Exploring the Relationship Between the Built Environment and Bike-Sharing Usage as a Feeder Mode Across Different Metro Station Types in Shenzhen

Abstract

Bike-sharing has been widely recognized for addressing the “last-mile” problem and improving commuting efficiency. While prior studies emphasize how the built environment shapes feeder trips, the effects of station types and spatial heterogeneity on bike-sharing and metro integration remain insufficiently explored. Taking the urban core area of Shenzhen as a case study, this paper examines how the built environment influences such integration during morning peak hours and how these impacts differ across station types. First, we proposed a “3Cs” (convenience, comfort, and caution) framework to capture key built environment factors. Metro stations were classified into commercial, residential, and office types via K-means clustering. Subsequently, the ordinary least squares (OLS) regression model and the multiscale geographically weighted regression (MGWR) model were employed to identify significant factors and explore the spatial heterogeneity of these effects. Results reveal that factors influencing bike-sharing–metro integration vary by station type. While land-use mix and enclosure affect bike-sharing usage across all stations, employment and intersection density are only significant for commercial stations. Furthermore, these influences exhibit spatial heterogeneity. For instance, at office-oriented stations, population shows both positive and negative effects across areas, while residential density has a generally negative impact. These findings enhance our understanding of how the built environment shapes bike-sharing–metro integration patterns and support more targeted planning interventions.

1. Introduction

With the acceleration of China’s urbanization, issues such as traffic congestion and high energy consumption have become increasingly severe. As an effective way to ease environmental pressure and improve commuting efficiency, shared bikes help meet residents’ mobility needs and support multimodal transfers [1]. Their speed advantage over walking makes them a better solution to the first-and-last-mile problem, thereby enhancing transit accessibility and equity [2]. As one of China’s first low-carbon pilot cities and a model for public transportation development, Shenzhen has witnessed shared bicycles emerge as a vital mode of transportation for commuters. According to data from the Shenzhen Municipal Transportation Bureau, approximately 52.3% of daily bike-sharing trips serve as a feeder mode to public transit, with usage primarily concentrated during weekday peak hours [3].

Given the context above, increasing attention has been given to feeder trips between bike-sharing and public transportation in recent years [4,5,6,7,8,9]. Among the factors shaping people’s travel behavior, the built environment has been identified as a key determinant influencing residents’ choice of feeder modes [7,8]. In this context, a growing number of studies have explored the relationship between the built environment and the integration of bike-sharing with metro systems, confirming that factors such as population density, location conditions, and land use around metro stations significantly influence the choice of bike-sharing as a feeder mode [4,6,7,10]. These studies have often measured built environment characteristics within the service areas of metro stations, analyzing them as a unified whole. However, as the direct environment in which cycling behavior occurs, the impact of both macro and micro road features on bike-sharing usage is significant and should not be overlooked [11,12], and effective urban design along cycling routes plays a key role in enhancing the comfort and convenience of cycling [13]. While many studies have considered road-related built environment variables, such as road density and intersection density [6,14], few have focused on the impact of road characteristics (i.e., sky openness, traffic safety infrastructure, greenery) surrounding metro stations on commuters’ feeder behaviors.

Furthermore, given that different types of metro stations, such as commercial-oriented, residential-oriented, and employment-oriented stations, exhibit variations in location, interchange level, passenger volume, and travel patterns [15,16], the surrounding built environment, such as land use, road pattern, and streetscape, also differs across station types [17,18]. Existing studies have proved that both station characteristics and its surrounding environment are associated with commuters’ feeder choice [5,7,17]. Therefore, categorizing metro stations within the study area is essential, as it allows for a more comprehensive consideration of how station-specific characteristics mediate the effects of the built environment on feeder behavior. However, there is limited research on how built environment effects vary across different stations, which limits the ability to provide targeted urban design and transport planning recommendations that address the specific needs of different station contexts.

To fill these gaps, we focus on the urban core area of Shenzhen, China, which is the city’s concentration area for metro stations and bike-sharing usage. This study examines how the surrounding road environment influences bike-sharing and metro integration during morning peak hours across different types of metro stations by identifying key influencing factors and their spatial heterogeneity. The main research questions are as follows: (1) How do the built environment and feeder characteristics around different metro stations vary? (2) What built environment factors around various types of stations have a significant impact on the integration of bike-sharing and metro usage? (3) What is the spatial heterogeneity of these influential features? To solve these problems, we first develop a “3Cs” framework (convenience, comfort, and caution) to comprehensively assess the impact of the road built environment around metro stations on the integration of bike-sharing and metro. Second, we use K-means clustering to categorize metro stations based on their surrounding transportation characteristics, population density, and land use. Third, the OLS regression model is applied to identify significant factors influencing bike-sharing around different station types. Finally, we employ a multiscale geographically weighted regression (MGWR) model to further examine the spatial heterogeneity of these factors.

The remainder of this paper is organized as follows: Section 2 presents a literature review on bike-sharing–metro integrated use and the measurement of the built environment; Section 3 describes the study area, data collection, and major methods; Section 4 introduces the results of the metro station classification, OLS regression model, and the MGWR model; Section 5 presents the discussion of the results. Finally, Section 6 concludes the paper with the main findings, policy recommendations, and limitations.

2. Literature Review

2.1. Built Environment Determinants of Bike-Sharing–Metro Integrated Use

2.1.1. Built Environment Correlates of Cycling Behavior

The integration of bike-sharing and metro systems is influenced not only by urban infrastructure and economic policies but also closely associated with the built environment characteristics surrounding metro stations [19,20]. In terms of built environment research, Cervero et al. initially explored the impact of the built environment on travel behavior through the “3Ds” framework, including density, diversity, and design, in 1997 [21]. Building upon this foundation, Cervero et al. later expanded the framework to the “5Ds” by incorporating distance to transit and destination accessibility [22]. By 2010, Ewing and Cervero further developed the widely cited “6Ds” framework, which added the category of demand management and has since become one of the most commonly used theories in built environment studies [23,24].

