Scale Effects and Regional Disparities of Land Use in Influencing PM_2.5 Concentrations: A Case Study in the Zhengzhou Metropolitan Area, China

Abstract

Land use has been demonstrated to have an important influence on PM_2.5 concentrations; however, how the scale effects and regional disparities in land use influence PM_2.5 concentrations remains unclear. This study investigated the scale differences in spatial variations in PM_2.5 concentrations, in spatial associations between PM_2.5 concentrations and land use, and explored the effects of the spatial heterogeneity and action scale of land use on PM_2.5 concentrations. The main findings indicated greater intra-unit variation at small scales and greater inter-unit variation at large scales. PM_2.5 concentrations had a positive association with the surrounding cultivated land and artificial surface, and had a negative association with surrounding forest and grass; the positive spatial association between PM_2.5 concentrations and the surrounding artificial surface was stronger at small scales. Cultivated land and forest negatively influenced PM_2.5 concentrations generally. Artificial surfaces showed a strong positive influence on PM_2.5 concentrations in most urban areas. The action scale of cultivated land in influencing PM_2.5 concentrations was the largest (4698.05 m). The findings provide a new interpretation of the relationship between PM_2.5 concentrations and land use, and may contribute to effective policy making from the perspective of land use planning to PM_2.5 pollution control and abatement.

1. Introduction

Air pollution is an important environmental problem globally. In many large cities and metropolitan areas, air pollution is a serious concern that has long been a focus of both the government and the public [1,2,3]. PM_2.5 is a kind of major air pollutant; it refers to particulate matters whose aerodynamic diameter is less than 2.5 microns. Land is the spatial carrier of emission sources and factors influencing PM_2.5 concentrations, and is considered to have an important influence on PM_2.5 concentrations [4,5]. Rapid urban expansion and a transformed land use structure, together with the distribution of emission sources and local meteorological conditions within the rural–urban territorial system, have had significant effects on regional or global patterns of PM_2.5 concentration [6,7]. China’s vast urban expansion is advancing and is expected to result in new influences on the air environment. Moreover, China’s government has proposed the ambitious goal of essentially eliminating heavy pollution during the 14th Five-Year Plan (2021–2025). Thus, a thorough and systematic investigation of the relationship between PM_2.5 concentrations and land use is urgently needed.

Multidisciplinary researchers have conducted numerous studies on the relationship between PM_2.5 concentrations and different land use types. Artificial surfaces or construction land concentrates many primary and secondary anthropogenic emission sources [8] and is positively correlated with PM_2.5 concentrations [9,10,11]. Large-area artificial surfaces cause urban heat islands, which can further transport air pollutants to metropolitan areas and worsen urban air pollution [12]. Forest is generally considered to have an inhibitory effect or negative influence on PM_2.5 concentrations [13,14,15,16]. However, inconsistent findings have been reported regarding the influences of cultivated land and grassland on PM_2.5 concentrations. For example, Lin et al. (2020) and Lu et al. (2018) have found that cultivated land and grassland have a significant negative influence on PM_2.5 concentrations on the national scale [10,17], whereas Nguyen et al. (2015) have reported that grassland cannot control suspended particulate matter in Beijing [18], and Hong et al. (2021) have found that PM_2.5 concentrations are negatively correlated with cultivated land and positively correlated with grassland in the Pearl River Delta Region, China [9]. Furthermore, Yang et al. (2018) have investigated the relationship between PM_2.5 concentrations and integrated land ecosystems, and documented that land ecosystems are second only to climate among the major factors driving China’s spatial pattern of PM_2.5 concentrations [19]. Previous studies have provided a basis for understanding the relationship between PM_2.5 concentrations and land use.

However, multiscale characteristics are common in geospatial or ecological patterns [20,21], and the relationships between geographical elements also have potential scale differences [22]. As a geographical phenomenon, the spatial distribution of PM_2.5 concentrations shows significant multiscale differences, such as regional differences [13], rural-urban differences [23], and intra-city differences [8,24]. These multiscale spatial differences may potentially have relationships with land use and its spatial patterns and structures. In this regard, Feng et al. (2017) have examined the correlation of landscape indices with PM_2.5 concentrations in buffers with different distances and assessed the differences among six urban agglomerations in China [13]. Multiscale and regional differences in the relationships between PM_2.5 concentrations and different land use types remain important scientific problems worthy of further investigation.

