PCA-Based Identification of Built Environment Factors Reducing PM_2.5 Pollution in Neighborhoods of Five Chinese Megacities

Abstract

Air pollution, especially PM_2.5 pollution, still seriously endangers the health of urban residents in China. The built environment is an important factor affecting PM_2.5; however, the key factors remain unclear. Based on 37 neighborhoods located in five Chinese megacities, three relative indicators (the range, duration, and rate of change in PM_2.5 concentration) at four pollution levels were calculated as dependent variables to exclude the background levels of PM_2.5 in different cities. Nineteen built environment factors extracted from green space and gray space and three meteorological factors were used as independent variables. Principal component analysis was adopted to reveal the relationship between built environment factors, meteorological factors, and PM_2.5. Accordingly, 24 models were built using 32 training neighborhood samples. The results showed that the adj_R² of most models was between 0.6 and 0.8, and the highest adj_R² was 0.813. Four principal factors were the most important factors that significantly affected the growth and reduction of PM_2.5, reflecting the differences in green and gray spaces, building height and its differences, relative humidity, openness, and other characteristics of the neighborhood. Furthermore, the relative error was used to test the error of the predicted values of five verification neighborhood samples, finding that these models had a high fitting degree and can better predict the growth and reduction of PM_2.5 based on these built environment factors.

1. Introduction

Air pollution, especially PM_2.5 pollution, still seriously endangers the health of urban residents in China. According to the 2020 World Air Quality Report published by IQAir, China is 14th in the rankings for poor air quality among the 106 countries that have been given air quality monitoring stations by the WHO. In particular, the middle and lower reaches of the Yangtze River are among of the most polluted areas in China, where a large number of residents live. Serious PM_2.5 pollution has resulted in respiratory issues, asthma, and even death [1]. To maintain the basic requirement of respiratory health, it is urgent to improve air quality.

Spatial-temporal variations in PM_2.5 and the impact factors on PM_2.5 have attracted much attention in recent years; among these studies, PM_2.5 levels were mainly affected by socioeconomic environments [2], climatic conditions [3], and urban physical environments [4]. Human activities, such as traffic emissions, industrial activities, and coal consumption, cause severe PM_2.5 pollution [5,6]. At the same time, the loss of natural land and increase in artificial ground cover exacerbate the problem [7]. In addition, the change in urban land cover and its spatial pattern leads to environmental problems, especially urban heat island effects that always contribute to gathering atmospheric pollutants or forming secondary pollutants, thereby strengthening PM_2.5 pollution [8]. Investigating the temporal variations in PM_2.5 and its mitigation approach is essential to improve the human settlement environment.

The built environment, usually measured by factors from various perspectives including land cover type, land use, urban form at the urban scale and public space, layout of buildings and roads at the neighborhood scale [9], is one of the important factors affecting PM_2.5; however, the key factors remain unclear. Obvious differences in PM_2.5 levels are found across urban land cover patterns [10]. In addition, urban landscape patterns or structures, including the composition and configuration of built environments, have aroused interest by using a series of relevant metrics to investigate their effects on PM_2.5 [11,12,13,14], although the influence of individual metrics may differ among studies. Moreover, urban morphology has attracted increasing attention because of its strong impacts on PM_2.5. At the city scale, the spatial format of urban built-up areas, such as the size, compactness, fragmentation, and complexity of the morphology of urban areas, influences the city’s average PM_2.5 level [15,16,17]. At the local scale, the PM_2.5 concentration varies from neighborhood to neighborhood [18]. The difference in street canyon characteristics, building layout, and spatial form is one of the important factors that significantly influence PM_2.5 levels [19,20,21]. Local climate zones (LCZs), a concept aimed at classifying local built environment features, have been widely used in urban climate studies [22,23,24]. Ten built-up LCZs and seven land cover LCZs can be provided for measuring different built environments and natural environments, respectively. Indicators, such as the sky view factor, aspect ratio, impervious surface fraction, pervious surface fraction, and height of roughness elements, are frequently used for determining these LCZs [25,26,27]. However, they are rarely involved in the PM_2.5 field. Neighborhood-level PM_2.5 pollution should be a main concern because it is closely related to people’s daily lives. In a high-density neighborhood environment, densely constructed buildings block air ventilation and consequently impede pollutant dispersion [28]. Particular attention should be given to the mitigation of PM_2.5 by optimizing the neighborhood-level built environment, which constitutes the basic fabric in a city.

