Non-Linear Response of PM 2.5 Pollution to Land Use Change in China

: Land use change has an important inﬂuence on the spatial and temporal distribution of PM 2.5 concentration. Therefore, based on the particulate matter (PM 2.5 ) data from remote sensing instruments and land use change data in long time series, the Getis-Ord Gi* statistic and SP-SDM are employed to analyze the spatial distribution pattern of PM 2.5 and its response to land use change in China. It is found that the average PM 2.5 increased from 25.49 µ g/m 3 to 31.23 µ g/m 3 during 2000-2016, showing an annual average growth rate of 0.97%. It is still greater than 35 µ g/m 3 in nearly half of all cities. The spatial distribution pattern of PM 2.5 presents the characteristics of concentrated regional convergence. PM 2.5 is positively correlated with urban land and farmland, negatively correlated with forest land, grassland, and unused land. Furthermore, the average PM 2.5 concentrations show the highest values for urban land and decrease in the order of farmland > unused land > water body > forest > grassland. The impact of land use change on PM 2.5 is a non-linear process, and there are obvious differences and spillover effects for different land types. Thus, reasonably controlling the scale of urban land and farmland, optimizing the spatial distribution pattern and development intensity, and expanding forest land and grassland are conducive to curbing PM 2.5 pollution. The research conclusions provide a theoretical basis for the management of PM 2.5 pollution from the perspective of optimizing land use.


Introduction
The city is the densest place for human activities and is the space where air pollutants are most likely to accumulate [1,2]. Since the 1980s, with the rapid urbanization process and social and economic development, peoples' material wealth and living standards have been improved in China, but this has also brought a series of environmental problems [3,4]. Moreover, it highlights the serious contradiction between accelerating the urbanization process and abiding by ecological and environmental protection. Developed countries such as Europe and the United States have experienced air pollution problems for more than 100 years. Pollution increased intensively in China's economically developed regions in the past 20-30 years [5][6][7][8], especially PM 2.5 (Particulate matter with aerodynamic diameter ± 2.5 µm) pollution [9][10][11][12]. At present, China has become one of the most polluted areas in the world in regards to PM 2.5 pollution [1,11]. It is particularly prominent in the Beijing-Tianjin-Hebei region to the east of Hu Line, Yangtze River Delta, Chengdu-Chongqing Economic Zone, Guanzhong Economic Zone, Central Plains, and Harbin-Changchun urban agglomeration. They have become the worst-hit areas [1,9]. Furthermore, because PM 2.5 can remain in the atmosphere for a longer time than PM 10 and total suspended particulate matter (TSP), and contains sulfates, nitrates, dust, polycyclic aromatic hydrocarbons, and heavy metals that are toxic to the human body [13], it significantly affects human that R 2 is 0.81 [9]. Van Donkelaar et al. estimate PM 2.5 by combining Aerosol Optical Depth (AOD) retrievals from the NASA's Medium Resolution Imaging Spectrometer (MODIS), Multi-angle Imaging Spectrometer (MISR), and Wide Field Ocean Observation Sensor (SeaWiFS) satellite instruments and coincident aerosol vertical profiles with the GEOS-Chem (http://geos-chem.org/, accessed on 10 March 2021) chemical transport model, and subsequently calibrated to regional ground-based observations using Geographically Weighted Regression (GWR) [49]. Ground-based PM 2.5 measurements over mainland China were obtained from http://beijingair.sinaapp.com/, accessed on 10 March 2021, Taiwanese PM 2.5 measurements were downloaded from https://taqm.epa.gov.tw/taqm/ tw/YearlyDataDownload.aspx. For further details, see references [50,51]. These have been effectively applied to national and regional scale air pollution studies [9,52].

Land Use Dataset
Land use data originate from the CCI-LC global land cover product developed by the European Aviation Agency (http://maps.elie.ucl.ac.be/CCI accessed on 4 June 2020, 2020). The dataset covers the world with a time span of 1992-2018, of which the 2000-2015 data format is TIFF, the 2016-2018 data format is netCDF, the spatial resolution is 300 m, and the coordinate system is WGS-1984. ArcGIS 10.3.1 software and Chinese administrative boundaries are employed to tailor CCI-LC products and extract the land use data during 2000-2016. As the CCI-LC maps are designed to be globally consistent, the type of land counts 22 classes, and each class is associated with a ten values code. In this study, we re-combine the types of land use according to research needs. The classification rules are: urban areas of CCI-LC maps are reclassified into new urban land; water bodies of CCI-LC maps are reclassified into new water bodies; grassland of CCI-LC maps is reclassified into new grassland; cropland, rainfed and cropland, irrigated or post-flooding of CCI-LC maps are merged into new farmland; permanent snow and ice, bare areas, lichens and mosses of CCI-LC maps are merged into unused land; and the rest are classified as forest land.