In this study, we took the Zhengzhou metropolitan area, an emerging metropolitan area located in the polluted Northern China, as the study region. We focused on scale effects and regional differences to explore the influences of land use on PM_2.5 concentrations. We investigated the scale differences in spatial variations in PM_2.5 concentrations and in spatial associations between PM_2.5 concentrations and land use using semi-variogram and bivariate indicators of spatial associations at different scales, respectively. We then explored the effects of the spatial heterogeneity and action scale of land use on PM_2.5 concentrations by constructing a Multiscale Geographically Weighted Regression (MGWR) model. The findings provide a new interpretation of the relationship between PM_2.5 concentrations and land use. The main findings and implications are described in the discussion and conclusions sections.

2. Materials and Methods

2.1. Study Area

Zhengzhou metropolitan area is an emerging metropolitan area in inland China. It is located in the central region of the North China Plain and consists of five cities—Zhengzhou, Kaifeng, Xinxiang, Jiaozuo, and Xuchang—and 30 counties or county-level cities (Figure 1). Among the five cities, Zhengzhou is one of the nine national central cities of China. The metropolitan area was proposed in the “Development Planning of Central Plains Urban Agglomeration” and approved by the State Council of China in 2016. It covers a total area of 31,226 km². In recent years, this area has experienced rapid economic development and urbanization. In 2020, the total population of the metropolitan area was 31.61 million and the GPD reached CNY 2296.22 billion. This area is located in the most severely air-polluted area of China, in northern China. In 2020, the annual average PM_2.5 concentrations in Zhengzhou, Kaifeng, Xinxiang, Jiaozuo, and Xuchang were 51 μg/m³, 55 μg/m³, 52 μg/m³, 56 μg/m³, and 53 μg/m³, respectively. Furthermore, Zhengzhou, Kaifeng, Xinxiang, and Jiaozuo are air pollution transmission channel cities of Beijing-Tianjin-Hebei; therefore, strict policies on the discharge of air pollutants should be implemented.

2.2. Data and Processing

PM_2.5 data were gathered from the NASA Socioeconomic Data and Applications Center, which released the global annual PM_2.5 grids from MODIS, MISR, and SeaWiFS aerosol optical depth, v4.03 (1998–2019) [25]. This data set is aerosol optical depth retrievals from multiple satellite algorithms and calibrated based on Aerosol Robotic Network (AERONET) observations [25,26]. The total column measures of aerosol were associated with near-surface PM_2.5 concentrations with the GEOS-Chem chemical transport model and were calibrated with the Geographically Weighted Regression (GWR) model on the basis of global ground-based measurements [26]. We calculated annual PM_2.5 concentrations based on PM_2.5 concentrations from 1 January 2019 to 31 December 2019 at 42 monitoring sites and extracted PM_2.5 concentrations from the global grid annual PM_2.5 concentrations produced in 2019. We preliminarily verified the grid data using the unitary linearity regression. The derived verified R² is 0.71 and the statistical slope is 0.51 (lots of missing values during the heavy pollution period in December resulted in the low slope). In addition, Hammer et al. (2020) verified the data and documented that the annual PM_2.5 estimates were highly consistent with global ground monitors (R² = 0.81; slope = 0.90). The data comprise annual average PM_2.5 concentrations with a spatial resolution of 0.01 × 0.01° and were distributed as GeoTIFF files. The most recent data are the annual average PM_2.5 concentrations for 2019. We extracted a subset of the global PM_2.5 concentrations. The average, maximum, and minimum value of PM_2.5 concentrations is 64.95 μg/m³, 81 μg/m³, and 49 μg/m³ in 2019 in the study area.

