Correlation between Land Use Pattern and Urban Rail Ridership Based on Bicycle-Sharing Trajectory

Abstract

As a form of rapid mass transportation, urban rail systems have always been widely used to alleviate urban traffic congestion and reconstruct urban structures. Land use characteristics are indispensable to this system and correlate with urban ridership. Dock-less bicycle-sharing expands the station service coverage range because it integrates public transportation with an urban rail system to create a convenient travel model. Consequently, the land use pattern with dock-less bicycle-sharing is associated with urban rail ridership. This paper measures the correlation between land use and urban rail ridership based on the trajectory of dock-less bicycle-sharing, which precisely reflects the travel behavior of passengers along the trip chain. The specific relationship has been determined using the random forest model. This paper found that the land use pattern could better explain the egress ridership during morning peak hours. In particular, it could explain 48.46% of the urban rail ridership in terms of egress, but the explicability for the ingress ridership slightly decreased to 36.88%. This suggests that the land use pattern is related to urban rail ridership. However, the impact situation varies, so we should understand this relationship with greater care.

1. Introduction

As a rapid mass transport system, urban rail is used worldwide and has become the basis of the daily lives of city dwellers [1]. It is expected to meet the majority of travel demands [2] and alleviate urban traffic congestion because it provides a high-capacity, dependable, safe, and efficient mode of transportation [3]. The ridership levels of the system might bear relation to the urban rail system’s operating performance, and ample evidence has shown that land use characteristics correlate with urban rail ridership [4,5,6,7].

Regarding land use patterns within station catchment areas, we could understand how land use characteristics of density and diversity affect urban rail ridership and even attempt to decipher the effect of specific land use patterns. However, most studies identify this correlation based on land use spatial structure, and few studies can quantitatively measure land use characteristics based on travel behaviors or demand data [8]. Individual travel routes represent the land use patterns of passenger departure/arrival locations. They provide more detail to identify urban rail ridership and land use patterns under tangibly defined trip chains.

The systems of dock-less bicycle-sharing developed in recent years could solve the “first–last mile” in urban transportation and effectively extend the hub transport service radius [8]. The concept significantly enhances convenience due to not having a fixed pile. In 2019, in China, there were almost 300 million registered users and 19.5 million shared bikes covering 360 cities across the country [9]. In addition, bicycle-sharing is an effective transit mode within the metro system, one which directly connects to land use and further assists in determining the relationship between urban rail ridership and land use based on travel behavior and demand.

The dock-less bicycle-sharing trajectory data from between the urban rail station and the riders’ destinations (and vice versa) contributes to our accuracy in identifying land use characteristics depending on the spatial distribution features of people who travel by bicycle [8]. This paper recognizes the correlation between land use patterns and urban rail ridership based on origin–destination pairs. In this sense, this paper will solve the following questions: (1) What is the land use pattern in bicycle-sharing origin—destination areas? (2) What is the relationship between land use pattern and urban rail ridership? (3) Does this correlation differ depending on whether the trip is to an access station or an egress station? According to this study, the relationship between land use and urban rail ridership can be seen in the way people use bicycle-sharing to travel.

2. Literature Review

Land Use and Transport Strategies (ILUT) reflect efforts to curb the growth of car-oriented travel by creating spatial patterns and environments. The corresponding land use management strategies are known as New Urbanism, Smart Growth (in the USA), or Transit-Oriented Development (TOD) [10]. Land development patterns in urban areas can substantially affect travel demand, further impacting transport model usage [11]. In Seoul, land use mix diversity positively relates to subway ridership but has different spatial ranges [7]. In Kuala Lumpur, commercial, workplace, and public service facilities affect urban rail ridership in many ways [5].

Indeed, many people devote time to identifying the correlations between the built environment and either the urban rail system/bicycle-sharing system ridership or the performance of an integrated urban rail and bicycle-sharing system [1,12,13,14,15]. Nevertheless, few identify this relationship with the dock-less bicycle-sharing trajectory, in which it is an intermediate mode to connect urban rail stations and land use patterns. This would give us a thorough understanding of travel patterns and the journey chain between the urban rail system and bicycle-sharing behavior.

Bicycle-sharing is a critical component of sustainable transportation [16], a model that is flexible, low-cost, environmentally friendly, and one that has gradually been adopted by people for their daily commutes. Even in small areas, cycling as an entry/exit mode to/from the metro system improves transport accessibility within urban rail station catchment areas [17]. The land use characteristics within the catchment areas of urban rail stations, such as residential and workplace density, will positively influence the density of bicycle–metro transfer cycling trips [1].