Hot Spot Analysis (Getis-Ord Gi*)
The hot spot analysis employed the Getis-Ord Gi* [53,54] statistic to produce a hot (high PM 2.5 values) and cold (low PM 2.5 values) spot PM 2.5 pollution map. It is widely used in the analysis of the spatial agglomeration of geographic features. Getis-Ord Gi* is formulated as: where x j is the PM 2.5 values for city j; w i,j is the spatial weight between city i and j, the w i,j are constructed using the rook's case method (polygon features that share a boundary are neighbors); X is the arithmetic average of x j ; and n is equal to the total number of cities. The Gi* statistic returned for each city is a z-score. For statistically significant (p-value) positive z-scores, the larger the z-score is, the more intense the clustering of high PM 2.5 values (hot spot). For statistically significant (p-value) negative z-scores, the smaller the z-score is, the more intense the clustering of low PM 2.5 values (cold spot). The p-value is a probability representing confidence levels [55], and in this study we set p-value equal to 0.05.

Semi-Parametric Spatial Durbin Model (SP-SDM)
The spatial lag model, spatial error model, and spatial Durbin model (SDM) are commonly used as spatial analysis tools [56,57]. As emissions from other neighboring areas affect PM 2.5 pollution levels, SDM can be applied to effectively measure its impact [31,56]. It can also be employed to measure the indirect effects of exogenous variables such as initial conditions and control variables in a region. Therefore, SDM is utilized in this manuscript to analyze the relationship between PM 2.5 and land use change. However, the homogeneity of the investigated parameters is assumed in the traditional SDM [58]. The effects of explanatory variables on the explained variables are the same in all regions, and the spatial heterogeneity and non-linearity of regional growth behavior cannot be recognized [59]. In order to measure whether there are spatial heterogeneity and non-linear characteristics, two different forms of spatial econometric models are constructed through the generalized additive models (GAM) [60] and SDM for comparisons. g 1 (µ) is an ordinary non-linear model, and g 2 (µ) is SP-SDM, with the calculation formula as follows: where g(µ) is a link function, µ = E(PM 2.5 /α 1 , α 2 , . . . . . . , α m ), β represents autoregressive coefficient, α i represents the land use type, W is n*n order spatial weight matrix, and K = 4, Wα i and WPM 2.5 is the α i and PM 2.5 spatial lag variables, respectively, ε is a random error, and s(·) is the smooth function of connecting explanatory variables. The spdep and mgcv packages in the R 3.6.3 version [61] (a free software environment for statistical computing and graphics) are used to calculated SP-SDM. The analysis showed that China's PM 2.5 governance had achieved certain results, but further efforts are needed to reduce PM 2.5 pollution and achieve sustainable development goals by 2035. get 2 in WHO transition period) was 17.25%-20.22%. The proportion of cities with an av-erage yearly PM2.5 concentration between 25 μg/m 3 and 35 μg/m 3 (IT-1: the annual average limit of target 1 in WHO transition period) was 23.18%-25.34%. The proportion of cities with an annual average PM2.5 concentration greater than 35 μg/m 3 was 43.40%-47.71%. The analysis showed that China's PM2.5 governance had achieved certain results, but further efforts are needed to reduce PM2.5 pollution and achieve sustainable development goals by 2035.