Land use data were gathered from a global geo-information public product from GlobeLand30 (http://www.globallandcover.com/, accessed on 11 September 2020). This global land cover data product has a 30-m resolution. The most recent version of the data is GlobeLand30 2020. The data include 10 land cover classes: cultivated land, forest, grassland, shrubland, wetland, water bodies, tundra, artificial surface, bare land, and perennial snow and ice. The main images used to develop GlobeLand30 data were 30-m multispectral images, including TM5 ETM+ OLI multispectral images from Landsat, the Chinese Environmental Disaster Mitigation Satellite (HJ-1) multispectral images, and GF-1 multispectral images with 16-m resolution. The total accuracy of GlobeLand30 2020 is 85.72% and the kappa coefficient is 0.82.

The study region has eight types of land use: cultivated land, forest, grassland, shrubland, wetland, water bodies, artificial surface, and bare land. Referencing common land use/cover primary classes, we reclassified the above land use classes by merging forest and shrubland as forest, and merging wetland and water bodies as waters. To obtain PM_2.5 concentrations and the areas of different land use types at different scales, we created grids from 1 km to 10 km, 12 km, 15 km, and 20 km. We extracted the PM_2.5 concentration at each grid point of 1 km using the tool to extract multi values to points, and we extracted PM_2.5 concentrations at other scales and the area of each land use type at all scales using the zonal statistics in ArcMap software.

2.3. Methods

2.3.1. Nugget and Sill

In geostatistics, a semi-variogram is used to measure similarity changes with distance. Generally, it presents an increasing curve over distance. The nugget, sill, and range are the three key parameters of semi-variograms, which depict the variations in observed values at different distances. In general, a semi-variogram ($\gamma \left(h\right)$) is calculated with the following formula:

$$\gamma \left(h\right)=\frac{1}{N\left(h\right)}{\displaystyle \sum }_{i=l}^n{\left[Z\left(x_i\right)-Z\left(x_i+h\right)\right]}^2$$

(1)

where h is distance (lag distance), N(h) is the number of point pairs within distance h, $x_i$ and $x_i+h$ represent the locations of one point pair, and $Z\left(x_i\right)$ and $Z\left(x_i+h\right)$ are the value of the regional variable at locations $x_i$ and $x_i+h$, respectively.

With the above formula, a series of discrete semi-variogram estimates can be calculated. To obtain the nugget and sill, a continuous empirical semi-variogram must be fitted. The common empirical semi-variogram models include the spherical model, exponential model, and Gaussian model. We used the spherical model to fit the semi-variogram. Its form is as follows:

$$\gamma \left(h\right)=\left\{\begin{array}{c}0\\ C_0+C\left(\frac{3h}{2\alpha }-\frac{h^3}{2{\alpha }^3}\right)\\ C_0+C\end{array}\right.,\begin{array}{c}h=0\\ 0<h\le a\\ h>a\end{array}$$

(2)

where $C_0$ is the nugget, $C_0+C$ denotes the sill, and $\alpha$ is the range.

Referencing Chen et al. (2019) [27], we used the nugget and sill to examine intra-unit variation and inter-unit variation at different scales simultaneously.

2.3.2. Bivariate Spatial Association

Bivariate spatial association is an extension of spatial correlation analysis [28]. It calculates a bivariate Moran I and can be used to examine spatial associations and explore the patterns of spatial aggregation between two variables. Its formula is:

$$I_{kl}^i=Z_k^i{\displaystyle \sum }_{j=l}^n{W}_{ij}{Z}_l^i$$

(3)

where W_ij is the spatial weight matrix, $Z_k^i=\left[x_k^i-\overline{x_k}\right]/{\sigma }_k$, $Z_k^i=[x_k^j-\overline{x_l}]/{\sigma }_l$, $x_k^i$, and $x_l^j$ represent variables k and l at locations i and j, and ${\sigma }_k$ and ${\sigma }_l$ represent the variance of x_k and x_l, respectively.

In general, I values range from −1 to 1 and reflect the spatial association between two variables. Higher positive values of I indicate stronger positive associations between two variables, whereas lower negative values indicate stronger negative associations between two variables.

