PM2SFCA: Spatial Access to Urban Parks, Based on Park Perceptions and Multi-Travel Modes. A Case Study in Beijing

Abstract

Assessing park accessibility plays an essential role in providing rational recreational services for residents in a city. The perceptions and comments of residents are also important nonspatial factors for accessibility. However, there are few accessibility studies that are combined with public perceptions. Addressing this deficit, this study proposes a perception-based, multi-travel mode, two-step floating catchment area (PM2SFCA) method to calculate park accessibility. First, we quantified the selection probability of residents to parks by integrating the Huff model and the people’s perceptions towards parks. Next, under four travel modes (walking, biking, driving and public transport), we combined the Huff model and the two-step floating catchment area method to compute park accessibility. Furthermore, the Gini coefficient and the Pearson correlation coefficient were used to illustrate the proposed method compared with the traditional E2SFCA method. Based on the above, taking the area of Beijing within the Fifth Ring Road as a study area, this paper facilitated the accessibility computation. The results indicated that the spatial distribution patterns of accessibility differed greatly under the four travel modes. Even under the same travel mode, there was an uneven accessibility distribution. Areas with high accessibility were mainly concentrated in the north, and some marginal areas also presented higher accessibility to parks. The comparative analysis results suggest that our proposed method for accessibility measurements alleviates the underestimation and overestimation of accessibility values obtained by a traditional method such as the center and edge of the study area. The research explores a new research perspective for measuring park accessibility. Furthermore, this study offers better guidance for policymakers trying to optimize park spatial distribution issues.

1. Introduction

Urban parks are landscape features in a city that provide recreation, relaxation and communication services for people [1]. As one of the major infrastructures in cities, they are vital public spaces that can offer various services for residents and bring valuable ecosystem services [2]. In an urbanizing society, high-quality urban parks for residents are increasingly important [3]. However, some cities have a mismatch between the spatial patterns of park resources and the population [4]. Therefore, it is essential to allocate park resources rationally. In order to explore the spatially irregular distributions of parks, the use of assessments of the accessibility to parks has begun to increase.

Spatial accessibility, which assesses residents’ accessibility to a particular public service, has long been a research interest in many fields, such as urban planning, social work, public health and geography [5]. With a growing interest in spatial accessibility, the methods and models have constantly changed and improved. Hansen proposed the gravity model, which considered both the supply-side scale and the distance cost in 1959 [6]. Some researchers added the population scale factor, introducing the concept of the gravity potential model [7]. As a variation of gravity-based models, the two-step floating catchment area (2SFCA) method [8,9] is another increasingly popular method, which can appropriately measure potential spatial accessibility [10]. To better assess accessibility, previous studies took full consideration of the setting parameters of the 2SFCA method. In view of the distance decay, search radius and the quantification of supply or demand, Kanuganti et al. calibrated a distance decay function from the travel behavior of patients, named E2SFCA, which showed better results in finding healthcare accessibility in rural areas [11]. By applying a catchment area-based accessibility metric that captures diverse travel patterns, McGrail et al. provided accurate estimates of access to healthcare [12]. Some scholars incorporated multiple transportation modes into accessibility measurement and provided more realistic accessibility estimations [13]. To overcome the problem of overestimation in the 2SFCA, Luo generated a probability-based estimate of the selection weights [14]. Many scholars measured park accessibility according to the characteristics of parks, for example, using park size [15], park on-site amenities [16] and multiple entrances [17] as the metrics.