Spatial Distribution Pattern of PM2.5
As can be seen from Figure 2, the spatial distribution of PM2.5 shows concentrated regional convergence, with obvious spatial heterogeneity. High PM2.5 values are shown in the densely populated and relatively economically developed central and eastern regions and the Taklimakan Desert in Xinjiang. In contrast, low PM2.5 values are mainly distributed in the west & central areas and south Fujian with low population density and relatively backward economic development. Except for natural factors, the spatial distribution pattern of PM2.5 is roughly consistent with that of the population and economic patterns, indicating that human socio-economic activities significantly impact PM2.5 concentration.
The hot spot analysis method was employed to identify the hot spots (high PM2.5 values) and cold spots (low PM2.5 values) of PM2.5 distribution in China and further analyze the clustering characteristics of PM2.5. It is demonstrated that the spatial distribution of PM2.5 in China conforms to the characteristic analysis and clustering. Hot spots (high PM2.5 values) are mainly distributed in Kashgar, Aksu in western Xinjiang, the Beijing-Tianjin-Hebei region, Shandong Peninsula, the Central Plains, the middle reaches of the Yangtze River, the Yangtze River Delta, and other central and eastern urban agglomerations, especially the northern regions. They have experienced rapid industrialization and coal burning in winter, which has deteriorated air quality. Cold spots (low PM2.5 values) are mainly distributed in the north slope of the Tianshan Mountains, the Qinghai-Tibet Plateau, the west coast of the Taiwan Straits, the Yunnan-Guizhou Plateau, Hainan As can be seen from Figure 2, the spatial distribution of PM 2.5 shows concentrated regional convergence, with obvious spatial heterogeneity. High PM 2.5 values are shown in the densely populated and relatively economically developed central and eastern regions and the Taklimakan Desert in Xinjiang. In contrast, low PM 2.5 values are mainly distributed in the west & central areas and south Fujian with low population density and relatively backward economic development. Except for natural factors, the spatial distribution pattern of PM 2.5 is roughly consistent with that of the population and economic patterns, indicating that human socio-economic activities significantly impact PM 2.5 concentration.
The hot spot analysis method was employed to identify the hot spots (high PM 2.5 values) and cold spots (low PM 2.5 values) of PM 2.5 distribution in China and further analyze the clustering characteristics of PM 2.5 . It is demonstrated that the spatial distribution of PM 2.5 in China conforms to the characteristic analysis and clustering. Hot spots (high PM 2.5 values) are mainly distributed in Kashgar, Aksu in western Xinjiang, the Beijing-Tianjin-Hebei region, Shandong Peninsula, the Central Plains, the middle reaches of the Yangtze River, the Yangtze River Delta, and other central and eastern urban agglomerations, especially the northern regions. They have experienced rapid industrialization and coal burning in winter, which has deteriorated air quality. Cold spots (low PM 2.5 values) are mainly distributed in the north slope of the Tianshan Mountains, the Qinghai-Tibet Plateau, the west coast of the Taiwan Straits, the Yunnan-Guizhou Plateau, Hainan Province, and the border area between the three northeastern provinces and Inner Mongolia. From 2000 to 2016, most of the hot spots remained stable and only a few areas changed. The Chengdu-Chongqing Economic Zone and Lanzhou-Xining Urban Belt changed from hot spots to insignificant regions, indicating that air pollution has improved. The continuous expansion of cold spots (low PM 2.5 values) in the urban agglomerations on the west coast of the Taiwan Straits, central Yunnan, and central Guizhou indicates that significant results have been achieved in air pollution control in these regions.