2.3.3. Multiscale Geographically Weighted Regression

The Multiscale Geographically Weighted Regression (MGWR) model is an extension of the well-known spatially varying-coefficient regression model, GWR [29]. The difference between MGWR and GWR is that MGWR improves the bandwidth (bandwidth is a key parameter in GWR, MGWR, and other GWR models; it means the spatial extent of data used to estimate a statistic. It can be regarded as the action scale or spatial scale of an independent variable significantly influencing a dependent variable) selection. In MGWR, different variables can use different bandwidth values to better reflect the spatial heterogeneity between variables and improve the accuracy of regression analysis. Thus, the model can capture scale differences in the relationships between the dependent variable and different independent variables [22]. The formula of MGWR is:

$$Y_i={\beta }_{bw,0}\left(u_i,v_i\right)+{\displaystyle \sum }_m{\beta }_{bw,im}\left(u_i,v_i\right)X_{im}+{\epsilon }_i i=1,\cdots ,n$$

(4)

where ${\beta }_{bw,m}$ represents the regression coefficient of variable m when the bandwidth is bw, (u_i, v_i) is the spatial position i of the variable, and ${\beta }_{bw,0}\left(u_i,v_i\right)$ is the local intercept.

An iterative back-fitting algorithm is generally used to solve coefficients of the MGWR model [29]. The convergence criterion is the score of change in the residual sum of squares (RSS). Its formula is:

$$SO{C}_{RSS}=\frac{\left|RS{S}_{new}-RS{S}_{old}\right|}{RS{S}_{new}}$$

(5)

where $RS{S}_{old}$ represents the RSS derived in the previous iteration and $RS{S}_{new}$ is the RSS in the current iteration.

In this study, we used the R package GWmodel to calculate the MGWR results [30].

3. Results

3.1. Scale Effects of Spatial Variations in PM_2.5 Concentration

On the basis of the PM_2.5 concentrations at all scales described above, we fitted the semi-variograms using the spherical model to examine the scale effects of spatial variations in PM_2.5 concentrations. The derived nugget and sill are shown in Figure 2. The nugget generally showed a decreasing tendency, thus, indicating obvious and large local small-scale differences (intra-unit variation). Moreover, an increased tendency was observed in scales 1–6 km, and slight fluctuations were observed from 6 km to 20 km, thus indicating that spatial differences between measuring units (inter-unit variation) increased with scale in 1–6-km scales and were larger and relatively stable when the measurement units were larger than 6 km. In metropolitan areas or large cities, compact traffic or industrial sites can cause high local pollution, whereas PM_2.5 concentrations are low in green spaces, parks, and schools. Thus, spatial heterogeneity in emissions leads to small-scale differences in metropolitan areas. Urban areas have concentrated industry, transportation, and energy consumption; moreover, urban heat island effects can cause atmospheric pollutants to collect outside cities in the main urban area. The potential differences between central urban and suburban or rural–urban differences should be obvious, as revealed by the sill at large scales.

3.2. Scale Differences in Spatial Associations between PM_2.5 Concentrations and Land Use

From Figure 1, we can find that land use shows discrepant spatial distribution in the study region. Cultivated land has the largest area in the region, followed by artificial surface and forest. Artificial surface is mainly distributed in urban areas of cities, county-level cities, and counties, especially in the urban areas of Zhengzhou. Forest is mainly distributed in mountainous area in the northwest and west. There is some grassland, which is mainly distributed adjacent to forest. There is a strip of water in the middle of the region; it is the Yellow River.

Land use has been demonstrated to have direct or indirect effects on PM_2.5 concentrations and their spatial patterns. We revealed the scale differences in PM_2.5 concentrations. We next sought to determine the relationships between each land use type and PM_2.5 concentrations in a metropolitan area, particularly the spatial association and its scale differences. We calculated the bivariate Moran I of each land use type and PM_2.5 concentration at different scales. The area of waters and bare land was low, so these two types of land use were not analyzed. Figure 3 shows the results on the scale differences in the bivariate Moran I between PM_2.5 concentrations and the area of cultivated land, forest, grassland, and artificial surface. The bivariate Moran I of PM_2.5 concentrations and the area of cultivated land and artificial surface were positive at all scales, whereas the bivariate Moran I of PM_2.5 concentrations and the area of forest and grass were negative at all scales. These findings suggested that high PM_2.5 concentrations are associated with adjacent cultivated land and artificial surface at all scales, whereas low PM_2.5 concentrations are associated with adjacent forest and grass at all scales.

