schools, kindergartens, and childcare centers), Healthcare facilities (e.g., dentists, clinics, pharmacies, and hospitals), and Social facilities (e.g., libraries, art centers, and social centers). Figure 3 shows that the area near the facilities has better built environment that make cyclists feel less perceived distance which lead to a higher willingness-to-cycle. These examples point out the importance for urban planners to understand how the built environment at the street level influences the perceived distance of routes, and potentially the actual geographic boundaries of facility catchment areas that people are willing to cycle to. We anticipate that these insights will be valuable for other accessibility analyses focused on cyclists, such as identifying transit or food deserts or developing a transport poverty index for neighborhoods.

Figure 3.

Willingness-to-cycle to different Points of Interest.

Conclusion

In this study, a comprehensive analysis of cyclists’ route choices by integrating GPS data and street view images is conducted to understand the impact of the street-level built environment on cycling route choice behavior. Our findings confirm that multiple factors shape route choices, including road visibility, non-motorized transport (NMT) facilities, greenery, and other built environment characteristics. Consistent with previous research, our results highlight that cyclists prefer routes with proper NMT infrastructure, reduced vehicular presence, and enhanced environmental aesthetics, such as greenery and blue views. Interestingly, while the presence of green and blue spaces positively influences route selection, an uneven distribution of such elements can negatively impact cycling preferences. Since roads with minimal traffic lights enhance travel continuity, the presence of traffic lights decreases the preference of cyclists. Our analysis further emphasizes that cyclists prioritize routes that balance travel efficiency and comfort. Shorter distances remain a primary determinant of route choice, yet factors such as greenery, blue views, and the presence of urban design elements significantly contribute to a positive cycling experience.

From a policy perspective, our findings provide deep insights into cyclable city concept for urban planners and transportation officials. Investing in dedicated cycling infrastructure, particularly in areas with high vehicular traffic, can significantly enhance route attractiveness. Additionally, incorporating green and blue spaces into urban cycling networks can further encourage cycling by improving the overall user experience. However, attention must be given to the equitable distribution of such elements to ensure consistent exposure throughout a route.

Despite its contributions, this study has several limitations. While the study successfully extracted and quantified several built environment features, other potentially influential factors, such as noise levels, air quality, and social perceptions of safety, were not considered. Second, Utrecht is highly favorable to cyclists, which may not be able to represent the route choice behavior in other cities. More cities will be included in the future work in order to identify the similarities and differences. Third, more advanced method, such as adaptive path size logit (Duncan et al., 2020) and machine learning (Magnana et al., 2022; Marra and Corman, 2025), could be applied to better explain the data. Finally, route choices may vary under different weather and road conditions, as well as across age groups, are not considered in this study. Future research could expand on these aspects to provide a more comprehensive understanding of route choice determinants.

Footnotes

ORCID iD

Chaoru Lu

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was partially supported by the Direktoratet for høyere utdanning og kompetanse (HK-dir) (UTF-2020/10115) and GRC pilot: Big data technology-enabled sustainable and social just cities (No. 352058).

Declaration of conflicting interests

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

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.

Notes

Author biographies

Dr. Chaoru Lu is an Associate Professor specializing in smart mobility at the Department of Civil Engineering and Energy Technology at Oslo Metropolitan University (OsloMet) in Norway. He holds a Ph.D. from Iowa State University. His research focuses on intelligent transport systems, traffic flow theory, autonomous vehicles, and transport electrification.

Dr. Daniel Piatkowski is an Associate Professor of Integrated Land Use and Transportation Planning at Oslo Metropolitan University. He holds a Ph.D. from the University of Colorado Denver. His research focuses on sustainable mobility, active transportation, and how human behavior intersects with urban technologies.

Dr. John Östh is a Swedish human geographer and professor specializing in spatial analysis, GIS, and urban analytics. He currently holds a professorship in Urban Analysis at Oslo Metropolitan University and serves as a faculty member at Uppsala University’s Department of Human Geography. His research focuses on mobility, segregation, and socio-spatial inequalities, utilizing big data and geostatistical methods.