Land use characteristics associated with bicycle-sharing ridership influence the arrival and departure ratios of the bicycle-sharing system [15]. However, this situation is not static; for example, the odds of cycling in Hong Kong were negatively correlated with the population density/number of bus stops [18]. The bicycle-sharing trip chain may also reflect the land use characteristics, with the most reliable bicycle-sharing trip chain for identifying residence compared with regional work and consumption [8]. In addition, the demand for weekday and weekend trips is unbalanced due to the variety of trip purposes [17].

Due to the technology involved in bicycle-sharing, data encompassing travel behavior includes, but is not limited to, factors such as boarding/alighting time, origin/destination place, and riding time. This information could be utilized to comprehend the land use characteristics and to influence urban rail ridership based on the trajectory of bicycle-sharing. Extremely high population density leads to a decline in bicycle-sharing–metro integrated use [19], and researchers found that the proportion of commercial and green space land use should be reasonable and that too high or too low a proportion does not benefit such multi-transport model usage.

According to the trajectory of dock-less bicycle-sharing, we could identify the precise location for transfers and measure the land use pattern of the location to reflect the relationship between land use pattern and urban rail ridership. Cycling trajectory connects the origin and destination locations that tangibly represent the land use pattern of the area, and bicycle flows were expected to decrease as we move further from the central business district (CBD) [15]. The trip chain of integrated urban rail and bicycle-sharing is an innovative method for determining the relationship between land use and urban rail station ridership. In comparison to identifying this relationship solely based on the spatial distribution characteristics of land use, bicycle-sharing affords greater precision.

Under cycling trajectory, the relationship between land use and urban rail ridership will fluctuate slightly. Commuters will switch from alternative modes of transportation during the morning and evening rush hours, and the new urban rail line will prompt more people who use automobiles to switch to public transportation [20]. During the morning rush hour, a metro station that serves residential areas may dispatch passengers, while the afternoon rush hour may attract a large number of passengers, whereas the opposite of this situation is possible in the commercial and business sector [21]. Improving bicycle pathways will significantly increase integrated use of the metro, but the effect will be more pronounced in suburban areas for morning feeder trips than in urban areas [22].

Indeed, spatiotemporal variation will result in modification of travel behavior. The analysis period for this paper is from 6:00 a.m. to 10:00 a.m., to demonstrate how land use pattern affects urban rail ridership in the morning based on a dock-less bicycle-sharing trajectory. This study will also attempt to determine the relevant behavior if there is a disparity between access/exit urban rail station travel behavior and land use pattern. This study will also identify the triangular relationship between land use pattern, urban rail ridership, and dock-less bicycle-sharing in the morning commute scenario.

3. Methodology

3.1. Study Area and Data

Xiamen, situated on the southeastern coast of China (Figure 1), is an island city with a population of 4.29 million (2019) with land area of 1700 km²; however, the majority of its urban area is located on the mainland (Xiamen Island) (Figure 2). Until 2021, the three metro lines had a combined length of 98.4 km (Figure 2) and an average daily passenger volume of about 314,000 people (data from “Xiamen transport development annual report 2020”). This annual report revealed that the likelihood of cycling on the mainland was significantly greater than in other areas of Xiamen. The slow modes of transportation, that is, walking/public transportation/bicycle and sharing bicycles and private vehicles, comprised 44.7% and 20.1% of daily trips, respectively. On the mainland, the proportion of slow transportation is 76%.

In 2018, using 10 m satellite images, OpenStreetMap, nighttime lights, POIs, and Tencent social big data as input features [23], land use data was used to create a new urban land use map for the whole of China. Adapted from the Chinese Standard of Land Use Classification are the Essential Urban Land Use Categories. This paper adopts the point of interest (POI) as additional data to reflect the land use pattern if the land use data are not covered. Figure 3 depicts the land use pattern on the mainland of Xiamen.

Due to data limitations, we were only able to collect ridership information for four urban rail stations on the mainland (Figure 4). Consequently, this paper selects them as specific analysis stations to establish this association during the morning rush hour (6:00 a.m.–10:00 a.m.). These four stations are located in the geographic center of Xiamen mainland (Wushipu, Lvcuo, Tangbian, and Jiangtou stations). Lvcuo is a transfer station; the remaining stations are on Lines 1 and 2 (Figure 4). Figure 4 depicts the point of interest (POI) kernel density distribution characteristics. The POI category includes residential, business office, healthcare facilities, grocery, restaurant, and leisure facilities. We could identify the city’s hot-spot area based on these categories of POI because they constitute the city’s daily functions. Three of the stations (Wushipu, Lvcuo, Jiangtou) are located in bustling areas, while the remaining station (Tangbian) is located on the central periphery. Due to the multiple characteristics of the built environment surrounding the urban rail station, the study of these four stations could yield objective results.