At 1–3-km scales, the bivariate Moran I of PM_2.5 concentrations and the area of artificial surface were larger but showed a continuous decrease at 3–20-km scales, thus indicating that PM_2.5 concentrations have a stronger positive association with adjacent artificial surface at small scales and have a weaker association with adjacent artificial surface with increases in scale. The bivariate Moran I of PM_2.5 concentrations and the area of cultivated land generally increased with increasing scale, thus indicating that PM_2.5 concentrations have relatively stronger positive associations with adjacent cultivated land at larger scales. The bivariate Moran I of PM_2.5 concentrations and the area of forest were negative with larger absolute values and minor changes, thus indicating that PM_2.5 concentrations have relatively stronger negative associations with adjacent forest at all scales. The bivariate Moran I of PM_2.5 concentrations and the area of grassland generally decreased with increasing scale, thus indicating that PM_2.5 concentrations had a relatively stronger negative association with adjacent grassland at larger scales.

3.3. Influences of Regional and Scale Differences in Land Use on PM_2.5 Concentrations

Land use types vary among regions and have underlying relationships with the spatially discrepant pattern of PM_2.5 concentrations. The relationships between PM_2.5 concentrations and different land use types may also have scale differences, and the influences of the scale of land use on PM_2.5 concentrations may also differ. Therefore, we constructed an MGWR model to examine how the spatial heterogeneity of these relationships and the scale differences in land use influence PM_2.5 concentrations. Because of excessive null values for some land use types at 1-km and 2-km scales, we integrated the log of PM_2.5 concentrations as the dependent variable and the log of the area for each land use type at the 3-km scale as the independent variable. Because zero values were present for some land use types, we pretreated the data by adding 0.0001 to all original values for each land use type in calculating the logarithm. The general statistics for estimated coefficients are shown in

Table 1. Summary of the MGWR coefficient estimates.

Variable	Min	1st Quantile	Median	Mean	3rd Quantile	Max
Intercept	3.9302	4.1250	4.1668	4.1577	4.2025	4.3184
Cultivated land	−0.0333	−0.0058	−0.0019	−0.0030	0.0003	0.0116
Forest	−0.0151	−0.0045	−0.0025	−0.0030	−0.0009	0.0052
Grassland	−0.0175	−0.0010	0.0003	0.0007	0.0023	0.0175
Artificial surface	−0.0355	−0.0007	0.0013	0.0030	0.0051	0.0531

. The coefficients of cultivated land and forest were mainly negative, thus indicating that increases in the area of cultivated land and forest are conducive to lower PM_2.5 concentrations in this metropolitan area. The coefficients of grassland and artificial surface were mainly positive, thereby indicating that increases in the area of grassland and artificial surface are associated with greater PM_2.5 concentrations.

Figure 4 shows the specific spatial distribution of the MWGR coefficients for the above four variables. As shown in Figure 4a, cultivated land has negative effects in most regions. In several regions, mainly those far from the main urban areas of the five cities, such as the north and east of Xinxiang, the east and southwest of Kaifeng, the east of Xuchang, the border between Jiaozuo and Xinxiang, and the southeast of Zhengzhou, cultivated land showed positive effects on PM_2.5 concentrations. Forest showed negative effects in a larger area (Figure 4b) and showed positive effects mainly in the north of Kaifeng; the southeast of Xinxiang; and the border areas of Zhengzhou, Kaifeng, and Xuchang. Grassland showed a strong negative influence mainly in the southeast of the metropolitan area and showed a positive influence in the northwest of the region; the border region of Kaifeng and Xinxiang; and the border region of Zhengzhou, Kaifeng, and Xuchang (Figure 4c). In some areas, such as those along Zhengzhou’s borders with Jiaozuo and Xinxiang, and most regions in the southwest of the study area, the coefficients of grassland were close to 0, thus indicating that the influence of grassland on PM_2.5 concentrations was not sensitive and was weak in these regions. Figure 4d shows the coefficients of artificial surface, indicating a strong positive influence in most main urban areas of the five cities and their governed counties and county-level cities, particularly in Xinxiang, Jiaozuo, and some counties and county-level cities in the two cities. Along the regions of Zhengzhou and Kaifeng’s border with Jiaozuo and Xinxiang, which is crossed by the Yellow River, and in some regions in the southwest of Zhengzhou, the coefficients of artificial surface were close to 0, thus indicating that PM_2.5 concentrations were not sensitive to artificial surface and were relatively less affected by artificial surface in these regions.