A more accessible area indicates a higher potential for access to public facilities [18]. In addition to the spatial factors (traffic distance or time–cost), nonspatial factors are also crucial in examining accessibility, such as the self-attributes of residents (economic income, age, ethnicity, gender, individual wishes and preferences) [19]. Li et al. considered the impact of travel behavior to assess the spatial equity of parks in Nanjing [20]. Guo et al. estimated park access and explored whether older people could experience equitable park access [21]. Hughey examined the relationship between park availability and characteristics including neighborhood disadvantage and racial/ethnic identity [22]. Access to parks for children and teenagers was investigated in Rigolon’s study [23]. Hu et al. explored the spatial relationship between park accessibility and population density [24]. Under the improvement of people’s ability to obtain information, spatial factors such as distance place fewer constraints on park access, and the role of subjective choice becomes more obvious [25]. Person perception refers to the different mental processes that we use to form impressions of other people or places [26]. The public perceptions towards urban parks are emotional connections between humans and nature [27]. Furthermore, a considerable amount of research studies have investigated the relationship between humans and parks. It was found that visiting parks is a key way for urban residents to enhance positive emotions [28]. Researchers used social media (Dianping) data to investigate the potential factors affecting public satisfaction with urban parks [29]. Wang et al. studied the tourists’ emotional cognition regarding their memories of Beijing’s historic sites through interviews [30]. Guo et al. measured the specific factors affecting the radius of park services [31]. Exploring the physical activity and events in park areas, the literature has investigated the popular activities among urban green space visitors, respectively [32,33]. Roberts determined the emotional response of visitors in park areas [34].

Despite existing studies of the original and improved 2SFCA method having many advantages in measuring potential spatial accessibility, most studies on park accessibility in different cities have shown that accessibility is distributed unevenly across spaces, that areas with high accessibility were the least distributed [35,36,37] and that access to park spaces has constantly changed at different periods [38,39]. However, they ignored the public perception, which would lead to bias in accessibility calculations. Based on the above, this study attempted to measure park accessibility from the perspective of humans towards parks. Crowd-sourced text data have the potential to provide a broader overview of human attitudes and perceptions to explain human behavior. Therefore, taking the area of Beijing within the Fifth Ring Road as a study area, we proposed a PM2SFCA method to implement the calculation of park accessibility. Meanwhile, the Pearson and Gini coefficient comparison was used to verify our proposed method. Combining the online reviews, real travel time and survey data, the accessibility results under four travel modes presented a significantly different pattern. In the north and some peripheral areas, areas have relatively high accessibility. The framework of this paper is as follows (Figure 1):

The remainder of this paper is organized as follows. Section 2 introduces the study area and datasets. Section 3 introduces the methods for measuring park accessibility. Section 4 shows the results of spatial characteristics and a comparative analysis of accessibility. The final section discusses the advantages, implications and limitations of our proposed method and concludes our findings.

2. Study Area and Datasets

2.1. Study Area

As the political, cultural and international exchange center of China, Beijing is an ancient city with a long history, between longitudes 115.25° E and 117.30° E and latitudes 39.28° N and 41.25° N [40]. A growing imbalance between the supply and demand of public services exists in this city due to rapid urbanization. To optimize the spatial structure, the Beijing government started the construction of urban parks. The permanent population was 21.705 million by the end of 2015. The total administrative division area was 16,410 km², the green space coverage was 813.05 km², of which park areas constituted 295.03 km² [41]. Many historical parks were built during the Ming and Qing dynasties (such as the Summer Palace and Beihai Park), and many public parks were built after the founding of the People’s Republic of China (such as Zizhuyuan Park and Yuyuantan Park). In this study, we limited the study area to within the Fifth Ring Road, which is the most densely populated area (Figure 2).

2.2. Datasets

The data we collected included the traffic analysis zone (TAZ), Ctrip online reviews and real travel times (

Table 1. The information of datasets.

Data	Source	Characteristic
Traffic analysis zone	Based on the sub-district	Spatial unit with population information
Ctrip online reviews	https://www.ctrip.com (accessed on 2 December 2021)	People’s attitudes and perceptions of one place
Real travel times	Baidu map	Time people consumed from one place to another

).

2.2.1. Traffic Analysis Zone

As a spatial unit, the TAZ contains information on socio-economic attributes such as area, population and land use. Compared with the sub-districts, the unit area of the TAZ is smaller, thus, it can reflect the distribution of accessibility in more detail. After cutting the sub-district administrative units by main roads, we obtained the TAZ. According to the area ratio of the TAZ to the sub-district and the population of the sub-district [42], this paper collected demographic attributes of the TAZ. The study area was divided into 280 TAZs, as shown in Figure 3.