The spatial distribution of PM 2.5 has strong heterogeneity due to different emission intensities of pollution gas and meteorological conditions in different areas of China, and different contributions by different types of PM 2.5 precursors. PM 2.5 pollutants in Xinjiang and Qinghai Qaidam Basin mainly came from sand dust aerosol. However, PM 2.5 pollutants in eastern China and urban agglomerations mainly came from anthropogenic emissions. The reaction between nitrogen dioxide and sulfur dioxide in water absorbed by Remote Sens. 2021, 13, 1612 6 of 13 PM 2.5 is the main formation path of sulfate during fog and haze. Nitrogen oxides not only lead to an increase of PM 2.5 concentration, but aerosols from high emissions of agriculture, industrial production, and airborne dust also results in the rapid generation of sulfates by their unique chemical pathways [38], which is one of the main reasons for the increased PM 2.5 concentration.  The spatial distribution of PM2.5 has strong heterogeneity due to different emission intensities of pollution gas and meteorological conditions in different areas of China, and different contributions by different types of PM2.5 precursors. PM2.5 pollutants in Xinjiang and Qinghai Qaidam Basin mainly came from sand dust aerosol. However, PM2.5 pollutants in eastern China and urban agglomerations mainly came from anthropogenic emissions. The reaction between nitrogen dioxide and sulfur dioxide in water absorbed by PM2.5 is the main formation path of sulfate during fog and haze. Nitrogen oxides not only lead to an increase of PM2.5 concentration, but aerosols from high emissions of agriculture, industrial production, and airborne dust also results in the rapid generation of sulfates by Land use space is the underlying surface of the atmospheric environment, which can directly or indirectly affect the temporal and spatial distribution pattern of PM 2.5 . We calculated the land use area of different types in each city from 2000 to 2016 by ArcGIS software. The Spearman correlation coefficient method [62] is employed to determine the correlation between PM 2.5 and land use types.
It is found that PM 2.5 was positively correlated with urban land, farmland, and water bodies, and the correlation coefficients were 0.34 (p = 0.01), 0.047 (p = 0.01), and 0.067 (p = 0.01), respectively. PM 2.5 was negatively correlated with forest land, grassland and unused land, and the correlation coefficients were −0.438 (p = 0.01), −0.265 (p = 0.01) and −0.441 (p = 0.01) respectively. Land use data were resampled to make its spatial resolution consistent with the grid PM 2.5 data. Then, the mean concentration of PM 2.5 on different land types was calculated by using the resampled land use data of 2016. As shown in Figure 3, the average PM 2.5 concentrations show the highest values for urban land and decrease in the order of farmland > unused land > water body > forest > grassland. Because the unused land includes sandy land, dust in the desert is the cause of high PM 2.5 concentrations. The highest value being under urban land indicates the city as a major source area of PM 2.5 in China.
ware. The Spearman correlation coefficient method [62] is employed to determine the correlation between PM2.5 and land use types.
It is found that PM2.5 was positively correlated with urban land, farmland, and water bodies, and the correlation coefficients were 0.34 (p = 0.01), 0.047 (p = 0.01), and 0.067 (p = 0.01), respectively. PM2.5 was negatively correlated with forest land, grassland and unused land, and the correlation coefficients were −0.438 (p = 0.01), −0.265 (p = 0.01) and −0.441 (p = 0.01) respectively. Land use data were resampled to make its spatial resolution consistent with the grid PM2.5 data. Then, the mean concentration of PM2.5 on different land types was calculated by using the resampled land use data of 2016. As shown in Figure 3, the average PM2.5 concentrations show the highest values for urban land and decrease in the order of farmland > unused land > water body > forest > grassland. Because the unused land includes sandy land, dust in the desert is the cause of high PM2.5 concentrations. The highest value being under urban land indicates the city as a major source area of PM2.5 in China.