Previous studies have explored the relationship between the urban built environment and PM_2.5 from different aspects, yet there are still shortcomings. First, the lack of systematic investigation of the built environment may result in the unclear key factors that significantly influence PM_2.5. Second, most studies focus on individual cities instead of regions, which may limit the applicability of the research findings. Considering the complexity of the built environment in a neighborhood, traditional stepwise regression analysis has some limitations, including the potential collinearity among multiple variables and the possibility of removal of some predictive variables significantly related to dependent variables [29]. Principal component analysis (PCA) has been adopted to convert complex variables into new variables that contain most of the original information and are independent of each other, thereby reducing the predictors’ collinearity in PM_2.5 simulation [30].

For this, this study explored key built environment factors influencing PM_2.5 in common neighborhoods that are abundant in cities. We focused on 37 neighborhoods located in five megacities in the middle and lower reaches of the Yangtze River to better understand the influence mechanism of the built environment on PM_2.5 pollution. Principal component analysis (PCA) was conducted before regression analysis to increase its performance. All principal component variables were retained for regression analysis to establish the optimal influencing factors. Therefore, key factors can be obtained.

2. Materials and Methods

2.1. Study Area and Neighborhoods

Five megacities (Wuhan, Hefei, Nanjing, Shanghai, and Hangzhou) in the middle and lower reaches of the Yangtze River were selected as the study area (Figure 1); they have the same climate conditions. Due to the extreme high-density construction, these cities experience more serious PM_2.5 pollution than other cities in this region [31]. The similar features of the five cities provide sufficient comparability for this study, including urban form, landform, population, and PM_2.5 pollution characteristics.

On average, each city has 10 national PM_2.5 monitoring stations that present a relatively uniform distribution throughout the constructed areas. A neighborhood was then defined as a square domain with an area of 1 km² (1 km × 1 km) centered at a monitoring station because 1 km is a common size for a neighborhood division in China and was frequently used in previous studies [32,33]. In addition, the air environment within this size plays an important role in people’s quality of daily life [34]. In consideration of the unique locations of 13 neighborhoods and subsequent special land use, such as close to pollution sources (construction sites) and natural land (urban parks or scenic regions), 37 neighborhoods were selected for analysis in this study. Detailed information on the 37 cities is shown in Table S1.

2.2. PM_2.5 Data Source and Processing

Our previous studies examined the influences of neighborhood green space and urban morphology on PM_2.5 separately, providing evidence for the effects of urban green space coverage and morphological patterns and gray space forms on PM_2.5 [18,32,35]. As a series of studies, hourly PM_2.5 concentrations from 2016 to 2017 were collected from monitoring stations to ensure the consistency of PM_2.5 data. These data in different cities were monitored with the same standard. Because of the different locations of 37 neighborhoods, the uncontrolled factors (such as weather conditions, PM_2.5 background concentration of five cities) were kept at similar levels to remove the external impact factors as much as possible and to greatly minimize evident differences, thereby performing an intercomparison. Consistent with our previous studies [32,35], three relative indicators, the range (C_in and C_de), duration (Δt_in and Δt_de), and rate (C_in’ and C_de’), of the increase and decrease in PM_2.5 concentration were calculated. To investigate pollution-level differences in the effect of the neighborhood-level built environment on PM_2.5, four pollution levels, including slight (PM_2.5 level ranging from 75 μg/m³ to 114 μg/m³), moderate (PM_2.5 level ranging from 115 μg/m³ to 150 μg/m³), heavy (PM_2.5 level greater than 150 μg/m³), and overall pollution, were analyzed based on Chinese ambient air quality standards. The average of slight, moderate, and heavy pollution levels of observations was defined as the overall pollution level. The process of PM_2.5 data is shown in Figure 2. Consequently, PM_2.5 data of eight, three, and two days in 2016, and eight, two, and one day in 2017 were used for slight, moderate, and heavy pollution, respectively (Table S2).