2.2.2. Ctrip Online Reviews

As one of the largest online travel agencies in China, Ctrip serves hundreds of millions of people (https://www.ctrip.com (accessed on 2 December 2021)). People can book their travel tickets on this website. In addition, it offers many travel plans and much information on scenic spots, which are great references for travelers. Many users voluntarily share comments on the places they visit and their travel experiences on the travel website [43]. In this study, we used Ctrip as our data source and collected online reviews of urban parks in Beijing from 1 January 2002 to 31 August 2021. When collecting and processing the data, the duplicate reviews and reviews with no text content were removed. To ensure adequate data samples, we chose the urban parks that are closest to people’s lives, with more than 100 online reviews. On this basis, 49 urban parks and 44,380 Chinese reviews were collected for further analysis. The size of the parks range from 2.11 acres to 302 acres. Large parks are mainly located in the urban peripheries, while small parks are mainly located in the inner city (Figure 3).

2.2.3. Real Travel Times for Different Travel Modes

The real travel times from each TAZ to each park were obtained using the route planning service of Baidu Maps, which, combined with real-time traffic, provides users with route planning services covering domestic and foreign countries. Considering the road congestion and traffic flow, this would provide an objective value to avoid much of the road network processing work, and the results would be more reliable. We collected the estimated travel time during the off-peak periods for four travel modes. Samples of the travel time records are shown in

Table 2. Samples of the real travel time records.

TAZ Index—Park Index		1—1	1—2	…	2—1	2—2	…
Real Travel Time (s)	Walk	7770	8014	…	11,091	11,277	…
	Bike	3214	3748	…	4590	4854	…
	Car	1352	1141	…	1344	1219	…
	PT	3504	4144	…	3379	3983	…

. In Table 2, 1—1 means the real travel time between the first TAZ and the first park.

3. Methods

3.1. Sentiment Analysis of Park

As a method, sentiment analysis transforms ambiguous emotions into numerical sentiment scores, which can reveal public perceptions and attitudes towards a certain place or event. Based on Baidu’s immense data accumulation, Baidu’s language and knowledge has made a contribution to the development of cutting-edge natural language processing and knowledge graph technologies [44]. Natural language processing has opened several core abilities and solutions, with more than ten types of ability, such as sentiment analysis, address recognition and customer comments analysis [45]. Based on pre-trained, state-of-the-art machine learning models, the Baidu Natural Language Processing (NLP) platform (https://ai.baidu.com/tech/nlp (accessed on 10 June 2022)) can achieve high accuracy computations on sentiment analysis tasks. Thus, to quantify park perceptions, we used it to perform sentiment analysis on Chinese texts. We obtained the sentiment score for each online review by accessing their application programming interfaces (APIs). Sentiment scores from low to high (0–1) indicate negative, neutral and positive emotional polarities. For each park, the average sentiment score of all the reviews was used as a proxy for its perception. Based on the distribution of sentiment scores, we classified online reviews into three sentiment polarities by setting specific thresholds [46], as follows: positive (sentiment score ≥ 0.9), neutral (0.5 < sentiment score < 0.9) and negative (sentiment score ≤ 0.5). We named park sentiment as variable $P_j$ and used it in later analyses.

3.2. The PM2SFCA Method

Our proposed PM2SFCA method considers public perceptions and attitudes towards urban parks and different travel modes. As the component of 2SFCA, the demand coefficient represents the population located in the parks’ service areas. On the basis of the traditional 2SFCA method, we improved the demand coefficient by incorporating the Huff model. It added the consideration of the selection probability $Pro{b}_{ij}$ of residents towards parks. The detailed improvements and equations of spatial accessibility computations are as Equation (1).

3.2.1. Selection Probability by the Huff Model

First, previous research on search radiuses made the assumption that people visited parks with only a single travel mode [47] or used a uniform search radius [48], combining multiple travel modes. However, they ignored residents’ real-life travel behavior, as the acceptable travel time varies depending on the mode of transportation. Therefore, we referred to the method in this study [21] to determine the travel time threshold. The online questionnaire platform (https://wj.qq.com, accessed on 3 July 2022) is operated by China’s largest instant messaging company, Tencent. We collected 120 valid online and offline questionnaires about residents going to the park using different travel modes and the maximum tolerable travel time of each travel mode. The travel modes included walking, biking, car and public transport, the units of recorded contents were minutes. The average value of each result collected was used as the search radius specific to each travel mode. The results (

Table 3. The maximum acceptable travel time under four travel modes.