Non-Linear Response of PM2.5 to Land Use Change
The constructed SP-SDM was employed to measure the non-linear characteristics of PM2.5 and land use change. To make the data more stable, logarithmic transformation was undertaken for land use types. The results are shown in Table 1. It was demonstrated that the two models have passed the 1% significance level test in terms of the estimated and referenced degrees of freedom. R 2 of model 2 is 0.86, which is significantly greater than that of model 1 (0.66). However, this shows the contrary values in Generalized Cross-Validation (GCV) [63]. The smaller the value of GCV, the better the performance of the model, indicating that the impact of land use change on PM2.5 is statistically significant, and the fitting degree of SP-SDM is superior to that of the ordinary non-linear model. It can also be seen from Table 1

Non-Linear Response of PM 2.5 to Land Use Change
The constructed SP-SDM was employed to measure the non-linear characteristics of PM 2.5 and land use change. To make the data more stable, logarithmic transformation was undertaken for land use types. The results are shown in Table 1. It was demonstrated that the two models have passed the 1% significance level test in terms of the estimated and referenced degrees of freedom. R 2 of model 2 is 0.86, which is significantly greater than that of model 1 (0.66). However, this shows the contrary values in Generalized Cross-Validation (GCV) [63]. The smaller the value of GCV, the better the performance of the model, indicating that the impact of land use change on PM 2.5 is statistically significant, and the fitting degree of SP-SDM is superior to that of the ordinary non-linear model. It can also be seen from Table 1 that the p-value of the spatial lag variables of PM 2.5 (WPM 2.5 ) is significant, indicating that PM 2.5 has a spatial spillover effect [56]. The degree of freedom for model 1 is greater than 1, showing that the function is a non-linear curve equation (when the degree of freedom is 1, the function is a linear equation) [60]. Moreover, the more significant non-linear relationship indicates the non-linear response of PM 2.5 concentration to land use changes.
The vertical axis in Figure 4 is the linear prediction value of PM 2.5 , and the two horizontal axes are the land use type and its spatial lag variables respectively, reflecting the change characteristics of PM 2.5 concentration under the interaction of different land use scales and its spatial lag. In China, water bodies are mostly shown as long and narrow strips. Except for the northwest's sandy land, the unused land has a small distribution area in other places, especially on the urban fringe. This causes these two types of land to be prone to be polluted by adjacent land types. Therefore, the visual mapping is only performed for urban land, farmland, forest land, and grassland. It can be seen from Figure 5 that the PM 2.5 concentration value increases with the expansion of urban land and farmland and decreases with the expansion of forest land and grassland, indicating that the urban land and farmland have a positive contribution to PM 2.5 pollution. Forest land and grassland have a negative contribution to PM 2.5 pollution. inhibitory effect in adjacent areas. Moreover, the larger the scale of forest land and grassland, the more obvious the inhibitory effect. The large-scale and green production of farmland will help curb PM2.5 pollution. The impact of land use scale on PM2.5 pollution is a non-linear process, and different land use scales have different effects. As shown in Figure 5, as the main driver for air pollution, different land use types, patterns, and development intensities will obviously result in different distribution patterns of PM2.5 pollution [64][65][66][67]. In fact, the expansion of urban scale and the increase in urban land area will lead to the imbalance of artificial and natural surface structures. In the process of this change, the land pattern has also changed. The construction of new residential areas, commercial areas, or industrial parks has improved the population density and commuting distance, thus increasing pollution emission sources. Changes in cities' landscape pattern structure have also changed the microclimate environment, which is prone to the heat island effect and increased water evapotranspiration [68]. In addition, the increase in land use and development intensity, building density, and height enable This is consistent with the above correlation analysis results. Generally, PM 2.5 decreases with the increase of the spatial lag of urban land, forest land, and grassland, while PM 2.5 increases with the increase of the spatial lag of farmland. Specifically, when the spatial lag of urban land expands, the PM 2.5 concentration almost decreases linearly. After the spatial lag of urban land reaches a high value, as the urban land continues to expand, the PM 2.5 concentration appears again rising trend. Unlike construction land, when the spatial lag of forest and grassland increases, the PM 2.5 concentration shows a trend from decline to rise. When the spatial lag of forest and grassland reaches a high value, PM 2.5 shows a downward trend as the forest and grassland expand. It indicates that the expansion of urban land, forest land, and grassland in the surrounding area can suppress regional PM 2.5 pollution. However, as the local urban land scale expands, the construction scale in the surrounding regions also expands. This will increase PM 2.5 pollution. Air pollution in urban agglomerations in China is a typical phenomenon, indicating that urban land scale has an Environmental Kuznets Curve effect on PM 2.5 pollution. Reasonable control of the scale of urban land can achieve good environmental effects. Expansion of the forest land and grassland can not only suppress PM 2.5 pollution in the region, but it also has an inhibitory effect in adjacent areas. Moreover, the larger the scale of forest land and grassland, the more obvious the inhibitory effect. The large-scale and green production of farmland will help curb PM 2.5 pollution. The impact of land use scale on PM 2.5 pollution is a non-linear process, and different land use scales have different effects.