This study utilized dock-less bicycle-sharing ridership data and urban rail station temporary ridership data from the “Xiamen big data security open innovation applying competition”. Several million points of cycling location recording data and urban rail station ridership data were collected between 21 and 25 December 2020.

The distribution of ridership per station is depicted in Figure 5. This is the cumulative time-ridership for these five days, which represents the simultaneous ridership. It indicates that the peak ridership occurs between 8:00 and 9:00 a.m. Except for egress ridership between 6:00 and 7:00 a.m., Lvcuo is the station with the highest ridership. Between 7:00 and 8:00 a.m. and from 8:00 to 9:00 a.m., the ingress ridership at Jiangtou and Tangbian is greater than the egress ridership, and vice versa at the other stops. In addition, Lvcuo and Wushipu have higher outbound ridership than inbound ridership at the same time, except between 6:00 and 7:00 a.m.

3.2. Method

This paper examines the relationship between urban rail ridership and land use patterns by using the random forest model based on the temporary passenger flow and bicycle-sharing trajectory of four metro stations. The random forest model is popular for classifying and regressing multidimensional phenomena with high precision and a low likelihood of overfitting. In a variety of application scenarios, RF outperforms other prevalent machine learning models [24].

A random forest model investigates more nuanced associations between the outcome and explanatory variables [25], while nonlinear relationships, multimodal data, categorical and numerical features, missing values, and tolerance of random variables can all be handled by an RF model [26,27], and it has good accuracy [27,28]. Wen et al. used high-density traffic monitoring data and land use data to train random forest models capable of accurately predicting dynamic, link-level vehicle emissions [24]. Cheng et al. examined the effects of the built environment on elderly people’s walking habits using random forest and found that land use mix, to some extent, only increases older adults’ walking [25]. In addition, they also found that the random forest model outperformed linear regression.

Cheng et al. analyzed travel-mode choices using a random forest model and reported that it performed significantly better in predicting travel mode-choice, with higher accuracy and lower computational cost [29]. This study seeks to determine the relationship between urban rail ridership and land use patterns based on cycling trajectory, as demonstrated by the random forest model in Figure 6. Regarding this process, this study established the random forest modeling as shown in Figure 7. The land use pattern consists of Residential, Business office, Commercial service, Industrial, Administrative, Education, Sports and culture, and Park and green space; urban rail ridership is divided into ingress and egress ridership, and dock-less bicycle-sharing is the link between land use pattern and urban rail ridership.

Each tree is bootstrapped with different observations [29], and the candidate variables for each tree are selected at random from the entire range of variables. Bootstrap aggregating is a variant of the random forest model that considers every variable a candidate-splitting variable, thereby reducing bias. This study identifies the relationship between urban rail ridership and land use patterns based on cycling trajectory through bootstrap aggregating, given this context.

This study divided the dataset into test set and train set to prevent possible model overfitting [24], as the test set should not be used for model development. Therefore, 30% of the dataset was selected randomly as the test set, and the remaining data were the train set. This study first establishes the mode to determine or predict urban rail ridership from land use pattern. Then, it determines whether the variable importance due to OOB estimates of all deciduous trees in the forest can be averaged to obtain the generalized error estimation of the RF model [27]. This acknowledges the importance of land use pattern for urban rail ridership.

This study examined the relationship between urban rail station ridership and egress and ingress land use patterns. Ridership at urban rail station exits (entrances) is the dependent variable, while land use pattern is the independent variable. However, because only a small fraction of the commuters uses bicycle-sharing to (or from) medical, transit station, park, and green space areas, the rest of the land use pattern is chosen as the independent variable to measure the correlation between urban rail ridership and land use pattern in order to improve the effectiveness of modeling (

Table 1. Dependent and independent variables.

Dependent Variable	Independent Variable
Urban rail egress (ingress) ridership	Residential, Business office, Commercial service, Industrial, Administrative, Educational

).

Table 2. Urban rail station egress ridership distribution with land use pattern.