Dr. Karima Kourtit is at the Open Universiteit, Heerlen, the Netherlands. She was the lab owner at the Jheronimus Academy of Data Science (JADS) in the division of Smart Cities and Data Analytics. She holds two Ph.D.s in economics and geography (with distinction), and has a profound interest in regional and urban topics. Her research interest focuses on the emerging ‘New Urban World’. Her academic profile is characterized by a profound involvement in evidence-based urban and spatial research on urban development, smart city policy, digital technology, and data metrics.

Dr. Peter Nijkamp is a Dutch economist renowned for his extensive contributions to regional and urban economics, as well as economic geography. He served as a professor at Vrije Universiteit Amsterdam and is recognized as one of the world's most prolific economists, with a vast array of publications in spatial economics, transport systems, and sustainable development. In 1996, he was awarded the prestigious Spinoza Prize for his quantitative research in spatial economics. Throughout his career, Nijkamp has also held significant roles, including President of the Netherlands Research Council (NWO) and fellow of the Tinbergen Institute.

References

Allam

Nieuwenhuijsen

Chabaud

, et al. (2022) The 15-minute city offers a new framework for sustainability, liveability, and health. The Lancet Planetary Health 6: e181–e183.

Basu

Sevtsuk

(2022) How do street attributes affect willingness-to-walk? City-wide pedestrian route choice analysis using big data from Boston and San Francisco. Transportation Research Part A: Policy and Practice 163: 1–19.

Beecham

Yang

Tait

, et al. (2023) Connected bikeability in London: which localities are better connected by bike and does this matter. Environment and Planning B: Urban Analytics and City Science 50: 2103–2117.

Ben-Akiva

Bierlaire

(1999) Discrete choice methods and their applications to short term travel decisions. Handbook of Transportation Science. Springer Nature.

Bekhor

Ben-Akiva

Ramming

(2006) Evaluation of choice set generation algorithms for route choice models. Annals of Operations Research 144: 235–247.

Ben-Akiva

Bergman

Daly

, et al. (1984) Modelling inter urban route choice behaviour. In Papers presented during the Ninth International Symposium on Transportation and Traffic Theory held in Delft the Netherlands, 11-13 July 1984 .

Broach

Gliebe

Dill

(2010) Calibrated labeling method for generating bicyclist route choice sets incorporating unbiased attribute variation. Transportation Research Record: Journal of the Transportation Research Board 2197: 89–97.

Broach

Dill

Gliebe

(2012) Where do cyclists ride? A route choice model developed with revealed preference GPS data. Transportation Research Part A: Policy and Practice 46: 1730–1740.

Casello

Usyukov

(2014) Modeling cyclists’ route choice based on GPS data. Transportation Research Record 2430(1): 155–161.

10.

Chung

Yao

Pan

, et al. (2024) Understanding the route choice preferences of private and dock-based public bike users using GPS data in Seoul, South Korea. Journal of Transport Geography 116: 103845.

11.

de Corte

Cairns

Grieve

(2021) Stated versus revealed preferences: an approach to reduce bias. Health Economics 30: 1095–1123.

12.

de Jong

Böcker

Weber

(2023) Road infrastructures, spatial surroundings, and the demand and route choices for cycling: evidence from a GPS-based mode detection study from Oslo, Norway. Environment and Planning B: Urban Analytics and City Science 50: 2133–2150.

13.

Dijkstra

(1959) A note on two problems in connexion with graphs. Numer. Math 271: 269–271.

14.

Duncan

Watling

Connors

, et al. (2020) Path size logit route choice models: issues with current models, a new internally consistent approach, and parameter estimation on a large-scale network with GPS data. Transportation Research Part B: Methodological 135: 1–40.

15.