The Impact Mechanism of Land Use Change on PM 2.5 Pollution
As shown in Figure 5, as the main driver for air pollution, different land use types, patterns, and development intensities will obviously result in different distribution patterns of PM 2.5 pollution [64][65][66][67]. In fact, the expansion of urban scale and the increase in urban land area will lead to the imbalance of artificial and natural surface structures. In the process of this change, the land pattern has also changed. The construction of new residential areas, commercial areas, or industrial parks has improved the population density and commuting distance, thus increasing pollution emission sources. Changes in cities' landscape pattern structure have also changed the microclimate environment, which is prone to the heat island effect and increased water evapotranspiration [68]. In addition, the increase in land use and development intensity, building density, and height enable cities to accommodate more people. Still, it also increases the urban energy consumption, and the buildings are built higher and higher in the city. The effect of blocking and friction makes the wind flow through the city significantly weaker, thereby exacerbating PM 2.5 pollution [69]. On the contrary, the reasonable optimization of the land structure, distribution pattern, and development intensity can alleviate PM 2.5 pollution in the process of urban expansion. adjust. The city's real objective world is a three-dimensional space, and the distribution pattern and development intensity also affect the distribution pattern of PM2.5 pollution. Therefore, policy tools should be utilized for new urban development and urban reconstruction in old urban areas. Furthermore, the urban greening rate, the land use structure, development intensity, and distribution pattern should be gradually improved and optimized. The air duct should be reserved, and scattered ecological land shall be restored into an overall urban ecological network structure, which will exert a greater scale effect of ecological land. Thus, these improvements can play a role in reducing urban PM2.5 pollution.

Discussion
PM2.5 pollution is affected by both human-made and natural factors [70]. In the desert areas of northwest China, it is mainly affected by natural wind and dust [69]. However, it is mainly affected by human-made factors in eastern China, especially in urban agglomerations [3,31]. The empirical method of conditional mean regression has been employed The area that can be used for newly-built green space and grassland is limited in cities. It is very difficult to control PM 2.5 pollution by optimizing the land use structure alone. Government administrators will face the problem of where and how much to adjust. The city's real objective world is a three-dimensional space, and the distribution pattern and development intensity also affect the distribution pattern of PM 2.5 pollution. Therefore, policy tools should be utilized for new urban development and urban reconstruction in old urban areas. Furthermore, the urban greening rate, the land use structure, development intensity, and distribution pattern should be gradually improved and optimized. The air duct should be reserved, and scattered ecological land shall be restored into an overall urban ecological network structure, which will exert a greater scale effect of ecological land. Thus, these improvements can play a role in reducing urban PM 2.5 pollution.

Discussion
PM 2.5 pollution is affected by both human-made and natural factors [70]. In the desert areas of northwest China, it is mainly affected by natural wind and dust [69]. However, it is mainly affected by human-made factors in eastern China, especially in urban agglomerations [3,31]. The empirical method of conditional mean regression has been employed in most of the existing studies. It cannot essentially reveal the differences in PM 2.5 levels caused by the non-linearity and heterogeneity of different regions. SP-SDM is thus employed to effectively reveal the heterogeneity and non-linear impact mechanism of land use on PM 2.5 and the interaction of land use and its spatial lag on PM 2.5 . It is confirmed that PM 2.5 pollution has a spatial spillover effect, and that the effects of different land use scales on PM 2.5 are significantly different. It provides a certain reference for the management of PM 2.5 pollution and the regional joint prevention and control of PM 2.5 from the perspective of land use optimization. However, the "earth-atmosphere" system is a very complex system, and PM 2.5 pollution is the result of the interaction, synergy, and coupling of natural conditions, human activities, and land use changes. Due to the difficulty in obtaining high-precision PM 2.5 observational data and its complexity response to land use change, there are still many uncertainties. With the accumulation of high-precision PM 2.5 , land use changes, and individual-based social and economic data, the response process and mechanism of PM 2.5 pollution to land use changes should be focused on different time and space scales in the future.

Conclusions
Based on the remote sensing inversion of PM 2.5 data and land use change data in long time series, the Getis-Ord Gi* statistic and SP-SDM are employed to analyze the spatial distribution pattern of PM 2.5 and its response to land use change. The main conclusions are as follows: showing an annual average growth rate of 0.97%. It is still greater than 35 µg/m 3 in nearly half of the cities in China. (2) The spatial distribution pattern of PM 2.5 presents the characteristics of concentrated regional convergence. PM 2.5 is positively correlated with urban land and farmland, while it is negatively correlated with forest land, grassland, and unused land. Furthermore, the average PM 2.5 concentrations show the highest values for urban land and decrease in the order of farmland > unused land > water body > forest > grassland. (3) The impact of land use change on PM 2.5 is a non-linear process, and there are obvious differences for different land types. Moreover, it will also affect the surrounding areas.
Thus, reasonably controlling the scale of urban land and farmland, optimizing the spatial distribution pattern and development intensity, and expanding the forest land and grassland are conducive to curbing PM 2.5 pollution.