Assessing the Nonlinear Relationship between Land Cover Change and PM₁₀ Concentration Change in China

Abstract

Inhalable particulate matter (PM₁₀) is a major air pollutant that has significant impacts on environmental climate and human health. Land-cover change is also a key factor influencing changes in atmospheric pollution. Changes in land-cover types can lead to changes in the sources and sinks of air pollutants, thus affecting the spatial distribution of PM₁₀, which poses a threat to human health. Therefore, exploring the relationship between PM₁₀ concentration change and land-cover change is of great significance. In this study, we constructed an extreme randomized trees model (ET) based on ground PM₁₀ monitoring data, satellite-based aerosol optical depth (AOD) data, and auxiliary data including meteorological, vegetation, and population data to retrieve ground-level PM₁₀ concentrations across China. The coefficient of determination (R²), the mean absolute error (MAE), and the root mean square error (RMSE) of the model were 0.878, 5.742 μg/m³, and 8.826 μg/m³, respectively. Based on this, we analyzed the spatio-temporal distribution of PM₁₀ concentrations in China from 2015 to 2021. High PM₁₀ values were mainly observed in the desert areas of northwestern China and the Beijing–Tianjin–Hebei urban agglomeration. The majority of China showed a significant decrease in PM₁₀ concentrations. Additionally, we also analyzed the nonlinear response mechanism of the PM₁₀ concentration change to land-cover change. The PM₁₀ concentration is sensitive to forest and barren land change. Therefore, strengthening the protection of forests and desertification control can significantly reduce air pollution. Attention should also be paid to emission management in agricultural activities and urbanization processes.

1. Introduction

In recent years, with the rapid economic growth in China, urbanization and industrialization have been continuously advancing, leading to increasingly severe regional extreme weather [1] and atmospheric environmental pollution issues [2]. The problem of atmospheric pollution is particularly prominent and has received attention from numerous researchers. According to the 2021 China Ecological Environment Report, out of 339 cities nationwide, 121 cities still exceeded the environmental air quality standards, accounting for 35.7% of all cities. The proportion of polluted days in terms of environmental air quality was 12.6%, with inhalable particulate matter (PM₁₀) as the primary pollutant, accounting for 25.2% of the total number of exceeded days. This indicates that PM₁₀ remains a significant contributor to air pollution [3,4]. Meanwhile, the formation of land-cover types, including grassland, forest land, arable land, construction land, wetland and water bodies, wasteland, etc., can reflect the influence of natural and anthropogenic factors on surface cover conditions, directly or indirectly representing the biological and non-biological characteristics of the land [5]. The atmosphere can affect climate and ecological environment by altering the radiation balance of the land–atmosphere system. The transition of land-cover types also impacts the atmospheric environment [6]. This transformation alters the spatial and temporal distribution of near-surface air pollutants by affecting the emission, absorption, and purification processes of pollutants [7], posing a threat to human health and well-being [8]. Therefore, quantitatively exploring the relationship between PM₁₀ concentration changes and land-cover changes is crucial.

Acquiring atmospheric particulate matter concentration data is a fundamental prerequisite for studying air pollution. Currently, atmospheric particulate matter monitoring is mainly conducted through ground-based monitoring in China. After decades of development, China has established a nationwide atmospheric monitoring network consisting of both national and local monitoring stations. However, due to China’s vast territory, monitoring stations are relatively dispersed, with the majority concentrated in major and medium-sized cities. Relying on the existing station distribution and monitoring technology, it is challenging to achieve large-scale and continuous monitoring of China’s atmospheric environmental quality, requiring significant manpower and resources [9,10]. The rapid development of satellite remote-sensing technology has made it possible to conduct rapid, large-scale, and periodic dynamic monitoring of the atmosphere using satellite sensors. This not only facilitates the acquisition of real-time information but also reduces monitoring costs. Currently, a large number of domestic and foreign researchers are conducting extensive study on the retrieval of PM₁₀ concentrations using multisource remote-sensing data. They attempt to establish models between aerosol optical depth (AOD) and ground-level PM₁₀ concentrations. Initially, PM₁₀ concentrations were estimated through linear models such as the multiple linear regression model and geographic weighted regression model [11,12]. However, linear models exhibit significant uncertainty when faced with big data and complex nonlinear problems. In contrast, machine-learning models often perform with a higher computational efficiency and can express more complex relationships when solving pollution concentration estimation [13,14,15]. To further improve the accuracy of the model, meteorological auxiliary data such as the height of the boundary layer, humidity, temperature, and so on, which are related to the particulate matter, are added in the retrieval process, which can effectively improve the accuracy of the estimation results [16]. These studies provide an important foundation for simulating large-area, long-time series of ground-level PM₁₀ concentration data. However, the existence of issues such as the characteristics of the models, the selection of auxiliary parameters, and the number of monitoring stations has introduced a lot of uncertainty into the retrieval process, leading to discrepancies in different research results. Therefore, there is still a need for further exploration in PM₁₀ concentration inversion work to select more appropriate models and auxiliary parameters to achieve high-quality prediction.