2.3. Built Environment Variables and Meteorological Factors

Built environment factors, including green space and gray space, were included in this study, as well as three important meteorological factors: atmospheric temperature (T_a), relative humidity (RH), and wind velocity (V). A total of 22 variables were selected for analysis based on the quantity and spatial pattern of green and gray spaces (

Table 1. Indicators for model building.

Dimension	Independent Variables			Dependent Variables
	Green Space	Gray Space	Meteorological Factors
Quantity	TCR (x1), GCR (x2)	HSCR (x10)	Ta (x20), RH (x21), V (x22)	Cin (y1), Δtin (y2), Cin’ (y3), Cde (y4), Δtde (y5), Cde’ (y6)
Spatial Pattern	Core (x3), Islet (x4), Perforation (x5), Edge (x6), Loop (x7), Bridge (x8), Branch (x9)	BD_1 (x11), BD_2 (x12), BD_3 (x13), FAR (x14), H (x15), Hσ (x16), BEI (x17), SVF (x18), RD (x19)

). Detailed information on the 22 variables is shown in Tables S3–S5. The determination of these variables also took into account their potential impacts on PM_2.5 and the representativeness of built environment characteristics in neighborhoods.

The measurements and computations of built environment factors were based on a GIS vector dataset generated with a high-precision digital map (Google Earth, 2017) of each city, which ensured high accuracy of the calculation. On the one hand, the neighborhood green space cover ratio (GCR) and tree cover ratio (TCR) were two variables measuring the quantity of green space in the neighborhoods. The spatial pattern of green space was measured by morphological spatial pattern analysis (MSPA), which can provide seven types of green space patterns, including the core, islet, perforation, edge, loop, bridge, and branch [32].

On the other hand, the neighborhood hard space cover ratio (HSCR) was used as the quantity variable for gray space. Spatial pattern variables of the gray space were selected with a consideration of the density, vertical morphology, and spatial layout of gray spaces in neighborhoods. First, building density was considered one of the most important density variables, which was further classified into three categories, including building densities of one to three floors (BD_1), four to nine floors (BD_2), and more than nine floors (BD_3) [35]. In addition, as one of the major pollutant sources in neighborhoods, roads are a special kind of gray space, which may reflect the degree of traffic emissions. Road density (RD) was calculated as in many previous studies [36,37]. Second, the floor area ratio (FAR), mean building height (H), and standard deviation of building height (H_σ) were adopted to measure the vertical morphology of gray space in neighborhoods. FAR reflects the development intensity of the neighborhood. The larger the FAR is, the greater the development intensity and the higher the height of buildings. A neighborhood with a higher building height usually has a lower wind velocity near the ground, which results in better ventilation conditions [38]. H_σ represents the variation of building height. The higher the H_σ, the dispersion of the building height is greater. Third, the building evenness index (BEI) and sky view factor (SVF) were chosen to represent the layout of buildings in the neighborhood. BEI is a reflection of the difference in buildings’ flat form. A neighborhood with a higher BEI value implies a more uneven flat form of buildings. SVF is an important index reflecting built environment geometry [39]. A lower value of SVF indicates a more closed neighborhood space. The 3D building models were adopted for the calculation of SVF based on ArcGIS software according to the method of Gal’ et al. [40].

2.4. Analytical Methods

2.4.1. PCA Analysis

PCA was performed before establishing a regression model to involve all built environment factors and meteorological factors in the models as much as possible. The method can also remove the collinearity of independent variables to a certain extent. The principle of PCA is to convert a large number of factors into new principal factors through certain calculation methods and retain most of the information contained in the initial factors [41]. The principal factors are independent of each other and have no correlation, thereby eliminating the collinearity between factors when performing regression analysis. The relationship between the principal factors and the initial factors is as follows:

$$P_i={{\displaystyle \sum }}_{i=1}^n{l}_{ni}{x}_i$$

(1)

$$l_{\mathrm{ni}}=A_{ni}/\sqrt{{\lambda }_i}$$

(2)

where P_i is the i-th principal factor, n is the number of principal factors, equal to 22, l_ni refers to a principal component loading of x_i, and λ_i denotes an eigenvalue of the i-th principal factor.