In terms of convenience-related factors, Ni and Chen confirmed that residential and employment density can promote bike-sharing usage [25]. However, some studies have also found that these factors may inhibit bike-sharing in certain contexts [6,26]. Apart from specific land use types, the overall degree of land-use mix can also encourage residents’ cycling behavior [26,27]. Population size has also been shown to significantly affect cycling, but there is no consensus on the specific direction of this effect [4,8]. Furthermore, road-related variables such as road and intersection density can also influence cycling behavior, as supported by earlier studies [6,14].

Although the aforementioned studies have provided a relatively comprehensive consideration of the impact of the built environment, they remain focused on objective factors, often overlooking cyclists’ perceptual experiences on the cycling route, including comfort and safety [8,12,28]. In recent years, an increasing number of scholars have incorporated perceived built environment factors into their research. Specifically, a study in India found that the comfort of the in-car environment and journey safety can affect commuters’ choices between using cars or public transportation [29,30]. Meanwhile, a more aesthetically pleasing and safer street environment can stimulate people’s desire to walk [31]. Furthermore, existing studies have demonstrated that perceived factors such as cycling safety [8,10] and street design (e.g., greenery, enclosure, and shade) [12,32] also significantly shape residents’ cycling behavior, indicating that the impact of psychological perception factors on people’s travel behavior cannot be ignored. For example, Bai et al. discovered that greater enclosure could promote cycling behavior [33], since well-enclosed spaces often suggest a sense of safety and vitality [32], and thus encourage more physical activity [34]. While Song et al. suggested that enclosure could adversely affect the cycling experience [12]. Moreover, sky openness and greenery also have psychological value [33], by improving the mood of residents [35,36]. Previous studies have shown a positive correlation between sky openness, street greenness, and cycling [12,33,37]. In addition, with the emergence of big data methods such as street view imagery and its integration with machine learning (ML), the correlation between these perceived built environment factors and cycling can be further confirmed. For example, Ito and Biljecki combined street view imagery and computer vision (CV) techniques to incorporate micro-level factors into an assessment system for bikeability [38]. Bai et al. used street view imagery and ML methods to confirm that streetscapes promote the use of bike-sharing as a feeder mode among students [37]. Although the existing literature has considered the influence of objective and subjective built environment on cycling behaviors, few studies have integrated them to provide a comprehensive assessment of their impact on feeder trips.

As for built environment data extraction, when analyzing the built environment around metro stations, previous studies tend to establish a buffer centered on the metro station entrance. This buffer is then analyzed as a whole to examine how built environment factors such as population density, road density, and the land-use mix affect feeder trips [5,6,7]. The research on how both the subjective and objective built environments of roads affect cycling efficiency and integrated travel remains limited [11,12].

2.1.2. Metro Station Types and Travel Behavior

Existing studies often classify metro stations based on their inherent characteristics, such as location [17], passenger volume [16,39], and interchange level [16,17]. These characteristics have been shown to be closely associated with the travel behavior of residents in the surrounding areas [17,40,41]. Due to variations in location and urban functions, different types of stations are often associated with significant differences in the surrounding built environment, such as land use, facility density, and street network structure [15,17,18,42]. Previous studies have shown a high degree of consistency between residents’ feeder behavior and the built environment factors around metro stations [15,17]. In particular, according to typical riding distance for shared bike users [43], for bike-sharing and metro integration, the area within 800–2000 m of a metro station is typically regarded as the major attractive space that generates a large number of potential demands [9,10,44], and studies have found that land use, population density, road network characteristics, and perceived built environment factors within this area are significantly associated with bike-sharing usage. For example, Ma and Dill found that within a half-mile radius (approximately 804.67 m) around metro stations, the number of retail stores and road density can promote bike-sharing usage; Sun et al. found that within a 2000 m radius around metro stations, road density, POI characteristics, and land-use mix are closely associated with bike-sharing–metro feeder use [7]; Guo and He studied the 800 m buffer area, finding a strong correlation between a perceived built environment factor (traffic safety) and bike-sharing volume around the station [8]. However, these studies analyze all metro stations in the study area as a homogeneous whole, without categorizing metro stations by functional type. Therefore, this may overlook both the influence of the metro station itself on bike-sharing–metro integration usage and the differences in the surrounding built environment across different types of metro stations, making it difficult to implement more targeted policies around each station type.

To better explore the impact of the built environment around different types of stations on the use of bike-sharing–metro integrated usage, it is essential to first classify the metro stations in the study area. Clustering techniques such as K-means, expectation maximization, and hierarchical clustering are commonly employed to categorize metro stations within a region [17]. For instance, Wu et al. identified five TOD types in Wuhan based on the Place–Node model and K-means clustering [18]. Yang et al. categorized Beijing metro stations into seven categories using the K-means method by taking the station passenger flow characteristics and land use characteristics as the clustering criteria [45]. Xia and Gai, based on the distribution of population distribution, development intensity, public transportation connection, road network length distribution, station location attributes, and station passenger flow data of six dimensions of clustering factors, applied the K-means method to categorize Shenyang metro stations into five categories: residential, commercial and business, comprehensive development, industrial, and transportation hub [46].

2.2. Evaluation for Impact of Built Environment

When exploring the impact of built environment elements on bike-sharing usage, it is essential to first identify the significant factors that affect the dependent variable. Specifically, Wang et al. and Guo et al. used a regression model to test the relationship between the built environment and the cycling volume around metro stations [44,47]. However, using only the traditional regression model would ignore the spatial heterogeneity of the effects and could not further explore the relationship between variables that would be affected by different geographic locations [45].