The bandwidths for each variable derived from the MGWR model and the model diagnostic are shown in

Table 2. Bandwidth and diagnostics for MGWR and GWR.

Arguments	MGWR	GWR
Cultivated land	4698.05 (m)	11,050.98 (m)
Forest	3285.87 (m)	11,050.98 (m)
Grassland	3630.26 (m)	11,050.98 (m)
Artificial surface	2297.57 (m)	11,050.98 (m)
AIC	−26,879.98	−17,467.29
RSS	0.1015	1.8842
Adjusted R2	0.9863	0.8824
R2	0.9940	0.8893

. The bandwidths directly reflect the differing scale of action of the different variables. The bandwidth of cultivated land was the largest (4698.05 m), thus indicating that the scale of action of cultivated land was larger than that of other land use types, and the coefficient of cultivated land was relatively stable in space. The bandwidth of artificial surface was the smallest (2297.57 m), thus indicating that the scale of action of artificial surface was relatively small, and PM_2.5 concentrations would present significant spatial heterogeneity with changes in artificial surface. In contrast to the GWR model, which uses only a single bandwidth for variables (Table 2), the MGWR model can use different bandwidths to calculate the coefficients for variables and, consequently, can achieve a better fitting result. The diagnostic comparison of MGWR and GWR indicated that the MGWR model had better results in terms of AIC, RSS, and (adjusted) R². Furthermore, we also constructed the MGWR model based on PM_2.5 concentrations and land use at the 5-km scale. We derived similar results for the bandwidths and coefficients of each variable. The bandwidth of cultivated land was also the largest (7052.08 m), the bandwidth for artificial surface was also the smallest (3815.00 m), and the bandwidths for forest and grassland were close at 5277.32 m and 5216.83 m, respectively.

4. Discussion

From the perspective of scale effects and regional disparities, this study reanalyzed the influences of land use on PM_2.5 concentrations in the Zhengzhou metropolitan area. The nugget and sill at different scales reflected the scale differences in spatial variations in PM_2.5 concentrations. The intra-unit variation at 1–3-km scales was more obvious than that at larger scales, and the inter-unit variation at 6–20 km was greater and showed small fluctuations with changes in scale. The results of bivariate Moran’s I showed that PM_2.5 concentrations were positively associated with adjacent area of cultivated land and artificial surface, and negatively associated with adjacent area of forest and grass at all scales; at small (1–4 km) scales, PM_2.5 concentrations had a stronger positive association with adjacent artificial surface. The MWGR coefficients showed the discrepant regional influences and scale of action of each land use type on PM_2.5 concentrations. Cultivated land showed the largest scale of action (4698.05 m), whereas artificial surface showed the smallest scale of action (2297.57 m).

Theoretically, spatial pattern or surface combines large-scale spatial trends and small-scale variations, i.e., average process and residuals, in spatial statistics [31]. Moreover, in scale effects, measurements or statistics at different scales lead to discrepant statistical inferences regarding both single factors and among multiple factors—this characteristic is essential in geospatial data [21,32]. On different scales, the spatial variation characteristics of PM_2.5 concentrations and the influences of land use should differ. Previous studies have examined mainly spatial variation in PM_2.5 concentrations and their relationships with land use or other factors on national [14,33], regional [17], or city [8,34] scales. In this study, we quantified and empirically confirmed the multiscale spatial variations in PM_2.5 concentrations and the scale differences in their relationship with land use. The bivariate Moran I showed the global spatial association between PM_2.5 concentrations and land use, and the results were generally consistent with correlation results in previous studies [13,35].