Travel Mode	Walk	Bike	Car	PT
Search radius (min)	20	30	35	60

) showed that the maximum acceptable time was 20 min for walking, 30 min by bike, 35 min by car and 60 min by public transport.

The traditional 2SFCA method considered the scale factors of both the supply-side and the demand-side, as well as the distance cost factor. It has been widely used in a variety of studies. Luo and Wang initially assumed equal accessibility within a catchment to health care facilities [8]. However, it disregarded the competition among facilities, which may have resulted in errors. Wan enhanced the measurement by considering the competition effect and named it the 3SFCA method [49], which modified the population demand with a selection weight that only considered the distance factor and ignored the impact of the facility itself. However, the characteristics of the facilities were also important factors that affected people’s decisions. To resolve the influence of both distance impedance and demand site on spatial accessibility, we introduced the Huff model [50] into the 2SFCA, which quantified the selection probability of people on a service site among the multiple ones available. In this study, the Gaussian function [51] was adopted to continuously capture the distance decay effect of accessibility within a catchment. The formulas were as Equation (2), as follows:

$$Pro{b}_{ij}=\frac{P_j{G}_{ij}}{{{\displaystyle \sum }}_{t_{j,i}<t_0}{P}_j\times G_{ij}}$$

(1)

$$G_{ij}=\left\{\begin{array}{c}\frac{e^{-\left(\frac{1}{2}\right)\times {\left(\frac{t_{i,j}}{t_0}\right)}^2}-e^{-\left(\frac{1}{2}\right)}}{1-e^{-\left(\frac{1}{2}\right)}}, t_{i,j}\le t_0\\ \\ 0 , t_{i,j}\ge t_0\end{array}\right.$$

(2)

where $Pro{b}_{ij}$ is the possibility of visiting park $j$ for residents in TAZ $i$; $t_{j,i}$ denotes the travel time from $i$ to $j$, and the Gaussian function ($G_{ij}$) is the friction of time distance; $C_j$ is the perception of park $j$; and $t_0$ is the travel time threshold.

We present an example in Figure 4 to illustrate Equation (1). For residents in TAZ $i$, they can visit three parks—$A,B,C$—with different park perceptions, $P_A,P_B,P_C$ (0.9, 0.8, 0.7). The distance decay $G_{ij}$ was indicated between TAZ $i$ and park $j$ (0.3, 0.5, 0.8). Based on the above, the selection probability of residents in TAZ $i$ to park $A$ was (0.3$\times 0.9$)/(0.3$\times 0.9+0.5\times 0.8+0.8\times 0.7$).

3.2.2. Park Accessibility

The first step of the traditional 2SFCA method was that for each park $j$, we searched all population locations ($i$) within a time/distance ($t_0$) threshold from park $j$, which formulated the catchment area for park $j$. Populations at $i$ were weighted by a distance decay function. The sum of weighted populations was calculated as the potential demand of park $j$.

In this study, we integrated the Huff model, and the supply-to-demand ratio $R_j$ of park $j$ was based on the 2SFCA method and was computed as Equation (3), as follows:

$$R_j=\frac{S_j}{{{\displaystyle \sum }}_{t_{j,i}<t_0}{P}_j\times Pro{b}_{ij}\times G_{ij}}$$

(3)

where $S_j$ is the size (in km²) of park $j$; $P_j$ is the population of demand center $j$, falling within the catchment of supply center $i$; $Pro{b}_{ij}$ is the selection probability of visiting park $j$ for residents in TAZ $i$; $t_0$ is the travel time threshold between supply center $i$ and demand center $j$; and $G_{ij}$ is the friction of time distance.