Therefore, studies have started to introduce geographically weighted regression (GWR). For example, Zheng et al., when studying the impact of the built environment on ride-hailing in Chengdu, in addition to applying the OLS regression model to select the variables, also established a GWR model to investigate the impact of variables in different geographical locations [48]. Li et al. similarly combined the OLS and GWR models to explore the effects of factors such as the proportion of transportation facility coverage, population, and land use on the use of shared bicycles in Shanghai and its spatial heterogeneity [49]. However, GWR uses the same bandwidth for each variable [50], which may reduce the robustness of the model [45], while multiscale geographically weighted regression (MGWR) uses variable bandwidths and different geographic weighting matrices for each geographic object, which avoids the degradation of the performance of the dependent variable and its correlations due to overfitting and underfitting of the variables, as compared to the GWR method using fixed bandwidths and differentiated weighting matrices [51].

3. Methods

3.1. Study Area

The urban core area of Shenzhen was selected as the study area (Figure 1). Shenzhen is a mega-city in southern China, located in the Pearl River Delta near Hong Kong. Its urban core area comprises the three central districts of Luohu, Nanshan, and Futian, as well as Xin’an and Xixiang Streets of Bao’an District, Minzhi and Longhua Streets of Longhua District, and Bantian, Buji, Jihua, and Nanwan Streets of Longgang District. These areas cover four key functional centers: Qianhai, Nanshan, Futian, and Luohu, which are densely populated and have a high demand for bike-sharing usage.

As a transit-oriented city, Shenzhen possesses a highly modernized rail network, with over 40% of daily trips made via the metro system [7]. It is also one of the largest cities that have experienced an increase in DBS usage [10]. According to the Shenzhen Bike-Sharing Development Assessment Report, as of April 2019, the city had two major bike-sharing operators (“ofo” and “Mobike”) managing a fleet of 480,000 bikes, serving 26.47 million registered users with an average of 849,000 daily rides. Of these rides, 54.43% were for commuting, and 45% took place during morning and evening peak hours, underscoring the role of dockless bike-sharing (DBS) in complementing public transit for first- and last-mile connectivity. This integration is further highlighted by the fact that 98.6% of shared bicycles are concentrated in metro-dense areas. The two main operators charge between 0.6 yuan (approximately $0.08) and 2.5 yuan (approximately $0.35) per hour.

3.2. Data

This study includes two main datasets: the bike-sharing datasets and the built environment datasets.

The bike-sharing records data were obtained from the Shenzhen Government open data platform, which included vehicle IDs, start and end locations (X and Y coordinates), and start and end times (accurate to the second). The data were collected during the morning peak hour (07:00–09:00 AM) over three weekdays from 24 August to 26 August 2021, which were considered representative of typical weekday bike-sharing usage patterns [8]. During this period, the average temperature ranged between 29 °C and 30 °C under cloudy conditions, which were suitable for cycling. A total of 1,186,362 records of bike-sharing trips were selected, offering a comprehensive view of cycling activity in the study area.

As for the source of built environment data, population data were obtained from Worldpop (https://hub.worldpop.org/), which provides a 100 m precision raster dataset of the population. POIs data, including residential communities, restaurants, companies, and shopping and entertainment facilities, were sourced from GaodeMap (https://lbs.amap.com/) using an API interface; metro line and station data were obtained from BaiduMap via an API interface (https://lbsyun.baidu.com/). Road information that covers all levels of roads in the study area was collected from OpenStreetMap (https://www.openstreetmap.org/), which is a reliable open data source for road-related elements [8,37]. Finally, we sourced the street-view image (SVI) data from BaiduMap with Python3.9 to measure the micro-level built environment elements of the cycling route.

3.3. Variables

3.3.1. Measure of the Dependent Variable: Bike-Sharing–Metro Integrated Use

The dependent variable was the volume of bike-sharing usage as a feeder mode on the road segment over three days, which was measured in the following three steps:

Step1: Identifying integrated usage

In this phase, a 100 m buffer was established for each metro station entry, since parking or picking up vehicles within 100 m of the station entry is considered effective feeder use [9,44]. GIS was then applied to crop out the bike-sharing data with the start and end points within the 100 m buffer, and duplicated data were deleted. Additionally, empirical studies suggest that shared bikes around metro stations are usually used for short-distance feeder trips, with travel distances usually ranging from 200 to 3000 m [52,53]. Since the original bike-sharing data did not provide direct information on travel distance, usage time was used as a proxy. Based on the average cycling speed in Shenzhen [3], trips with durations between 96 and 1440 s were selected.

Step2: Cleaning road data

According to previous travel behavior studies, the 85th percentile value of the cumulative distribution of transfer distance is often used to define the search radius. Guo and He demonstrated that the distribution of 85% cumulative transfer distances for access and egress usage of shared bikes in Shenzhen refers to 1960 m and 2040 m, respectively [9]. Therefore, in this study, an average value of 2000 m was chosen as the radius to measure the built environment elements of roads around the metro station. Based on this buffer, bike-accessible roads were then extracted and segmented at intersections, and road centerlines were derived, resulting in a total of 4568 road segments for subsequent analysis.

Step3: Calculating bike-sharing volume

Subsequently, the selected data were then imported into GIS based on the start and end points. Using the shortest path-matching tool in network analysis, the OD trajectories of bike-sharing users were reconstructed based on the actual road network. The frequency of shared bicycle passages on each road was then summarized, representing the bike-sharing flow associated with feeder trips.