Station	Time	Ridership	Residential	Business Office	Commercial Service	Industrial	Administrative	Educational
Wushipu	6:00 to 7:00 a.m.	140	4	0	1	1	0	0
Wushipu	7 to 8 a.m.	944	16	5	7	2	0	0
Wushipu	8:00 to 9:00 a.m.	1164	33	15	12	1	0	1
Wushipu	9:00 to 10:00 a.m.	769	13	5	6	0	0	0
Lvcuo	6:00 to 7:00 a.m.	59	16	2	2	0	2	0
Lvcuo	7:00 to 8:00 a.m.	791	22	4	6	0	2	0
Lvcuo	8:00 to 9:00 a.m.	1963	42	18	18	4	0	0
Lvcuo	9:00 to 10:00 a.m.	1055	30	20	16	0	0	0
Tangbian	6:00 to 7:00 a.m.	33	1	0	0	0	0	0
Tangbian	7:00 to 8:00 a.m.	385	8	2	3	0	0	1
Tangbian	8:00 to 9:00 a.m.	616	9	7	4	4	0	1
Tangbian	9:00 to 10:00 a.m.	324	3	4	0	0	0	0
Jiangtou	6:00 to 7:00 a.m.	14	3	1	2	0	0	0
Jiangtou	7:00 to 8:00 a.m.	250	8	3	3	1	0	0
Jiangtou	8:00 to 9:00 a.m.	338	13	5	8	7	0	0
Jiangtou	9:00 to 10:00 a.m.	236	8	3	4	2	1	0

and

Table 3. Urban rail station ingress ridership distribution with land use pattern.

Station	Time	Ridership	Residential	Business Office	Commercial Service	Industrial	Administrative	Educational
Wushipu	6:00 to 7:00 a.m.	181	9	0	1	0	0	0
Wushipu	7:00 to 8:00 a.m.	524	34	5	1	0	0	1
Wushipu	8:00 to 9:00 a.m.	653	33	6	6	3	0	1
Wushipu	9:00 to 10:00 a.m.	299	33	7	3	4	0	0
Lvcuo	6:00 to 7:00 a.m.	290	18	2	2	0	0	0
Lvcuo	7:00 to 8:00 a.m.	648	52	20	6	0	2	2
Lvcuo	8:00 to 9:00 a.m.	737	100	30	30	10	4	4
Lvcuo	9:00 to 10:00 a.m.	437	30	8	10	4	0	0
Tangbian	6 to 7 a.m.	180	1	1	1	0	0	0
Tangbian	7:00 to 8:00 a.m.	444	6	1	0	0	0	0
Tangbian	8:00 to 9:00 a.m.	550	5	0	1	0	0	0
Tangbian	9:00 to 10:00 a.m.	257	2	2	1	0	0	0
Jiangtou	6:00 to 7:00 a.m.	164	2	1	0	1	1	0
Jiangtou	7:00 to 8:00 a.m.	366	22	2	6	4	1	0
Jiangtou	8:00 to 9:00 a.m.	653	26	6	5	2	3	0
Jiangtou	9:00 to 10:00 a.m.	209	5	0	3	1	0	0

display a portion of the five-day dataset that this study analyzed, and Appendix A provides additional information regarding the trip chain between the urban rail station and bicycle-sharing.

3.2.1. Bootstrap Aggregating Calculation

The first step is resampling with replacement, and Equation (1) is as follows:

$${\left\{x_i^{{\ast }^b},y_i^{{\ast }^b}\right\}}_{i=1}^n,b=1,\cdots ,B$$

(1)

where $\ast$ represents the bootstrap sample, and $n$ is per sample capacity.

The second step is evaluating the decision tree without trimming based on bootstrap sample B (${\left\{x_i^{{\ast }^b},y_i^{{\ast }^b}\right\}}_{i=1}^B$). The predicted result is as Equation (2):

$$\widehat{f_{bag}}\left(x\right)=\frac{1}{ b}\sum _{b=1}^B\widehat{f^{\ast b}}\left(x\right)$$

(2)

The third step is a majority vote, and the highest category will win, as in Equation (3):

$$\widehat{f_{bag}}\left(x\right)=argmax\sum _{y\in \left\{1,\cdots ,K\right\} b=1}^{B}I\left[y=\widehat{f^{\ast b}}\left(x\right)\right]$$

(3)

where hypothesis $y\in \left\{1,\cdots ,K\right\}$, there are K categories; I$\left(•\right)$ is the indicator function that determines whether $y=\widehat{f^{\ast b}}\left(x\right)$: if yes, then record 1, otherwise, record 0. Regarding this, many weaker classifiers can be combined to produce a powerful committee.