Before PCA, it is necessary to carry out standardization due to the various dimensions of each factor in green space, gray space, and meteorology. The Kaiser–Meyer–Olkin (KMO) test and Bartlett sphere test were performed for 22 factors. PCA can be carried out only when the KMO value is greater than 0.5 and the Bartlett sphere test is significant (p-value < 0.01). PCA was then carried out using standardized factors. According to the number of independent variables, a total of 22 principal factors can be obtained. The variance and eigenvalue of the principal factor reflect their contribution to the initial factor. The greater the value, the greater the contribution and the information containing the initial factor.

2.4.2. Regression Models and Evaluation

As Zhai et al. [29] suggested, there may be some problems that the contributions of the predictors truly driving PM_2.5 variations were unclear when using anterior principal components as explanatory variables without explicit rules and standards. To better understand the relationship between the built environment and PM_2.5 and establish regression models, this study carried out stepwise regression analysis involving all principal factors, which can have a screening process for the principal factors and obtain the principal factors that have a significant impact on the dependent variables [29]. The verification of regression models was in accordance with the common methods used in relevant research fields in which the neighborhood samples were divided into test samples and verification samples [35]. The selection of two types of samples should not only consider that there are enough test samples to establish the regression model but also take a certain number of verification samples for validation. Therefore, one neighborhood sample in each city was randomly selected for validation, including WH4, HF5, NJ3, SH4, and HZ4. The remaining 32 neighborhood samples were test samples for the construction of the regression model.

The accuracy of the regression model was measured by comparing the difference between predicted values and actual values of the dependent variable. The relative error (RE) was used to evaluate the accuracy of the predicted values of the PM_2.5 indicators of the five validation samples.

$$\mathrm{R}{\mathrm{E}}_i=\frac{\left|{y_i}^{\prime }-y_i\right|}{y_i}\times 100\%$$

(3)

where RE_i is the relative error of the i-th validation sample and y_i′ and y_i are the predicted value and actual value of a PM_2.5 indicator of the i-th validation sample, respectively.

3. Results and Discussion

3.1. Results of PCA

3.1.1. Overall Characteristics of PCA

PCA had pollution level differences due to the difference in meteorological factors included at the different pollution levels. At the overall pollution level, the KMO value of 22 standardized factors was 0.630, and the significance of the Bartlett spherical test was 0.000, which met the requirements of PCA. As shown in Figure 3a, with the increase in the number of principal factors, the eigenvalues and variance showed a downwards trend. The eigenvalue of principal factors reflects their contribution. The eigenvalues of the first six main factors were greater than 1, and the cumulative contribution of these six principal factors was 85.7%, which can explain most of the information of the initial factors. When the number of principal factors reached 12, it could basically contain 98% of the information of the initial factors. The contribution of the first principal factor (P₁) to the total variance was 35.2%. There was little difference between the loadings of P₁, most of which were between 0.2 and 0.3. P₁ was negatively correlated with the green space factor and positively correlated with the gray space factor. Therefore, P₁ reflects the great difference between the green space and the gray space of the neighborhood. The contribution of P₂ to the total variance was 18.7%. The higher loading of P₂ was higher than 0.3, which was mainly positively correlated with TCR (x₁) and the Edge (x₆), reflecting the large-scale green space of the neighborhood. The contribution of P₃ to the total variance was 11.2%. The higher loading of P₃ was more than 0.4, which was significantly positively correlated with H (x₁₅) and H_σ (x₁₆), reflecting the higher building height in the neighborhood and their greater height differences. P₄ contributed 9.2% to the total variance. Relative humidity (x₂₁) and wind speed (x₂₂) were two principal factors for P₄, representing meteorological factors. P₅ and P₆ contributed 6.4% and 5% to the total variance, respectively, and GCR (x₂) and Loop (x₇) were principal factors.

The variance and eigenvalue showed similar trends at different pollution levels, which were similar to those of the overall pollution (Figure 3b). The KMO values of 22 standardized factors were 0.621, 0.615, and 0.612 for slight, moderate, and heavy pollution, respectively, and the significance of the Bartlett spherical test was 0.000. The eigenvalues of the first six principal factors were greater than 1, and the cumulative contributions of these six main factors were 85.3%, 86.0%, and 83.9%. When the number of principal factors reached 12, they could contain 98% of the information of the initial factors.