Gao

Fang

(2025) Pedaling through the cityscape: unveiling the association of urban environment and cycling volume through street view imagery analysis. Cities 156: 105573.

16.

García García

Salmerón Gómez

García Pérez

(2024) A review of ridge parameter selection: minimization of the mean squared error vs. mitigation of multicollinearity. Communications in Statistics - Simulation and Computation 53: 3686–3698.

17.

Halldórsdóttir

Rieser-Schussler

Axhausen

, et al. (2014) Efficiency of choice set generation methods for bicycle routes. European journal of transport and infrastructure research 14(4): 332–348.

18.

Hankey

Zhang

HTK

, et al. (2021) Predicting bicycling and walking traffic using street view imagery and destination data. Transportation Research Part D: Transport and Environment 90: 102651.

19.

Heinen

van Wee

Maat

(2010) Commuting by bicycle: an overview of the literature. Transport Reviews 30: 59–96.

20.

Hood

Sall

Charlton

(2011) A GPS-based bicycle route choice model for San Francisco, California. Transportation Letters 3: 63–75.

21.

Howard

Sandler

Chu

, et al. (2019) Searching for MobileNetV3. Proceedings of the IEEE/CVF International Conference on Computer Vision. 1314–1324.

22.

Hsueh

Lin

(2023) Influential factors of the route choices of scooter riders: a GPS-based data study. Journal of Transport Geography 113: 103719.

23.

Ito

Biljecki

(2021) Assessing bikeability with street view imagery and computer vision. Transportation Research Part C: Emerging Technologies 132: 103371.

24.

Ito

Morgan

Lovelace

(2023) Where to invest in cycle parking: a portfolio management approach to spatial transport planning. Environment and Planning B: Urban Analytics and City Science 50: 1438–1454.

25.

Khatri

Cherry

Nambisan

, et al. (2016) Modeling route choice of utilitarian bikeshare users with GPS data. Transportation Research Record: Journal of the Transportation Research Board 2587: 141–149.

26.

Lee

(2021) Analyzing the effects of green view Index of neighborhood streets on walking time using Google Street View and deep learning. Landscape and Urban Planning 205: 103920.

27.

Knap

Ulak

Geurs

, et al. (2023) A composite X-minute city cycling accessibility metric and its role in assessing spatial and socioeconomic inequalities – a case study in Utrecht. the Netherlands. Journal of Urban Mobility 3: 100043.

28.

Koo

Guhathakurta

Botchwey

(2022) How are neighborhood and street-level walkability factors associated with walking behaviors? A big data approach using street view images. Environment and Behavior 54: 211–241.

29.

Krenn

Oja

Titze

(2014) Route choices of transport bicyclists: a comparison of actually used and shortest routes. International Journal of Behavioral Nutrition and Physical Activity 11: 1–7.

30.

Scott

Dalumpines

(2018) Understanding bike share cyclist route choice using GPS data: comparing dominant routes and shortest paths. Journal of Transport Geography 71: 172–181.

31.

Magnana

Rivano

Chiabaut

(2022) Implicit GPS-based bicycle route choice model using clustering methods and a LSTM network. PLoS One 17: 1–20.

32.

Majumdar

Mitra

(2018) Analysis of bicycle route-related improvement strategies for two Indian cities using a stated preference survey. Transport Policy 63: 176–188.

33.

Marra

Corman

(2025) Modelling route choice in public transport with deep learning. Transportation (Ams).

34.

Meister

Liang

Felder

, et al. (2024) Comparative study of route choice models for cyclists. Journal of Cycling and Micromobility Research 2: 100018.

35.

Menghini

Carrasco

Schüssler

, et al. (2010) Route choice of cyclists in Zurich. Transportation Research Part A Policy Pract 44: 754–765.

36.

Morrison

(2000) Aggregation biases in stated preference studies. Australian Economic Papers 39: 215–230.

37.