3.2.2. Regarding Bootstrap Aggregation

This study found that the selected land use pattern could explain 48.46% of the egress ridership and 36.88% of the ingress ridership at 6:00–10:00 a.m. This means that the impact is in the disparity of different types of travel behavior.

3.2.3. Out-of-Bag Observation Calculation

Out-of-bag observation is used to assess test errors. The OOB MSE (out-of-bag indicates mean squared error) is as shown in Equation (4):

$$MS{E}_{OOB}\equiv \frac{1}{n}\sum _{i=1}^n{\left(\widehat{y_{i,OOB}}\ne y_i\right)}^2$$

(4)

where $\widehat{y_{i,OOB}}$ is the OOB prediction for the $i\mathrm{th}$ observation, and $y_i$ is the actual observation.

By using the OOB observation calculation, the egress and ingress random forest regression of $MS{E}_{OOB}$ are found to be 94,721.31 and 24,285.72, respectively. This indicates that predicting urban rail station egress ridership has a smaller error compared to ingress ridership.

3.2.4. Measuring Variable Importance

The random forest model involves several decision trees that could not reflect how each dependent variable’s importance differs from a single decision tree. Hence, it should measure variable importance per variable, presenting marginal effects of every variable. Thus, the partial dependence plot method was applied:

$$\widehat{\varphi }\left(x_1,x_2\right)\equiv \frac{1}{n}\sum _{i=1}^{n}f\left(x_1,x_2,x_{i3}\cdots ,x_{ip}\right)$$

(5)

Table 4. Variable importance measuring results.

Items	Variable Importance Index		Land Use Pattern
	% IncMSE	IncNodePurity
Ingress	14.62	904,731.42	Residential
	13.36	573,238.60	Business office
	−0.51	248,966.18	Commercial service
	2.72	85,867.69	Industrial
	−3.40	42,504.33	Administrative
	8.28	125,132.17	Educational
Egress	16.13	2,997,219.70	Residential
	31.79	11,841,480.70	Business office
	−0.49	2,071,268.70	Commercial service
	1.54	581,978.20	Industrial
	−1.82	468,586	Administrative
	19.84	1,832,630.90	Educational

displays the variable importance measuring results; IncNodePurity represents the residual sum of squares that measures the total increase in the homogeneity of the data samples by splitting them for a given variable. IncMSE is the mean squared error, which measures the effect on the predictive power when the value of a specific original variable is randomly permuted [30]. Indeed, these two coefficients also represent variable importance, and this paper mainly considers the IncMSE (mean squared error) as the assessment standard. Hence, the most important land use is residential and business office for ingress and egress ridership, respectively (Table 4).

4. Origin-Destination Pair Identification

The significance of identifying the trip from the urban rail station to the destination via bicycle-sharing is highlighted in this study. The dock-less bicycle-sharing trajectory dataset recorded the cyclists’ locations. Therefore, each day, there are more than two million points of bicycle-sharing data. There is also a unique user ID for identifying the cyclist’s route, which helps us draw the cycle route for the totality. Only one point linking time-series connections is required to complete the cycling route.

In addition, according to Guo et al., the 100 m buffer of metro entrances are the most popular parking area around metro stations [31,32,33]. This study concluded that cycling trips with an origin or destination within 100 m of the entrance to an urban rail station are properly considered bicycle–metro transferring trips.

Figure 8a,b displays the trip chain of metro and bicycle-sharing from 6:00 to 7:00 a.m. to tangibly illustrate the cycling database that selects a one-day (21 December 2020) cycling trajectory as a sample to reflect more details. Four colors represent distinct analysis stations and correspond to cycling trajectories that pass the exit or entrance of urban rail stations. Regarding the trip chain, we were able to accurately identify the land use pattern of transitory visitors leaving or entering the area. This provides a novel method for understanding the relationship between urban rail stations and land use patterns based on the actual trip chain. Tu et al. identified the spatial variations that influence public ridership for bus, metro, and taxi via vehicle (bus and taxi) GPS trajectories and smart card data [33]. They found that employment and mixed land use have bilateral effects on bus and metro ridership levels but positive effects on taxi ridership.

Figure 8a,b depicts the transitory exit and entrance situation at an urban rail station. From 6:00 to 7:00 a.m., twenty-one people (approximately 67.7%) arrived from residential areas, followed by three people from business offices and commercial services (approximately 9.7%). Around sixteen (approximately 64%) transit passengers used bicycle-sharing to travel home from the urban rail station, leaving four individuals from commercial services.