3.1.2. Principal Factors Composition

The principal factor (P_i) is composed of 22 initial factors (x_i) with a certain proportion coefficient (principal component loading l). The absolute value of the loading represents the contribution of the initial factor to a P_i, and the positive and negative values represent the influence mode of the initial factor on a P_i. Therefore, the principal component of a P_i is reflected by the initial factor with relatively high loadings. There were certain differences in the higher loading between different principal factors. Taking P₁ as an example, the relationship between it and the initial factor at the overall pollution level is as follows:

P₁ = −0.209x₁ − 0.18x₂ − 0.293x₃ + 0.216x₄ − 0.248x₅ − 0.053x₆ − 0.131x₇ + 0.116x₈ + 0.176x₉ + 0.285x₁₀ + 0.012x₁₁
+ 0.235x₁₂ + 0.238x₁₃ + 0.315x₁₄ + 0.111x₁₅ + 0.083x₁₆ + 0.125x₁₇ − 0.325x₁₈ + 0.252x₁₉ − 0.281x₂₀ − 0.195x₂₁ +
0.23x₂₂

(4)

where x₁, x₂, x₃, …, x₂₂ are the standardization values of the initial factor.

P_i reflects the relationship between each principal factor and its main initial factors. The composition characteristics of each P_i can be sorted and summarized according to the corresponding loading. At the overall pollution level, the loading of 22 principal factors was between −1 and 1. The maximum absolute value was 0.685 and the minimum value was 0.0004. The loading of each principal factor at different pollution levels was basically consistent with this, indicating that the composition characteristics of each principal factor were similar.

3.2. Construction and Verification of Models

Based on 22 principal factors, 24 regression models of six PM_2.5 relative indicators were carried out for four pollution levels, which included principal factors that significantly influenced PM_2.5. Regression models of six PM_2.5 relative indicators at the overall pollution level passed the test of significance (

Table 2. Regression model of principal factors at the overall pollution level.

Dependent Variable	Principal Factors	Constant	p-Value	F-Value	Adj_R2
Cin	P3(0.058,0.429) ***, P4(0.060,0.361) **, P20(−0.521,−0.340) **	1.424 ***	0.002	6.492	0.347
Δtin	P1(−0.037,−0.182) *, P3(0.142,0.391) ***, P4(0.089,0.199) *, P5(−0.146,−0.283) **, P7(−0.118,−0.189) *, P13(−0.367,−0.251) **, P15(0.353,0.191) *, P16(−0.586,−0.284) **, P17(0.654,0.246) **, P18 (−0.729, −0.227) **, P22 (3.749, 0.295) **	7.725 ***	0.000	6.962	0.679
Cin’	P11(−0.021,−0.416) **, P16(0.027,0.298) *, P18(0.041,0.286) *	0.199 ***	0.015	4.128	0.232
Cde	P4(−0.010,−0.354) ***, P5(0.017,0.532) ***, P12(0.020,0.287) **, P16(0.045,0.346) ***, P21(−0.118,−0.256) **	0.539 ***	0.000	10.520	0.606
Δtde	P1(−0.053,−0.269) ***, P2(−0.069,−0.253) ***, P3(0.079,0.222) **, P4(0.200,0.457) ***, P5(−0.154,−0.302) ***, P10(0.150,0.153) *, P13(0.430,0.299) ***, P15(0.713,0.393) ***, P17(0.788,0.301) ***, P18(0.449,0.142) *, P20(−0.692,−0.172) **	6.508 ***	0.000	13.228	0.813
Cde’	P3(−0.002,−0.265) ***, P4(−0.003,−0.374) ***, P5(0.004,0.415) ***, P10(−0.004,−0.194) **, P13(−0.010,−0.327) ***, P15(−0.012,−0.326) ***, P16(0.008,0.176) *, P17(−0.021,−0.385) ***, P18(−0.017,−0.251) ***	0.100 ***	0.000	12.783	0.774

Note: ***, **, and * indicate that the factors passed the test of significance at 1%, 5%, and 10%, respectively, and the numbers in brackets indicate the regression coefficient and standardization coefficient, respectively.