3.3.2. Measure of the Independent Variable: The Built Environment

This study employs built environment factors as independent variables, which are conceptualized based on the “3Cs” framework. To comprehensively identify the built environment features across the road, a 50 m linear buffer was created for each road segment, which basically includes the roadway and its surrounding buildings, open spaces, traffic facilities, and green areas that may affect cycling behavior [54]. For population density, road density, and intersection density, we created 500 m buffers around the road centerline because we needed to consider the surrounding contexts [38]. Before the analysis, the built environment data were coordinate transformed to ensure that all the data were under the same coordinate system, before being clipped based on the mentioned 50 m or 500 m buffer.

We used the summarized built environment and travel behavior studies as a basis to construct the “3Cs” framework, which comprehensively combines both objective and subjective built environment factors [8,55,56] with commuters’ demand [57]. Under this framework, the built environment variables in this study were categorized into convenience, comfort, and caution, containing 13 variables in total (

Table 1. The “3Cs” built environment framework.

Category	Variable	Definition
Convenience	Population density	Population density within a 500 m circular buffer around the road centroid (persons/km2)
	Residential density	Number of residential facilities within a 50 m linear buffer per kilometer of road length (facilities/km)
	Employment density	Number of employment facilities within a 50 m linear buffer per kilometer of road length (facilities/km)
	Catering density	Number of catering facilities within a 50 m linear buffer per kilometer of road length (facilities/km)
	Shopping and entertainment density	Number of shopping and entertainment facilities within a 50 m linear buffer per kilometer of road length (facilities/km)
	Land-use mix	POI diversity within a 50 m linear buffer
	MHD	Road network centrality and accessibility calculated by the sDNA model
	Road density	Total road length within a 500 m circular buffer around the road centroid divided by the buffer area (km/km2)
Comfort	Sky openness	% of pixels classified as the sky
	Greenery	% of pixels classified as trees
	Enclosure	% of pixels classified as buildings, columns, walls, and trees
Caution	Intersection density	Number of intersections within a 500 m circular buffer around the road centroid (intersections/km2)
	Traffic safety infrastructure density	% of pixels classified as pillars, guardrails, and street lights

).

(1)

Convenience

The category of convenience includes variables related to urban density, POIs, and location, which reflect the commuting efficiency and practicality of using shared bikes as a feeder mode.

Population density serves as a direct indicator of the potential demand for feeder services [1]; and the density of employment, residential, catering, shopping, and entertainment facilities, whose spatial distributions were quantified by POI data, is expected to enhance the cycling convenience and integrated use [9]. The formula is as follows:

$$\mathsf{\rho }={\displaystyle \frac{{\mathrm{N}}_{\mathrm{i}}}{\mathrm{A}}}$$

(1)

$${\mathsf{\rho }}_{\mathrm{i}}={\displaystyle \frac{{\mathrm{N}}_{\mathrm{i}}}{\mathrm{L}}}$$

(2)

where $\mathsf{\rho }$ is the density of population; N_i is the number of people or the POI in the buffer; A is the area of the corresponding buffer; and L is the length of the road segment.

Land-use mix has been widely used to measure the functional complexity in built environment studies and is expected to promote the integrated usage of bike-sharing [58,59]. In this study, 4 primary types and 31 subtypes of POI data were used to measure this variable. Specifically, the number and type of POI in each buffer were calculated using GIS. Subsequently, the proportion of each type in each buffer was calculated to arrive at a measure of the land-use mix, also known as the Shannon–Wiener diversity index (SHDI) [33], the formula is as follows:

$$\mathrm{H}=-{\sum }_{\mathrm{m}=1}^{\mathrm{n}}({\mathrm{P}}_{\mathrm{m}}·\ln {\mathrm{P}}_{\mathrm{m}})$$

(3)

where H is the land-use mix degree; n is the number of all POI types contained in the buffer; and ${\mathrm{P}}_{\mathrm{m}}$ is the proportion of type m in the whole buffer.

Road density can reflect the optionality for cycling path [60], the formula is as follows:

$${\mathrm{d}}_{\mathrm{road}}={\displaystyle \frac{\mathrm{L}}{\mathrm{A}}}$$

(4)

where ${\mathrm{d}}_{\mathrm{road}}$ is road density; L is the total length of roads in the buffer; and A is the buffer area.

MHD is computed based on the sDNA model, reflecting the centrality and accessibility of each road segment [54].

(2)

Comfort

Comfort refers to cyclists’ subjective perception of the built environment, specifically including sky openness, greenery, and enclosure.

Sky openness represents the portion of the sky visible from the ground [33]; low openness can lead to a sense of oppression. Greenery enhances the attractiveness of the landscape [28] and provides shade for the bike lane. Enclosure describes the sense of confinement created by vertical elements like buildings, walls, and trees that define the street edges [12,61]. In line with Jia et al., we calculated the proportion of pixels classified as buildings, columns, walls, and trees to quantify this variable [62].

To effectively measure the comfort-related variables discussed above, we extracted street-view sampling points at 100 m intervals along the road segments and retrieved street view images (SVI) from BaiduMap for each point in four directions: 0°, 90°, 180°, and 270° [37]. The size of each SVI was 512 × 512 pixels, and a total of 47,684 valid SVIs were selected. To extract these elements and calculate the ratios of the pixels classified as them over the total number of pixels in the image [38], we applied the widely used Pyramid Scene Parsing Network (PSPNet) [12,37,63]. Using the model pre-trained on the ADE20K dataset, we quantified 150 objective elements in each street segment, ranging from natural objects such as the sky, grass, and trees, to public infrastructure elements like buildings, walls, and columns [12]. Elements unrelated to the variables were excluded, and a total of 11 elements were ultimately selected. Each variable was calculated as the sum of pixel proportions for all relevant visual elements according to its definition, and the pixel proportion was averaged across all sampling points within each street segment to obtain representative values. The formula for each point is as follows:

$$\mathrm{Pr}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^4{\mathrm{Pixel}}_{\mathrm{element}}}{{\sum }_{\mathrm{i}=1}^4{\mathrm{Pixel}}_{\mathrm{total}}}}$$

(5)

where Pr is the pixel ratio of each element, such as trees, grass, columns, and sky; ${\mathrm{Pixel}}_{\mathrm{element}}$ is the pixels of the element in one SVI; and ${\mathrm{Pixel}}_{\mathrm{total}}$ is the total pixels in this SVI.