As expected, the MGWR coefficients differed across regions. Cultivated land was found to have mainly negative influences on PM_2.5 concentrations, and grassland was found to have positive influences on PM_2.5 concentrations on a large area. Similarly, Hong et al. (2021) have found that cultivated land is negatively correlated with PM_2.5 concentrations and grassland is positively correlated with PM_2.5 concentrations [9]. Furthermore, forest land was found to have positive influences on PM_2.5 concentrations in a small area. The study area is located in China’s main grain-producing areas, in a temperate deciduous broad-leaf forest climate zone. Lush vegetation, including forest, grass, and crops, can absorb particulate matter and prevent its spread. However, a diminished adsorption capacity, increased bare land in winter, and fertilizer use in cultivated land can increase PM_2.5 sources or facilitate the formation of PM_2.5 [35,36,37]. Notably, the scale of action of each land use type on PM_2.5 concentrations was determined. Although PM_2.5 can diffuse under the influence of meteorological conditions, multiscale and spatially differential influences of land use should determine the long-term spatial pattern of PM_2.5 concentrations.

The results on bivariate Moran I at different scales reflect the scale sensitivity of global spatial association between PM_2.5 concentrations and land use. The spatial association between PM_2.5 concentrations and cultivated land and grassland tend to be more sensitive at smaller scales (Figure 3). The spatial association between PM_2.5 concentrations and artificial surface tend to be more sensitive at larger scales (larger than 8 km). Meanwhile, the spatial association between PM_2.5 concentrations and artificial surface present weak sensitivity at 1–20-km scales. Combining the GWR’s result, we can find that the coefficients of cultivated land and artificial surface are more sensitive across space, while the coefficients of forest and grassland showed present weak sensitivity. Scale effects are relatively more sensitive to changes in cultivated land far from the main urban areas and changes in artificial surface in most main urban areas. As mentioned above, land is the spatial carrier of both the emission sources and influencing factors of PM_2.5 concentrations, such as agricultural emissions from cultivated land, adsorption, and retarding effect of forest. Land use types and their combinations differ significantly across space, the effects of land use are inevitably sensitive to changes in land use over space. In other words, the long-term spatial pattern of PM_2.5 concentrations, such as the spatial distribution of annual PM_2.5 concentration is highly correlated with spatially different land use. Meanwhile, daily or short-term anomalies or fluctuations are more likely to be related to short-term changes in meteorological conditions.

On the basis of the main findings, this study indicates the following implications. First, co-ordinating the “source-sink” relationship of PM_2.5 is important. Chemical fertilizer usage should be reduced and a farmland shelterbelt project can be implemented around main urban areas to reduce agricultural emissions and block agricultural emission sources to transfer and concentrate into cities. The spatial distribution of industries could be adjusted and high pollution enterprises could be placed adjacent to forest. Planting of tree species with strong dust retention ability, such as cedar, white bark pine, and Pinus tabulaeformis, should be expanded and preferentially located in main urban areas, particularly industrial parks and major traffic routes. Second, air pollution control projects may consider changing or optimizing existing grass species. Although some studies have confirmed that grassland cannot effectively adsorb suspended particles and may even have a positive relationship with PM_2.5 concentrations, grassland can decrease ground dust sources blown in the air [18]. Our results also showed that grassland has a negative influence on PM_2.5 concentrations in some regions. To avoid the bareness of grass and an increase in airborne dust during heavy pollution winter, long-evergreen or cold-season grass should be considered for increased coverage in both urban and rural areas. Furthermore, in the studies on PM_2.5 concentration estimation and simulation, land use and other covariates are often integrated in uniform spatial units [38,39,40], whereas the scale differences in the relationships of PM_2.5 concentrations with different covariates have been less considered. Distinguishing the scale effects of different covariates in modeling may improve the accuracy of estimates and is worthy of further study in the future.

5. Conclusions

In conclusion, this study performed a quantitative analysis of scale effects and regional disparities on the influences of land use on PM_2.5 concentrations in the Zhengzhou metropolitan area, central China. The main findings enrich understanding of the relationships among PM_2.5 concentrations and land use, and provide a theoretical basis for regional air pollution control and model construction of PM_2.5 concentration estimation. The findings and implications are expected to contribute to land use policy making aimed at PM_2.5 pollution abatement.