). However, the principal factors included in the six models were different, indicating the complex impacts of different principal factors on the range, duration, and rate of PM_2.5 increase or decrease. The number of principal factors included in the regression model was 3~11. The more principal factors were included in the regression model, the adj_R² value was relatively higher. In these models, P₃, P₄, P_13, and P₁₇ were the four principal factors that appeared more frequently, indicating their significant effects on the increase/decrease in PM_2.5. Overall, these principal factors can explain approximately 60.6~81.3% of the PM_2.5 reduction indicators but only approximately 23.2~67.9% of the PM_2.5 increase indicators.

PM_2.5 indicator regression models at different pollution levels also passed the test of significance (Table S6). First, although there were great differences in the principal factors included in the different models, some principal factors had a high frequency and great impacts on PM_2.5. However, these principal factors varied based on pollution levels. For example, P₃, P_1, and P₁₆ were important principal factors affecting the relative indicators of PM_2.5 at slight, moderate, and heavy pollution levels, respectively. Second, the explanation degree of these principal factors for different PM_2.5 indicators showed a similar trend at different pollution levels. The explanation degree of C_in was higher than that of C_in’, and Δt_in was generally between them. The explanation degree of C_de was lower than that of C_de’, and Δt_de was often in between. Nonetheless, the explanations of these principal factors for PM_2.5 increase and decrease indicators were different. At the slight pollution level, these principal factors explained relatively more (approximately 52~81%) of the PM_2.5 decrease indicators and less (approximately 16~49%) of the PM_2.5 increase indicators. At the moderate pollution level, the principal factors had a higher explanation for PM_2.5 increase indicators (approximately 70~84%), while the explanation for PM_2.5 decrease indicators was lower (approximately 60~62%). At the heavy pollution level, the difference in the explanation of PM_2.5 increase and decrease indicators by principal factors narrowed, focusing on 60~75%.

The prediction accuracy of the verification neighborhood samples was calculated for validation. Figure 4 shows the comparison between the predicted value and the actual value of PM_2.5 indicators. Generally, the predicted value and actual value of each PM_2.5 indicator were similar. There were individual samples with great differences between the predicted value and the actual value, which were mainly at heavy pollution, followed by moderate pollution. Furthermore, the accuracy of different PM_2.5 indicators of five verification samples was compared through the RE value via Equation (3). At the overall pollution level, the RE of the six PM_2.5 indicators of each verification sample was mostly less than 10%. The prediction error of different verification samples had great randomness. The maximum prediction error was C_in (33.3%) of sample HZ4, and the minimum was Δt_de in WH4, whose RE was 0.3%. At different pollution levels, the trend of prediction error of verification samples increased with the increase in pollution level, and the prediction error of each verification sample varied greatly.

The prediction error was relatively low at the overall pollution level because more days of data (24 days) were used for testing. With the increase in pollution level, the number of days used for verification was less, which is vulnerable to accidental or sudden factors outside the built environment, resulting in an increase in prediction error. Based on these models, although the short-term prediction error is unstable, it can still achieve high prediction accuracy for the long-term PM_2.5 change trend of the block. Therefore, it has high application value.

Data with fewer days used in the analysis usually lead to a lower R² and a greater RE. In this study, due to the limited number of heavy pollution days, only 3 days of data were used for analysis at this pollution level. However, 4-day PM_2.5 data and 3-day data monitored by instruments were used to analyze the effects of urban lake wetlands, neighboring urban greenery, and plant communities on PM_2.5, respectively [42,43]. There may be some accidental factors influencing the results by limited data, but it is enough for analysis.

3.3. The Influence of Built Environment on PM_2.5

Because the included principal factors varied from model to model (Table 2 and Table S6), the most important factors influencing PM_2.5 were obtained by counting the times of principal factors that significantly influenced PM_2.5 in 24 models of four pollution levels. As Figure 5 shows, the darker the color, the greater the frequency of a principal factor. First, C_in, Δt_in, and C_in’ were synthesized as the increase change of PM_2.5. The decrease change of PM_2.5 included C_de, Δt_de, and C_de’. The principal factors that had the most significant impact on the increase change of PM_2.5 were P₁ and P₃, which occurred seven times, followed by P₄, P₆, P₁₆, and P₁₇, which occurred six times. P₇, P_12, and P₁₈ occurred five times. The most significant principal factor affecting the decrease change of PM_2.5 was P₅, which occurred 10 times, followed by P₃ and P₄, which occurred eight times. The frequency of P₁, P₁₂, P₁₃, P_15, and P₁₆ was six times. Overall, P₁, P₃, P_4, and P₁₆ were important factors that significantly affected the growth and reduction of PM_2.5 at the same time. These principal factors reflect the differences in green and gray space, building height and its differences, relative humidity, openness, and other characteristics of the neighborhood. In addition, these principal factors with high frequency were the factors that contributed the most to the corresponding dependent variable in the model. This further indicated their important role in PM_2.5.