Following the second step of the 2SFCA method, the spatial accessibility at community $i$ was calculated as in Equation (4), as follows:

$$A_i={{\displaystyle \sum }}_{t_{j,i}<t_0}{R}_j\times Pro{b}_{ij}\times G_{ij}$$

(4)

where $Pro{b}_{ij}$ is the selection probability of visiting park j for residents in TAZ $i;$ $t_0$ is the travel time threshold between supply center $i$ and demand center $j$; $G_{ij}$ is the friction of time distance; $R_j$ is the supply–demand ratio of park $j$; and $A_i$ is the availability of the park area (in km²/10 thousand person) for each population location $i$;

3.2.3. Global Spatial Autocorrelation Analysis

As a spatial statistical method, spatial autocorrelation and local spatial autocorrelation are important concepts for measuring the degree of similarity between spatially proximate data, and indicating whether features are spatially dependent or independent [52]. According to Anselin [53], the local indicators of spatial association (LISA) are used to examine the spatial patterns of the data at the local level to identify the local clusters. Thus, in this study we used ArcGIS to evaluate local Moran’s I for each spatial unit and the statistical significance. The LISA was named local Moran’s I, which was calculated as in Equation (5) [54], as follows:

$$\begin{array}{c}{I}_i=\frac{x_i-\overline{x}}{{\mathrm{S}}^2}{\displaystyle {\displaystyle \sum }_{i,j=1}^n}{W}_{ij}(x_j-\overline{x)}\\ {\mathrm{S}}^2={\displaystyle {\displaystyle \sum }_{j=1}^n}{x}_j{}^2/\left(n-1\right)-x^{\overline{2}}\end{array}$$

(5)

where $n$ is the number of spatial units; $\overline{x}$ is the average value; and $x_i$ and $x_j$ are the attribute values of space objects at points $i$ and $j$. $W_{ij}$ is the spatial weight matrix which represents the distance relationship of points $i$ and $j$.

3.3. Comparative Analysis Method

To illustrate the newly proposed method, we made some comparisons of the new method (called PM2SFCA) with the traditional E2SFCA method. The difference between the two methods was the inclusion of a preference factor. In the E2SFCA method, we also estimated the accessibility values separately by travel mode and the different travel time thresholds were the same as the PM2SFCA method.

In order to demonstrate the degree of variability, the modified growth rate was used to describe the difference (Equation (6)). The large value meant a great difference.

$$D_i=(A_{i\left(PM2SFCA\right)}-A_{i\left(2SFCA\right)}) /A_{i\left(2SFCA\right)}$$

(6)

where $D_i$ is the difference value between the PM2SFCA method and the E2SFCA method; $A_{i\left(PM2SFCA\right)}$ is the accessibility result to each TAZ by using the PM2SFCA method, while $A_{i\left(E2SFCA\right)}$ is that of using the E2SFCA method.

Following the existing equity research on health care accessibility [55], we plotted the PM2SFCA output values against the E2SFCA. Furthermore, we calculated the Pearson correlation coefficient $R^2$ [56]. It was a representative way to measure the relationship between variables, with a value ranging from −1 (perfect negative correlation) to 1 (perfect positive correlation), with 0 representing no correlation between two objects. The relative formula is as follows:

$$R^2=\mathrm{cov}\left(X,Y\right) /{\mathsf{\sigma }}_x{\mathsf{\sigma }}_y$$

(7)

where $Y$ is the accessibility result of using the PM2SFCA method, while $X$ is that of using the E2SFCA method. ${\mathsf{\sigma }}_x,{\mathsf{\sigma }}_y$ is the standard deviation of $x$ and $y$. $\mathrm{cov}\left(X,Y\right)$ is covariance.

The Lorenz curve [57] and the Gini coefficient (GC) [11] were used to represent the relationship between the accessibility of urban park resources and the resident population. It is one of the most widely used measures of inequality [58]. Mathematically, the Gini coefficient depends on the size of the area between the Lorentz curve and the line of perfect equality. The area below the line of perfect equality (0.5 by definition), minus the area below the Lorenz curve, divided by the area below the line of perfect equality is the Gini index. It has a value between 0 (or 0%) and 1 (or 100%). In this study, a higher value indicated that the distribution of parks was extremely uneven, with 0 representing perfect equality and 1 representing perfect inequality. In this study, Equation (8) was used to calculate the Gini coefficient.

$$GC=\left| \left[{\displaystyle \sum }_{i=1}^n \left(A_{i-1}+A_i\right)B_i\right]-1\right|)$$

(8)

where $A_i$ is the cumulative percentage of rank-ordered park accessibility in TAZ, and $B_i$ is the cumulative percentage of residents in TAZ.