Olszewski

Wibowo

(2005) Using equivalent walking distance to assess pedestrian accessibility to transit stations in Singapore. Transportation Research Record: Journal of the Transportation Research Board 1927: 38–45.

38.

Pearson

Berkovic

Reeder

, et al. (2023) Adults’ self-reported barriers and enablers to riding a bike for transport: a systematic review. Transport Reviews 43: 356–384.

39.

Prato

Bekhor

(2007) Modeling route choice behavior: How relevant is the composition of choice set? Transportation Research Record 2003(1): 64–73.

40.

Ramirez Juarez

Grigolon

Madureira

(2023) Cyclists’ perception of streetscape and its influence on route choice: a pilot study with a mixed-methods approach. Transportation Research Part F: Traffic Psychology and Behaviour 99: 374–388.

41.

Salmerón

García

(2018) Variance inflation factor and condition number in multiple linear regression. Journal of Statistical Computation and Simulation 88: 2365–2384.

42.

Scott

Brown

(2021) Route choice of bike share users: Leveraging GPS data to derive choice sets. Journal of Transport Geography 90: 102903.

43.

Sener

Eluru

Bhat

(2009) An analysis of bicycle route choice preferences in Texas. Transportation 36: 511–539.

44.

Shah

Cherry

(2021) Different safety awareness and route choice between frequent and infrequent bicyclists: findings from revealed preference study using bikeshare data. Transportation Research Record: Journal of the Transportation Research Board 2675: 269–279.

45.

Stinson

Bhat

(2003) Commuter bicyclist route choice: analysis using a stated preference survey. Transportation Research Record: Journal of the Transportation Research Board 1828: 107–115.

46.

Thompson

Kim

Aloe

, et al. (2017) Extracting the variance in flation factor and other multicollinearity diagnostics from typical regression results. Basic and Applied Social Psychology 39: 81–90.

47.

Van Duppen

Spierings

(2013) Retracing trajectories: the embodied experience of cycling, urban sensescapes and the commute between “neighbourhood” and “city” in Utrecht, NL. Journal of Transport Geography 30: 234–243.

48.

Vanwolleghem

Van Dyck

Ducheyne

, et al. (2014) Assessing the environmental characteristics of cycling routes to school: a study on the reliability and validity of a Google Street View-based audit. International Journal of Health Geographics 13: 1–9.

49.

Verhoeven

Van Hecke

Van Dyck

, et al. (2018) Differences in physical environmental characteristics between adolescents’ actual and shortest cycling routes: a study using a Google Street View-based audit. International Journal of Health Geographics 17: 1–15.

50.

Winters

Teschke

Grant

, et al. (2010) How far out of the way will we travel? Built environment influences on route selection for bicycle and car travel. Transportation Research Record: Journal of the Transportation Research Board 2190: 1–10.

51.

Yang

Gidófalvi

(2018) Fast map matching, an algorithm integrating hidden Markov model with precomputation. International Journal of Geographical Information Science 32: 547–570.

52.

Yang

Sahlqvist

McMinn

, et al. (2010) Interventions to promote cycling: systematic review. BMJ 341: 870.

53.

Yap

Chang

J-H

Biljecki

(2023) Incorporating networks in semantic understanding of streetscapes: contextualising active mobility decisions. Environment and Planning B: Urban Analytics and City Science 50: 1416–1437.

54.

Zhang

Buehler

Broaddus

, et al. (2021) What type of infrastructures do e-scooter riders prefer? A route choice model. Transportation Research Part D: Transport and Environment 94: 102761.

55.

Zhang

Liu

, et al. (2024) The effects of street environment features on road running: an analysis using crowdsourced fi tness tracker data and machine learning. Environment and Planning B: Urban Analytics and City Science 51: 529–545.

56.

Zimmermann

Mai

Frejinger

(2017) Bike route choice modeling using GPS data without choice sets of paths. Transportation Research Part C: Emerging Technologies 75: 183–196.