(3)

Caution

Caution refers to the objective facility configurations and road characteristics, including two variables: traffic safety infrastructure density and intersection density. Traffic infrastructure facilities such as pillars, guardrails, and street lights can regulate traffic order and ensure cycling safety, which were calculated using the same method as comfort-related variables in this study. However, too many intersections can increase the risk of conflict between cyclists and other traffic flows [9], and the calculation formula for its density is as follows:

$${\mathrm{d}}_{\mathrm{Intersection}}={\displaystyle \frac{{\mathrm{N}}_{\mathrm{Intersection}}}{\mathrm{A}}}$$

(6)

where ${\mathrm{d}}_{\mathrm{Intersection}}$ is the intersection density; ${\mathrm{N}}_{\mathrm{Intersection}}$ is the number of intersections inside the buffer; and A is the buffer area.

3.4. Models

The study first categorized all metro stations in the urban core area of Shenzhen using K-means clustering into commercial-oriented, residential-oriented, and office-oriented stations. Next, we examined the built environment and identified significant factors influencing bike-sharing usage by constructing an OLS regression model. Finally, an MGWR model was employed to further examine the spatial heterogeneity of these factors across station types.

3.4.1. K-Means Clustering Algorithm

The K-means clustering method is widely adopted for categorizing data points into discrete groups according to their similarity [64]. The process of metro station classification in this study consists of four main steps: (1) selecting and calculating the clustering variables; (2) standardizing the variables; (3) performing factor analysis; and (4) applying the K-means algorithm.

First, we classified the clustering variables into three categories: transportation characteristics, population density, and land use [17,42]. Transportation characteristics include four factors: interchange level, road length, station location, and bus stop density, while land use includes POI-based factors representing residential, commercial, and office functions. In line with Xia and Gai, we used the number of transfer routes at each station to represent the interchange level and the distance from the city center to characterize station location [46], while other variables were calculated based on an 800 m buffer around the station.

After computing the variables, the data were standardized using the Z-score method in SPSS Statistics 26 software. Factor analysis was then conducted, extracting three common factors from the original eight variables to serve as input for the K-means clustering algorithm. Based on practical considerations, the clustering result with K = 3 was ultimately selected as the classification of metro stations. Finally, as stated by Abdullahi et al. [64], the K-means clustering algorithm was applied.

3.4.2. OLS Regression Model

To identify the built environment factors that significantly influence bike-sharing–metro integrated usage around different types of metro stations, an ordinary least squares (OLS) regression model was employed. This model is widely used in built environment and travel behavior studies to explore the relationship between dependent and independent variables [7,49]. Before conducting the regression analysis, multicollinearity among the built environment variables was assessed using the variance inflation factor (VIF) to ensure the reliability of the model and avoid the influence of collinearity [7].

Subsequently, the OLS model was employed to analyze the impact of 13 built environment factors on bike-sharing and metro integrated use around three types of metro stations. The formula is as follows:

$${\mathrm{Y}}_{\mathrm{i}}={\mathsf{\beta }}_0+\sum _{\mathrm{i}}{\mathsf{\beta }}_{\mathrm{i}}{\mathrm{x}}_{\mathrm{i}}+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(7)

where ${\mathrm{Y}}_{\mathrm{i}}$ represents the dependent variable; ${\mathrm{x}}_{\mathrm{i}}$ is the independent variable; ${\mathsf{\beta }}_0$ is a constant; and ${\mathsf{\beta }}_{\mathrm{i}}$ is the regression coefficient; and ${\mathsf{\epsilon }}_{\mathrm{i}}$ is the random error.

3.4.3. MGWR Model

The MGWR model is widely adopted when capturing spatial non-stationarity, as it accounts for the spatial heterogeneity associated with sample locations. Compared to the geographically weighted regression (GWR) model, MGWR enables the relationship between independent and dependent variables to change at different spatial scales [65,66], making it more suitable for revealing the spatial distribution of impact factors [67]. In this study, we applied MGWR2.2 software to perform the model, investigating the effects of the “3Cs” built environment factors on bike-sharing and metro trips in different geographic spaces. The specific formula is as follows:

$${\mathrm{y}}_{\mathrm{i}}={\sum }_{\mathrm{j}=1}^{\mathrm{k}}{\mathsf{\beta }}_{\mathrm{bwj}}({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}){\mathrm{x}}_{\mathrm{ij}}+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(8)

where in sample road segment i, ${\mathrm{y}}_{\mathrm{i}}$ represents the response variable. k is the number of road segments; ${\mathrm{x}}_{\mathrm{ij}}$ is the predictor variable j; (${\mathrm{u}}_{\mathrm{i}}$, v_i) is centroid coordinates of street segment i; ${\mathsf{\beta }}_{\mathrm{bwj}}$ is the bandwidth used in (${\mathrm{u}}_{\mathrm{i}}$, v_i); and ${\mathsf{\epsilon }}_{\mathrm{i}}$ is the random error.