4. Results and Analysis

4.1. Spatial Characteristics of Park Accessibility

4.1.1. Public Perceptions of Park

We obtained a total of 44,380 online review sentiment scores by using the Baidu Natural Language Processing platform. Figure 5 depicts the frequency of the sentiment scores. We found that most of the reviews were positive, accounting for 86.44% compared to 8.84% for negative. Overall, it demonstrated that people have a positive attitude towards urban parks. Figure 6 illustrates the area and perception scores of 49 parks. Almost all parks presented high scores. The values ranged from 0.828 to 0.973. In addition, we found other situations, where some parks in the center did not have the highest score (e.g., Beijing Working People’s Cultural Palace), on the contrary, some parks that are located in the urban peripheries had higher scores (e.g., Linglong Park and Jiangfu Park).

4.1.2. Park Accessibility under Four Travel Modes

Figure 7, Figure 8, Figure 9 and Figure 10 show the accessibility results under four travel modes (walking, biking, driving and public transport). The park accessibility value was divided into six grades, using the Jenks natural breaks method—the color from light to dark indicates the gradual increase in accessibility. The unit of value is km² per ten thousand persons. As shown in Figure 7, for the spatial distribution in the walking accessibility results at each grade, a quick, detailed observation suggested that most of the TAZs belonged to the “not accessible” type and only five TAZs belonged to the high accessibility category. The accessibility value ranged from 0 to 96. Most of the areas with low accessibility were within the Fourth Ring Road. The reason for this is that we set the travel threshold to 20 min, according to the real travel habits of residents. However, the travel mode of walking is rarely affected by traffic. When the supply facility is far from the TAZ, the accessibility was so low that the results had more blank areas. In the results for the bike travel mode (Figure 8), most TAZs belonged to the low accessibility category. The lowest value was 0 and the highest was 16, which was lower than the walking accessibility. This may be related to the travel threshold—the biking mode has a larger travel range, and more residents will fall within the catchment of the supply center. When the supply–demand ratio decreases, this leads to lower accessibility. As a result, there were more areas with high accessibility in the north than that in the south, due to the economic development and park distribution of Beijing. Spatial disparities in accessibility were different between ring roads. Areas with high accessibility were mainly concentrated inside the Second Ring Road. In the north-east and north-west of the study area, there were two significant high accessibility areas, most likely related to the sparse population there.

Compared to the walking and biking modes, the accessibility results of driving and public transport had fewer blank areas (Figure 9 and Figure 10). In the results of the driving mode, the maximum value of accessibility was seven, the reason was that the driving mode has the maximum speed among all travel modes. Therefore, it has a wider range of services. This will increase the population of the demand side, leading to lower accessibility. We can see that the number of areas with high accessibility decreases from north to south, and the number of areas with high accessibility in the west is more than that in the east. Between the Fourth Ring Road and the Fifth Ring Road, the high accessibility areas are basically concentrated. In the south-east of the study area, there are some high accessibility areas, which is mainly due to the dense population compared to the surrounding areas. The maximum accessibility value for the public transport mode (Figure 10) is 90. Overall, it had a lower and balanced accessibility. The reason for this is probably related to the characteristics of the public transport mode—instead of taking the nearest route to one place, it chooses the route with multiple stops. Therefore, people will spend more time on the road, resulting in most of the areas having lower accessibility results.

4.1.3. Descriptive Statistics and Global Spatial Autocorrelation of Accessibility Results

Table 4. The statistics on park accessibility under different travel modes.

PM-2SFCA					Not Accessible
Travel Mode	Min	Max	Mean	Std. Deviation	(TAZ Numbers) (Total 280)	Percentage of Low- Accessibility Areas (%)	Percentage of High- Accessibility Areas (%)
Walk	0.000	95.498	1.705	7.944	226	95.7	1.8
Bike	0.000	15.666	1.523	2.401	69	85.7	5.7
Car	0.000	6.274	3.318	0.688	1	37.1	23.6
PT	0.000	89.386	2.832	5.885	18	96.1	1.8

summarizes seven descriptive statistics of these results. They are largely influenced by the mode of transportation. The mean values of park accessibility under the four travel modes were 1.705, 1.532, 3.318 and 2.832, respectively, and the standard deviations of park accessibility were 7.944, 2.401, 0.688 and 5.885, respectively. The walking mode had the highest maximum value, followed by the public transport mode. The mean of the driving accessibility was greater than that of other travel modes’ accessibility, and it identified the least underserved TAZs. The walking accessibility showed the largest standard deviation and had the most non-accessible TAZs. This is because the two travel modes have different characteristics, the latter has more parking options and is a convenient travel mode, whereas the former does not, due to distance constraints. In terms of each travel modes’ accessibility, the driving mode had the most high-accessibility areas and the fewest low-accessibility areas, followed by the biking mode.