Between 7:00 and 8:00 a.m., as shown in Figure 9a,b, the number of riders increased to 76 egress and 128 ingress travelers. Approximately 56% of individuals go to residential areas, followed by commercial services (21.1%) and business offices (15.8%). In contrast, 87 people (67.9%) ride to the station from residential areas, followed by 18 (14%) and 10 (7.8%) from business offices and commercial services, respectively.

There are more cycling trajectories in Figure 10a,b between 8:00 a.m. and 9:00 a.m., when bicycle-sharing is most popular. During this time, 162 people practiced bicycle-sharing from the urban rail station, with the majority going to residential, business office, and commercial service areas, with 76, 36, and 32 riders, respectively. Two individuals attended an educational venue, while one attended a sports and cultural venue. In the opposite direction, the transitory population was primarily residential (one hundred fourteen people), business office (twenty-seven people), and commercial service (twenty-seven people). The rest consisted of industrial (eleven people), administrative (five people), educational (three people), and sports and cultural (two people).

Between 9:00 and 10:00 a.m., the number of cyclists declined compared to the preceding hour, as depicted in Figure 11a,b. Around 81 people (94%) of the egress ridership were transit riders who use bicycle-sharing to travel to residential, business office, or commercial destinations. In the opposite direction, approximately 90.1% originated from these three land use patterns, totaling 85. In addition, seven riders and one rider come from industrial and sports and culture sectors, respectively.

Using this sample of origin–destination pairs, we were able to determine the land use pattern characteristics of each traveler’s boarding or alighting location. Analyzing the correlation between urban rail ridership and land use pattern based on individual travel behavior, this paper required five days of bicycle-sharing data. This paper also applied point of interest (POI) data to identify the land use pattern for the area where land use data are missing.

5. Correlation Analysis between Land Use Ridership

5.1. Results of Trip Chain

Regarding the origin—destination pair,

Table 5. Urban rail and bicycle-sharing ridership.

Date	Items	Ridership	Number of Cyclists	Cycling Ratio	Average Cycling Ratio
21 December 2020	Ingress	6592	346	5.25%	5.33%
22 December 2020		6753	409	6.06%
23 December 2020		6621	139	2.10%
24 December 2020		6670	493	7.39%
25 December 2020		6751	394	5.84%
21 December 2020	Egress	9120	435	4.77%	4.46%
22 December 2020		9133	459	5.03%
23 December 2020		9095	134	1.47%
24 December 2020		9216	553	6.00%
25 December 2020		9210	465	5.05%

details the ridership of urban rail and dock-less bicycle-sharing. During the analysis period from 6:00 to 10:00 a.m., the number of outbound passengers exceeds the number of entrants. However, the average cycling ratio of ingress ridership is approximately one percent higher than that of egress ridership (4.46%).

According to

Table 6. Travel behavior of transit by bicycle-sharing.

Land Use Pattern	Number of Cyclists (Ingress)	Number of Cyclists (Egress)
Residential	955	1308
Business office	362	260
Commercial service	261	271
Industrial	119	71
Sports and cultural	38	59
Educational	11	48
Administrative	26	24
Medical	6	0
Transportation station	2	4
Park and green space	1	1

, the travel habits of individuals who utilize bicycle-sharing are as follows. The majority of bicycle-sharing trips depart (or arrive) at residential locations, followed by business office and commercial service locations. Occasionally, a few people travel to or from medical, transportation station, and park and green space land use patterns, which are ranked as the bottom three based on bicycle-sharing ridership. Appendix A provides more details regarding the trip chain between the urban rail station and bicycle-sharing.

The built environment of land use patterns is one of the factors that influences dock-less bicycle-sharing usage, along with weather [34,35], road infrastructure [12,36], personal characteristics [37], and so on. Figure 12 depicts the fluctuation in the goal cycling ratio on two different days, the 23rd and 24th, which have the lowest and highest cycling ratio. While the ratio of cycling to land use pattern is nearly stable (Figure 13), the percentage of cycling to (or from) residential, business office, and commercial service areas is always in the top three during the analysis period. This indicates that the fluctuation in the proportion of cyclists at urban rail stations is influenced by other factors.

5.2. Results of Correlation Analysis

This paper selected six types of land use pattern as the variable to measure the correlation between urban rail ridership and land use pattern in an attempt to improve the efficacy of modeling, with the variable-explained value nearly doubling when all land use patterns were considered. Urban rail ridership is significantly influenced by land use patterns, which explains why egress ability outperforms ingress. Regarding urban rail station egress ridership, the land use pattern could explain 48.46% of egress ridership during the 6:00–10:00 a.m. period, while the land use pattern could account for 36.88% of ingress ridership during the same period. As shown in

Table 7. The coefficient with random forest result.