4. Results

4.1. Statistics and Spatial Characteristics of Built Environment Variables

4.1.1. Metro Station Clustering Results

Based on the K-means results, metro stations in the study area can be categorized into 120 commercial-oriented stations, 28 residential-oriented stations, and 32 office-oriented stations (Figure 2), with distinct characteristics identified for each station type as follows (

Table 2. Z-scores for each cluster.

Clustering Factor	Cluster 1 (Commercial-Oriented Station)	Cluster 2 (Residential-Oriented Station)	Cluster 3 (Office-Oriented Station)
Interchange level	0.03	−0.43	0.28
Road length	0.01	−0.55	0.44
Location	0.25	−0.85	−0.19
Bus stop density	−0.28	1.72	−0.44
Population density	−0.031	1.67	−0.29
Employment density	−0.49	0.39	1.50
Shopping density	0.39	−0.16	−1.33
Residential density	0.43	−0.47	−1.20

).

Commercial-oriented stations are the most common type in the study area, mainly located in the southern part, including key interchange hubs such as Grand Theater Station and Bao’an Center Station, with high surrounding road density and strong transport connectivity. Surrounding these stations, POI categories such as commercial facilities (z-score: 0.40) and residential facilities (z-score: 0.42), which include both commercial and residential properties, exhibit comparatively higher z-score values.

Residential-oriented stations (e.g., Shanglilang Station and Gushu Station) are primarily located beyond the outer ring of the study area, characterized by fewer interchange stations, lower road density, and higher concentration of bus stops. The residential population (z-score: 1.67) is focused in these areas, with access flows during the morning peak. POI categories such as residential facilities (z-score: −0.46) and employment facilities (z-score: 0.39) exhibit comparatively higher values.

Office-oriented stations are mainly located around functional centers, close to commercial-oriented stations. These stations are characterized by higher interchange stations (e.g., Qianhaiwan Station and Futian Station), and dense road networks. Employment facilities (z-score: 1.50) are prominent in these regions, with large egress flows during the morning peak.

4.1.2. Spatial Characteristics of Bike-Sharing Volume and Built Environment Factors

Figure 3 presents the spatial distribution of bike-sharing–metro integrated usage across different station types. Spatially, high bike-sharing volume during the morning peak is concentrated in the Futian, Luohu, Nanshan, and Qianhai functional centers, primarily around commercial- and office-oriented stations. Furthermore, there are differences in bike-sharing usage around different types of metro stations. Specifically, during the study period, the average bike-sharing volume per road segment around office-oriented stations was 227 trips, followed closely by commercial-oriented stations with 221 trips, while residential-oriented stations had only 157 trips.

Figure 4, Figure 5 and Figure 6 show the spatial distribution of built environment variables under convenience, comfort, and caution categories (only some variables are shown; others are referred to in the Supplementary Materials). From the perspective of spatial distribution, most built environment variables exhibit significant spatial clustering.

Specifically, high population density and land-use mix are mainly distributed around the functional centers and the northeast of the study area, which closely resembles the spatial distribution of bike-sharing volume (Figure 4). Regarding comfort-related variables, while the spatial distribution of enclosure is relatively dispersed, areas with high sky openness are mainly located at the outer edges of the study area, where bike-sharing volume is low (Figure 5). For caution-related variables, high values of intersection density are similarly clustered around the functional centers, whereas the spatial distribution of traffic safety infrastructures does not show significant clustering (Figure 6).

4.2. Results of the OLS Model

The OLS regression model was constructed to examine the effect of the built environment variables on bike-sharing–metro integrated usage. Before employing the model, the variable inflation factors (VIF) were used to test multicollinearity among the independent variables. Generally, a VIF value of 10 or higher indicates the presence of severe multicollinearity among the variables. In this study, all VIF values were below 4, indicating no significant multicollinearity.

Table 3. Results of the OLS regression model.

	Commercial-Oriented Station		Residential-Oriented Station		Office-Oriented Station
	Coef.	p	Coef.	p	Coef.	p
Population density	0.156 ***	0.000	0.104 ***	0.000	0.216 ***	0.000
Residential density	−0.052 ***	0.010	−0.041	0.097	−0.063 **	0.012
Employment density	0.047 **	0.014	−0.033	0.177	−0.014	0.573
Catering density	−0.065 ***	0.003	−0.003	0.905	−0.088 ***	0.001
Shopping and entertainment density	−0.046 **	0.022	−0.016	0.565	−0.071 ***	0.004
Land-use mix	0.081 ***	0.004	0.088 **	0.011	0.125 ***	0.000
MHD	0.151 ***	0.000	0.103 ***	0.000	0.165 ***	0.000
Road density	0.133 ***	0.000	0.226 ***	0.000	0.145 ***	0.000
Sky openness	−0.027	0.155	0.013	0.573	−0.006	0.804
Greenery	−0.024	0.173	−0.039	0.081	−0.042	0.056
Enclosure	−0.036 **	0.050	−0.069 ***	0.004	−0.061 ***	0.008
Intersection density	0.084 ***	0.000	−0.012	0.597	0.045	0.063
Traffic safety infrastructure density	0.012	0.477	0.017	0.402	0.01	0.618

Note: *** and ** represent the significance levels of 1% and 5%, respectively.

displays the results of the OLS regression model. Results show that population density, land-use mix, MHD, and road density are significant in all types of stations, indicating that the convenience variables have a strong attraction for bike-sharing feeder use, while enclosure has a significant negative impact on all stations. Meanwhile, sky openness, greenery, and traffic safety infrastructure density are not significantly related to any station type. In addition, for commercial- and office-oriented stations, most convenience-related variables have a significant impact on bike-sharing usage around the station, with eight and seven significant variables, respectively. Most POI-related variables (residential, employment, and shopping and entertainment density) exhibit negative effects on bike-sharing usage, which deviates from our previous expectations. The number of convenience-related significant variables for residential-oriented stations is notably lower, with only four showing a significant effect.