Figure 11 reflects the local spatial autocorrelation in the local Moran cluster map under four travel modes, which examined the relationship among different neighborhood types. It can be classified into five types, namely: high–high cluster (HH) in the red color, low–low cluster (LL) in the blue color, high–low cluster (HL) in pink, low–high cluster (LH) in light blue, and non-significant in grey. We can see that among the four travel modes, HH values are mainly distributed in the north, while LL values are in the south. Meanwhile, some areas with LL clusters are also distributed in the north and south-east, under the biking and public transport modes. In the urban center, most areas had no significant cluster characteristics. There were six TAZs with HH clusters under the walking mode, but no areas with LL clusters. In the biking mode, 32 and 69 TAZs belonged to HH and LL clusters, respectively. The number of TAZs with an HH cluster increased to 59 in the driving mode, which was more than that in other modes. In terms of the public transport mode, it had the fewest areas with HH clusters and more TAZs with LL clusters.

4.2. Comparative Analysis with Traditional E2SFCA Method

4.2.1. Accessibility Comparison

As shown in Figure 12, the positive value (warm color) indicates that the accessibility result of PM2SFCA is larger than that of the traditional method in the same area, while the negative value (cold color) is opposite to the positive value. The larger the value, the greater the difference. Except for the walking mode, the distribution of the difference presented a ring form, as follows: the value gradually increases from the center to the periphery of the Fifth Ring Road. For the biking mode, most of the areas had a positive value (0–270%). Areas with a high positive value under the driving and public transport modes were mainly distributed along the edge of the study area. The public transport mode had the greatest accessibility differences, with the lowest value being −60% and the highest being 1600%. Inside the Fourth Ring Road, there were more areas with a negative value. It suggests that, although most of the urban parks are concentrated in the center of the study area, the accessibility result of the PM2SFCA was lower than that of the E2SFCA method in these areas. At the edge of the study area, the accessibility result of the PM2SFCA was larger than that of the E2SFCA method in these areas.

4.2.2. Pearson and Gini Coefficient Comparison

In this part of the study, we used a scatter plot and the Gini coefficient to compare the equity degree of park accessibility in different situations. Due to a limited number of walking accessibility results, we only show the scatter plot results (Figure 13) under three travel modes (biking, driving and public transport). The horizontal axis shows the results of the PM2SFCA method, while the vertical axis shows the results of the E2SFCA method. These results show that a small number of values fell below the 1:1 line in the E2SFCA and slightly more in the PM2SFCA. In comparison to the PM2SFCA, the E2SFCA has the potential to overestimate and underestimate spatial accessibility. According to the Pearson correlation coefficient, it makes sense that the two methods present a more positive correlation ($R^2$ = 0.838, 0.871) under the biking and public transport mode. However, it had a significantly less positive correlation under the driving mode.

Figure 14 shows eight Lorenz curves and Gini coefficients for two methods and four travel modes. The horizontal axis represents the cumulative percentage of residents, based on ascending order of accessibility. The vertical axis represents the cumulative percentage of park accessibility. The dotted line indicates the perfect equity. We can see that the Gini coefficient of PM2SFCA under the driving mode was lower than that under the other travel modes. For the biking and public transport modes, the Gini coefficient of PM2SFCA was larger than that of E2SFCA. In addition, under the public transport mode, the difference of the Gini coefficient for two methods was larger. Namely, our proposed method could better balance the accessibility of the marginal areas, avoiding the overestimation or underestimation of some areas under the E2SFCA.

5. Discussion and Conclusions

Public perceptions and comments are important information for the planning and management of urban parks. In addition, measuring the spatial accessibility of parks is an effective way to identify disparities between areas, which can be better managed in the future. In this paper, we chose the online reviews on Ctrip as our data source, combined with the real travel time and surveys on residents, which overcame the limitations of the traditional methods of data collection. We proposed a PM2SFCA method to estimate people’s spatial access to urban parks in Beijing.