Items	% Var. Explained	% IncMSE	IncNodePurity	Land Use Pattern
Ingress	36.88	14.62	904,731.42	Residential
		13.36	573,238.60	Business office
		−0.51	248,966.18	Commercial service
		2.72	85,867.69	Industrial
		−3.40	42,504.33	Administrative
		8.28	125,132.17	Educational
Egress	48.46	16.13	2,997,219.70	Residential
		31.79	11,841,480.70	Business office
		−0.49	2,071,268.70	Commercial service
		1.54	581,978.20	Industrial
		−1.82	468,586	Administrative
		19.84	1,832,630.90	Educational

, the variables have differing effects.

The out-of-bag (OOB) errors (Figure 14) stabilize after the number of decision trees exceeds 200, indicating that the OOB errors will not increase regardless of the decision tree size. This will not result in reducing or overfitting of OOB errors.

Figure 15 illustrates the varying significance of urban rail ridership and land use pattern. Residential and Business office are the most significant variables for ingress ridership, regardless of IncMSE (mean squared error) or IncNodePurity (residual sum of squares) (Figure 15a–d). The variable with the greatest effect on egress ridership is business office, followed by educational or residential land use, depending on the IncMSE coefficient or IncNodePurity coefficient, respectively. However, this paper focuses primarily on the IncMSE (mean squared error) as the criterion for evaluation 2 × 10⁴ × 10⁶.

The variable importance plot (Figure 15) simply ranks the variable below a level of significance. This paper measures the marginal effects of residential and business office variables using a partial dependence plot to provide more information about these variables and how they affect urban ridership. Typically, the residential impacting capacity increases gradually, with station ridership increasing for the ingress aspect (Figure 16). It is mitigated when station ingress ridership exceeds 500 per hour, with similar effects for station ingress and egress ridership. However, when station egress ridership exceeds 600 per hour, the residential impact for egress ridership suddenly decreases, and after that, the impacting capacity almost stabilizes.

After station ingress ridership reaches 480 per hour and station egress ridership reaches 1200 per hour, the business office variable’s ability to influence ingress and egress ridership gradually stabilizes and improves (Figure 17).

5.3. Discussion of Results

This study concludes that the land use pattern surrounding urban rail stations has a significant relationship with urban rail ridership. The degree of influence shows a disparity between access station and egress station, with egress having a higher correlation than ingress. Residential land use has a greater impact on access station ridership because most passengers leave their homes in the morning (between 6:00 and 10:00 a.m.). This corresponds with the normal morning motivation to go outside every day. In terms of egress ridership, however, the land use of business office is the most significant, as the majority of people who require dock-less bicycle-sharing transfers will cycle to the office.

During the morning rush hour, the majority of people leave their homes for work, while students head to school. These commuters comprise primarily employees and students, and they are sensitive to trip duration and overall cost. As a result, they favor cost-effective methods such as bicycle-sharing and urban rail systems to complete their trips. In the area of the case study, bicycle-sharing costs RMB 1.5 per fifteen minutes, urban rail costs RMB 2 for the first four kilometers, and the average ticket price per kilometer is RMB 0.5. Students and adults who register smart cards receive a discount of 90% and 50%, respectively. The integration system and discount promotion in Xiamen appear to be successful, with residential areas having the greatest impact on ingress ridership. Regarding egress ridership, business and education areas are the top two.

The distance and number of cyclists between the urban rail station and the primary land use pattern are depicted in Figure 18. During the study period, the average distance from a commercial service area to an urban rail station via bicycle-sharing was 2.24 km. This distance represents the maximum distance traveled by bicycle-sharing to reach an urban rail station from a commercial service area. In contrast, the shortest distance is approximately 1.36 km between the urban rail station and residential areas. The trajectory of bicycle-sharing contributes to understanding the correlation between land use pattern and travel behavior, which applies to urban rail stations.

Regarding this finding, we could speculate that commercial land use is farther than others because passengers must travel further to reach it. That does not strictly conform to the TOD concept that the core circle of land surrounding a station should be commercial [38].

In addition, this paper measured this relationship based on the trajectory of bicycle-sharing. It has demonstrated that travel demand alone cannot establish the relationship between urban rail ridership and land use patterns. It found that the average ratio of cyclists entering a station (entrance) is higher than the ratio of cyclists leaving a station (exit), but the land use pattern would suggest more urban rail exit ridership. Consequently, simply using travel demand or traffic mobility to reflect how land use influences urban rail ridership could result in bias.