4.3. Results of the MGWR Model

Based on the OLS results above, we then introduced the significant variables for each station type into the MGWR model along with the bike-sharing volume to further explore the spatial heterogeneity of the effects. Before introducing the model, a spatial autocorrelation analysis was conducted on the built environment factors. The results showed that all 13 independent variables had Moran’s I values greater than 0 and p-values less than 0.001, indicating a strong positive spatial correlation that meets the requirement for MGWR model construction.

The MGWR models of commercial-oriented stations, residential-oriented stations, and office-oriented stations are presented in Figure 7, Figure 8 and Figure 9, respectively. The standard deviation of regression coefficients reflects spatial heterogeneity, with larger values indicating stronger heterogeneity. Overall, the MGWR results reveal spatial heterogeneity in the effects of built environment variables around metro stations, with varying impacts across different areas and station types. As for convenience-related factors, population and road density show the highest spatial heterogeneity, while land-use mix and MHD exhibit minimal variation. Enclosure shows low heterogeneity across station types and intersection density significantly affects only commercial stations, with clear spatial variation. The results for each type of station are as follows:

Figure 7 shows the spatial distribution of regression coefficients for commercial-oriented stations. The effect of intersection density is the most significant and bidirectional, with coefficients ranging from −3.793 to 1.519, concentrated around Qianhai Center in the west and Luohu and Futian centers in the southeast. The effects of population density, catering density, and road density also show notable spatial heterogeneity.

Figure 8 shows the spatial distribution of regression coefficients for residential-oriented stations. Road density has the most significant spatial heterogeneity, with coefficients ranging from −0.949 to 3.556, and extreme values clustered around Qianhai Center in the western part of the study area. Population density follows with coefficients between −1.716 and 1.932, showing a spatial distribution similar to road density.

Figure 9 shows the spatial distribution of regression coefficients for office-oriented metro stations. Among all factors, road density exhibits the most significant spatial heterogeneity, with regression coefficients ranging from −0.987 to 3.446. The low and high values are mainly concentrated around functional centers. In contrast, residential density and enclosure show the weakest spatial heterogeneity, with coefficient ranges of −0.040 to −0.028 and −0.483 to 0.224, respectively. The spatial distribution of these coefficients demonstrates a “low-in-the-west, high-in-the-east” pattern.

5. Discussion

5.1. Impact of the Built Environment on the Integrated Use

As shown in Figure 10, among three categories of built environment factors, convenience-related variables have a greater impact on the integrated use of bike-sharing and metro, which is consistent with the earlier findings [9,68]. In high-population areas, strong demand for metro services increases the use of bike-sharing for “last-mile” travel. Contrary to previous expectations [7,60], most facility indicators, including catering, shopping and entertainment, and residential density have a negative impact on bike-sharing usage. On one hand, a high density of facilities may exacerbate road congestion during the morning peak hours, thereby reducing residents’ willingness to cycle. Another possible explanation is that, in this study, POI data were used to capture the distribution and attributes of facilities around metro stations. However, these data have limitations, as they do not capture the scale, quality, or operational status of facilities, causing low-attractiveness or inactive sites to be equally counted, which may misrepresent the true appeal of the surrounding environment. Finally, better spatial proximity and a well-connected street network enhance connectivity, providing cyclists with more route options.

Conversely, comfort and caution-related factors have a minimal impact on bike-sharing use for feeder trips. Enclosure exerts a significant negative impact on shared bike usage across all station types. While a certain level of enclosure may enhance the sense of safety and vitality [32], excessive enclosure during peak hours tends to increase the sense of crowding, thereby reducing the willingness to cycle. In contrast to some prior studies, most comfort-related factors, including sky openness and greenery, have no significant effect on the dependent variable [12,37]. Although the openness and greenery of streets can enhance the attractiveness of the streetscape and improve cyclists’ comfort and sense of well-being [33], during weekday morning peak hours, cycling is primarily for commuting purposes [69], with a greater emphasis on efficiency. Additionally, the bike-sharing data were collected during August, with an average temperature of approximately 29.5 °C. In such hot weather, the demand for open sky exposure among cyclists tends to be suppressed. Our study also revealed differences in the impact of built environment variables across metro station types. As supported by earlier studies [9,26], excessive residential, shopping and entertainment, and catering facilities reduce bike use around office-oriented and commercial-oriented stations during the morning peak, while having no significant impact on bike-sharing usage around residential-oriented stations. This may be attributed to differences in user travel behavior: commuters near office and commercial stations typically prioritize efficiency and time savings [17], as bike-sharing is often used for the “last mile” of their commute. Moreover, high residential density may intensify the pressure on cycling roads or cause a shortage of shared bikes, leading more people to choose walking or bus connections instead. This finding suggests that while mixed land uses along cycling routes can support access and vitality, excessive facility density may reduce travel efficiency. Finally, intersection density significantly enhances bike-sharing–metro integrated usage only around commercial-oriented stations. A possible explanation is that commercial areas generate diverse travel demands (e.g., commuting, early shopping, and deliveries), where a higher number of intersections provides more route options and better network connectivity.

5.2. Spatial Heterogeneity of the Impact of Built Environment on the Integrated Use

Figure 10 demonstrates the spatial heterogeneity of the impact of built environment variables on different types of stations. For commercial-oriented stations, intersection density exhibits the strongest spatial heterogeneity in its effect on the dependent variable. Although an earlier study demonstrated that intersections can attract more bike-sharing users [56], the MGWR result revealed spatial non-stationarity. Specifically, intersection density shows a strong positive effect around the four functional centers in the southern region, while in the northern areas, the promoting effect is relatively weak and may even become suppressive on some roads. The reason for this is that for high commercial density areas, intersections can improve roadway access and commuting efficiency; however, in lower traffic areas, excessive intersections can increase the length of unnecessary waiting time to cross the street.