As a result of the park perceptions, the perception score of the park had no obvious relationship with its area and location. The reasons for these results can be found in Zhang’s study [59]. Even in the fringe area, those parks with excellent environments are still well-received by people: “Good wine needs no bush”. Based on the multi-travel modes (walking, biking, driving and public transport), we collected the real-time travel times and measured park accessibility. It was found that different travel modes have significant differences in the accessibility spatial pattern. Furthermore, the study units presented different accessibility values for the same travel mode. These situations are similar to the previous accessibility study on shopping stores [42]. The accessibility value for the walking and public transport modes was high, while for the biking and driving modes it was low. There were more areas with accessibility results under the driving mode. Overall, the accessibility results were varied in different directions. Areas with high accessibility were mainly distributed in the north. The dense population has led to a sharp decrease in green spaces in the urban core. In some peripheral areas, all travel modes presented high accessibility. Therefore, the characteristics of the travel modes and population density were important factors that influenced the results. The walking and biking modes are rarely affected by traffic conditions, and both have similar travel routes, while the driving and public transport modes are easily affected by traffic. The accessibility results of areas were different among the five ring roads. The area inside the Second Ring Road presented relatively high accessibility. These areas are older urban areas, dating from the Ming and Qing dynasties. Until now, the green park spaces have been distributed evenly [60]. Multi-travel modes offer the real travel situations of people and provide a more realistic assessment of spatial access. Relevant departments could consider the aspect of road congestion to improve traffic accessibility. In the results of the local spatial autocorrelation analysis, the north and south urban peripheries had HH and LL value clusters, respectively. It concurs with the previous conclusion that there is “a significant imbalance in access to the park resources” [61].

In the results of the comparative analysis with the E2SFCA method, our proposed method showed the advantage of being simple and convenient to implement. Although a large quantity of data needs to be collected to get a more realistic result, the improvement in technology makes the procedure easier. Integrating the Huff model in formulating population selections on parks, and parks with different sentiment scores led to a change in accessibility. The results showed a significantly lower positive correlation under the driving mode. This is because people could obtain more access to parks by car. Therefore, using the new method, PM2SFCA, this travel type strengthened the changes in the accessibility results in different areas, which led to a less positive correlation. The PM2SFCA moderates the population demand on service sites. Many parks are located in the center of the study area; however, a huge population and perception factor moderate the high accessibility. Moreover, the edge areas with a small population and large parks do not have a low-accessibility value. To some extent, our proposed method alleviates the underestimation and overestimation of accessibility values by a traditional method. Regarding guidelines for designing parks or undertaking city planning in the future, the government should pay more attention to the perceptions of residents. The findings will be of practical value in assisting policymakers to ease spatial imbalance and provide people with better services and a higher quality of life.

However, there are also some limitations that we cannot ignore. First, due to the lack of data, we only collected the TAZs inside the Fifth Ring Road of Beijing, which may have resulted in the edge effect. In the procedure of determining the time threshold and calibrating the model, the sample size for the questionnaire and the parks choices being due to the reviews may have resulted in bias. In the next research study, we hope to collect more data to better demonstrate the accessibility results of Beijing. Second, a high park perception would result in a high selection probability in our model. Thus, the park perceptions varied more, which could lead to the difference in the accessibility results. However, the sentiment scores of the parks in Beijing were quite similar to one another, therefore, their inclusion in the model had little impact. We hope to detect park accessibility in different cities or countries in a subsequent study. Third, although we have considered people’s real travel situations by collecting questionnaires, residents’ preferences are different in other aspects. Therefore, in future research, we could survey the preferences of residents with various non-spatial questions, e.g., asking them how much time per week they can spend in the parks and which mode of transport they prefer. Moreover, we can explore the facts that influence public perception. Nevertheless, for the overall accessibility distribution, the influences of such limitations were small due to their global nature. In the future, for other public facilities whose reputations can easily affect residents’ choices (such as hospitals, shopping stores and libraries), we will try to apply our proposed method to evaluate accessibility.