Bicycle-sharing provides a practical and cost-effective means to connect the place of origin to the place of destination (urban rail station), which offers novel ways to more accurately measure the impact between land use patterns and urban rail ridership. This also reflects the importance of considering travel behavior when identifying this relationship behavior. The structure of urban space influences the transport system, and vice versa. Similar to Tobler’s First Law of Geography (everything is related to everything else, but near things are more connected than distant things), individual travel behavior may form a disparate transport phenomenon due to personal variation. The spatial heterogeneity of built environments will result in a different impact on urban rail ridership, so that travel behavior may have the same outcome.

Developing a livable city that offers more transportation options is one of the most important aspects of the livability principle of sustainable communities in urban areas affected by urban traffic congestion and rising pollution levels [39]. The integration of public transportation will provide convenient access corridors to residents who live close to the station. Accessibility will entice more individuals to use public transport and shift their travel behavior from automobility to public transportation. Urban rail and bicycle-sharing system ridership is influenced by the land use characteristics surrounding the urban rail station. It establishes a balanced node-place development model that contributes to the success of TOD projects and the development of a sustainable city.

6. Conclusions and Implications

Increasing urban rail ridership has the advantage of optimizing its operating performance and could increase the public transport system’s share of all travelers. Numerous individuals from various regions have demonstrated that the land use pattern within the station catchment area will impact ridership. However, there have been few efforts to examine this relationship based on travel behavior. Dock-less bicycle-sharing as a flexible and inexpensive mode of transportation improves transit accessibility and extends the coverage area of urban rail station services. The trip chain connects effectively to land use patterns and urban rail stations. This study analyzed the bicycle-sharing trajectory of travelers departing/arriving at urban rail stations and identified a correlation between urban rail entry/exit ridership and land use pattern using the random forest model. Specifically, it examined the effects of urban rail ridership on residential, business office, commercial service, industrial, administrative, and educational land use patterns in Xiamen, China.

The results suggest that land use patterns vary but have positive effects on station entry and exit ridership during the morning (6:00–10:00 a.m.). The R² for egress ridership is 48.46%, which is greater than the R² for ingress ridership, which is 36.88%. This suggests that the land use pattern could explain the egress ridership of urban stations more effectively than the ingress ridership. Residential land use has the greatest impact on ingress ridership, followed by business office land use. In contrast, the land use of the business office is the most easily explained in terms of egress ridership. Administrative land use is least able to explain ingress and egress ridership.

The following are the primary contributions of this study:

(1)

The bicycle-sharing trajectory revealed travel behavior that addresses the relationship between urban rail ridership and land use pattern behavior.

(2)

Ingress and egress urban rail ridership is better explained by bicycle-sharing trajectory than total ridership.

(3)

Different land use patterns provided inconsistent explanations for morning ingress and egress urban rail ridership (6:00–10:00 a.m.).

(4)

In general, residential and business office land use patterns have the greatest impact on urban rail station access ridership. On the other hand, the egress ridership, business office, and educational land use patterns are sometimes more influential.

(5)

Cost-effective integration of public transportation will increase ridership.

The implications of these findings for increasing urban rail ridership by land use pattern and bicycle-sharing were discussed. First, residential land use and business office land use have the greatest impact on urban rail ridership, particularly during the morning rush hour. Second, some commuters cycle to and from their urban rail station, and bicycle-sharing could improve the performance of that transit element. Third, different land use patterns have a distinct impact performance that should be taken into account with greater precision during enhancements to maximize profit. Finally, providing more transport options improves transit accessibility, which is conducive to increasing urban rail ridership, sustainable transport system development, and creating a livable city.

The limitations of this study can be addressed in future research. First, this paper focuses solely on morning travel; if data are available, it could also consider the rest of the day, as travel behavior may vary according to time of day. Second, this paper used a dock-less bicycle-sharing trajectory to measure the correlation between land use pattern and urban rail ridership; however, bicycle-sharing usage is affected by numerous factors, such as road infrastructure and parking space; thus, further research could analyze the built environment within an urban rail station catchment area more thoroughly. Third, the travel behavior of cyclists should be studied in greater depth, because personal characteristics will also impact bicycle-sharing usage, potentially affecting this triangular relationship. Fourth, more details should be considered in understanding the transfer activity with integrated metro-bicycle-sharing; while this study concluded that the urban rail station coverage range is 100 m in this transfer system, there is bias. Therefore, future research could consider this aspect more thoroughly.