For residential-oriented stations, road density exhibits the most noticeable spatial variation among all variables, with both significant positive and negative impacts occurring in areas with high road density (around the Qianhai functional center). These stations are located near densely populated communities with numerous schools, hospitals, and service facilities, resulting in strong demand for cycling-friendly environments. In areas with sparse road networks, higher road density can enhance connectivity and accessibility, while in traffic-congested zones, it may also lead to more intersections, resulting in longer waiting times and greater safety risks for cyclists [10]. Similarly, population density, land-use mix, and MHD show spatial non-stationarity, with the strongest positive and negative effects also clustered around the Qianhai functional center. This may be because high facility density draws large populations, resulting in congestion that simultaneously increases cycling demand and impedes its efficiency. Compared to other station types, the effect of enclosure on residential-oriented stations shows less spatial variation. For most residential stations, enclosure has a suppressive effect on bike-sharing usage, possibly indicating that commuters in residential areas place a higher value on cycling comfort and that highly enclosed streets may intensify the feeling of congestion, thus reducing their willingness to cycle.

For office-oriented stations, the results for population density, road density, and POI-related variables are similar to those for commercial-oriented stations. However, it is interesting to note that, the effect of land-use mix on feeder trips varies spatially among office-oriented stations, which is inconsistent with previous studies [26,70]. Specifically, this factor shows a significant positive effect around the functional centers of Futian, Luohu, Nanshan, and Qianhai, but exhibits a negative impact in the northern areas with relatively low bike-sharing volume. For core regions, a high functional mix allows residents’ living, working, and service needs to be met, reducing commuting distances and attracting more commuters to ride to their workplaces. In contrast, in non-core areas, a high functional mix may not generate strong biking demand due to insufficient infrastructure, such as limited bike-sharing stations and poor cycling roads [49], preventing POIs from converting into cycling needs.

6. Conclusions

This study explored how the built environment affects bike-sharing–metro integration during morning peak hours and the spatial heterogeneity of these influences, by employing the OLS regression model and the MGWR model. Particular emphasis has been given to exploring this relationship with a focus on various metro station types that are clustered based on transportation characteristics, population density, and land use. The key findings of the study are summarized as follows: (1) Bike-sharing as a metro feeder mode shows significant spatial agglomeration in Shenzhen’s urban core area. There are notable differences in bike-sharing volume across different station types. Specifically, during the morning peak period, bike-sharing volume is significantly higher around office-oriented and commercial-oriented stations compared to residential-oriented stations. (2) The spatial distribution of built environment variables is also closely related to metro station types and shows a high degree of correlation with the spatial pattern of bike-sharing usage. Convenience-related variables, such as population density, POI-related factors, and land-use mix, are concentrated around commercial- and office-oriented stations, where bike volume is high. Comfort-related variables are mainly centered around residential stations, while caution-related variables are more dispersed. (3) The built environment around different types of metro stations has different impacts on cycling behavior. Population density, road density, MHD, and functional mixing show significant positive effects for all types of stations, while enclosure generally inhibits feeder behavior. Unlike expectations, POI-related variables do not promote feeder rides around all types of stations and most comfort-related factors have a weak effect on morning peak bike-sharing feeder use. (4) The influence of the built environment on bike-sharing usage exhibits significant spatial heterogeneity. Specifically, population density and road density exert strong spatial non-stationarity across all types of metro stations, showing both positive and negative effects in different areas, while enclosure consistently has a negative effect in most regions. The areas with significant impacts are generally concentrated around the four functional centers mentioned above.

This study also presents categorized planning recommendations for various types of metro stations, emphasizing differentiated strategies for central areas and low bike-sharing volume areas. For commercial-oriented stations, enhancing the functional mix while ensuring appropriate facility density is essential. Commuter-focused amenities, such as convenience stores and restaurants, can be added without overcrowding the street. Additionally, intersection and road management in core zones is crucial for maintaining traffic order, along with ensuring an adequate supply of shared bicycles. For residential-oriented stations, planning efforts can focus on improving road density and connectivity. Furthermore, attention should also be paid to the psychological comfort of cyclists in the areas surrounding residential neighborhoods by reducing the sense of enclosure created by walls, buildings, and fences and enhancing micro-level street design to support a more comfortable cycling environment. Finally, as for office-oriented stations, it is recommended to improve traffic organization in areas with dense road networks, plan surrounding dining, shopping, and entertainment facilities more rationally, and enhance road connectivity in northern areas to improve the efficiency of last-mile cycling.

It is important to acknowledge several limitations of this study. First, certain variables within the “3Cs” framework may have been overlooked. Future research is encouraged to incorporate factors such as competitiveness and commuters’ socioeconomic characteristics. Furthermore, greater emphasis should be placed on psychological indicators, including cyclists’ perceived comfort, security, and aesthetic appreciation of the cycling environment. Second, this study used POI data to analyze the distribution of catering, shopping and entertainment, employment and residential facilities as well as land-use mix around metro stations. However, POI data are static and lack information on temporal changes, facility size, or quality, providing only a partial view of the built environment’s influence. Future research could incorporate more detailed and dynamic datasets, including land use, qualitative factors, and environmental audits, to better capture the characteristics of the facilities. Finally, in terms of methodology, this study primarily analyzes the linear relationships between the independent and dependent variables, as well as the spatial heterogeneity of these effects. The exploration of potential nonlinear relationships remains an area for further investigation in future research.