Second, each PM_2.5 indicator was analyzed separately to find the differences in their principal factors. For C_in, P₄ and P₆ had the most contributions, indicating that meteorological factors and the green corridor connecting green space internally would greatly influence PM_2.5 increase change. P₁ contributed the most to Δt_in. However, there were no obvious principal factors for C_in’ compared with C_in and Δt_in. As for C_de, P₅ and P₁₆ had the most contribution to it. Moreover, P₅ showed a more important role in Δt_de than C_de. Therefore, attention should be given to green space coverage and openness of neighborhood for PM_2.5 reduction. P₄ was the most important principal factor for C_de’.

To strengthen the application of the regression model based on the built environment of the neighborhood and provide a reference for the optimization strategy of the built environment, we selected several neighborhoods from five cities with strong and weak PM_2.5 reduction effects. A neighborhood with strong PM_2.5 reduction effects is characterized by a high value of C_de and C_de’ or a low value of Δt_de. Their effects on the increase and decrease of PM_2.5 were analyzed through the analysis of the scale and spatial form of green and gray spaces in the neighborhood.

Neighborhoods with strong PM_2.5 reduction capacity included WH6, HF5, NJ4, SH7, and HZ7 (Figure 6a). On the one hand, WH6 and NJ4 have large-scale green spaces, which play a great role in promoting the adsorption and reduction of PM_2.5 [44]. Meanwhile, the lower building density contributes to the diffusion of PM_2.5 in the neighborhood and promotes the decrease in PM_2.5 [45]. HZ7 has a stable regulatory effect on PM_2.5. It can both inhibit the increase in PM_2.5 and promote the decrease in PM_2.5. Although the building density of HF5 and SH7 is higher than that of other neighborhoods, the almost determinant building layout and height arrangement of HF5 are conducive to the formation of a ventilation corridor. The large building height difference and building shape uniformity index of SH7 promote the decline of PM_2.5.

Neighborhoods with weak PM_2.5 reduction capacity included WH2, HF7, NJ1, SH3, and HZ3 (Figure 6b). Among these neighborhoods, WH2, NJ1, and HZ3 have a common feature of low-rise and high-density, which is not conducive to ventilation in the neighborhood. The building density of SH3 is high, and there are some high-rise buildings. However, the scale of green space in these neighborhoods is generally small, which reduces the active adsorption of PM_2.5 by green space. HF7 is a high-rise, low-density neighborhood. Too many high-rise buildings aggravate the pollution of PM_2.5 in the neighborhood and are difficult to evacuate.

4. Conclusions

This paper constructed a prediction model of PM_2.5 increase and decrease dynamic changes in neighborhoods based on 22 factors of green space, gray space, and meteorological factors, revealing the comprehensive impact mechanism of neighborhood-level built environments on PM_2.5 and laying a foundation for proposing specific optimization strategies. The adj_R² of these models was concentrated in 0.6~0.8, with the highest value of 0.836, indicating that it can better fit the existing indicators. P₁, P₃, P_4, and P₁₆ were the most important factors that significantly affected the increase and decrease in PM_2.5 at the same time, which reflected the characteristics of the green–gray space difference, building height and its difference, relative humidity, and openness, respectively. Among the many indicators of green space, green space coverage was significantly conducive to PM_2.5 reduction. The green corridor connecting green space internally would greatly influence the change of PM_2.5 increase. For gray space, the openness of the neighborhood was important to PM_2.5 reduction. In addition, relative humidity and wind speed contributed more to the change of both the decrease and increase of PM_2.5 than temperature.

There are some issues that need to be further explored. This study used PM_2.5 monitoring data for analysis, lacking exploration of PM_2.5 sources. Further study requires discussion on the emissions in the urban zones because of the different pollution sources in the five cities [46]. Data from more days, especially heavy pollution levels, can be included for more accurate research.