Research on land-cover change has achieved considerable progress, including studies on the spatial distribution of land cover and its impacts on global climate change and atmospheric pollution [17,18,19]. There is a close relationship between land cover and air pollutants. Changes in land-cover types can alter the emission or absorption of various gases in the atmosphere [20] and modify the hydrological and thermal conditions [21], thereby affecting the atmospheric environment and leading to air pollution issues. This has significant implications for ecosystems and human health. In studies examining the response of land-cover change to atmospheric pollution, the focus has primarily been on the relationship between land-cover change and pollutants such as ozone, nitrogen dioxide, and fine particulate matter (PM_2.5) [22,23,24]. However, there has been relatively less research on the relationship between PM₁₀ and land cover. Furthermore, when investigating the relationship between atmospheric pollutants and different land-cover types, the emphasis has been on the impact of individual land-cover types on air pollutants [25,26]. In areas with complex spatial distributions, changes in air pollutants are influenced not only by a single land-cover type but also by the combined effects of multiple land-cover types. Research studies commonly employ linear regression methods, including panel regression, multiple linear regression, and geographic weighted regression, to explore the relationship between land-cover change and variations in air pollutant concentrations. These methods are utilized to calculate the correlation between different land-cover types, landscape indices, and air pollutants, facilitating the assessment of their mutual influences [27,28,29]. However, the impact of land cover on air pollutant concentrations is not linear but a complex nonlinear relationship. Thus, solely relying on linear regression methods for quantitative analysis may result in biased calculation results. Thus, when discussing the nonlinear response mechanisms of air pollutants to different land-cover types, nonlinear models can be employed to more accurately capture the correlation between them [30,31]. Although studies of PM₁₀ concentrations and land-cover change are well established in their respective fields, few in-depth studies have been conducted on the relationship between the two. Currently, there is still a lack of quantitative research on the nonlinear relationship between land-cover type changes and PM₁₀ concentration variations at large spatial scales. This could lead to government and other relevant departments being unable to promptly capture the changes in PM₁₀ concentrations caused by land-cover type changes, thus hindering the implementation of macroscopic policy adjustments and environmental governance measures.

This study proposes to construct a machine learning-based model for estimating ground-level PM₁₀ concentrations in China using multisource remote-sensing data, with methods such as machine learning and spatial analysis as theoretical foundations. It aims to estimate long-term time series of ground-level PM₁₀ concentrations and analyze the spatiotemporal variations in PM₁₀ across China from 2015 to 2021. Additionally, we explored land-cover change over a seven-year period and quantitatively explored the nonlinear response of PM₁₀ concentrations to land-cover changes in China using a generalized additive model (GAM). This research is expected to provide scientific basis and guidance for the formulation of sustainable land-use policies and management measures to mitigate environmental pollution problems caused by changes in land-cover types.

2. Materials and Methods

2.1. Study Area

This study focuses on the land region of China, covering a latitude range from 7° N to 53° N and a longitude range from 72° E to 136° E. Figure 1 illustrates the overall spatial distribution of elevation in China. The spatial distribution of elevation in China exhibits diversity and complexity. The western region is dominated by the Qinghai-Tibet Plateau, characterized by significant altitude variations including high mountains, plateaus, and basins. In contrast, the eastern region is mainly composed of plains and hills, such as the North China Plain, the Yangtze River Basin, and the Pearl River Delta. The southwestern region comprises the Yunnan-Guizhou Plateau and mountainous terrain, characterized by complex topography and serving as important forest and water sources. The northeastern region includes the Northeast Plain, as well as mountain ranges like Changbai Mountain, boasting vast plains and abundant forest resources. These diverse terrain features have significant impacts on climate, water resources, and the ecological environment [32].

2.2. Data Source

2.2.1. PM₁₀ Monitoring Stations

To better monitor changes in air quality, China has established national automatic air quality monitoring sites. The ground-level PM₁₀ concentrations monitoring data used in this study are sourced from the China National Environmental Monitoring Center (https://www.cnemc.cn/, accessed on 28 March 2024). As of 2021, there are over 2000 monitoring stations nationwide in operation. Figure 1 illustrates the distribution of ground-based air quality monitoring stations. These stations can record the concentration values of pollutants measured at the monitoring site within 1 h, 8 h, and 24 h. The concentration values for 1 h are directly measured, while those for 8 h and 24 h are the sliding averages calculated for the preceding 8 h or 24 h up to the current time (

Table 1. Air quality monitoring data.

Date	Hour	Type	1001A	1002A	1003A	1004A	…	2846A
1 January 2016	0	AQI	245	119	250	243	…	153
1 January 2016	0	PM2.5	195	90	200	193	…	117
1 January 2016	0	PM2.5_24h	67	50	69	65	…	52
1 January 2016	0	PM10	200	-	209	209	…	-
1 January 2016	0	PM10_24h	148	-	110	139	…	68
1 January 2016	0	SO2	26	22	31	7	…	16
1 January 2016	0	SO2_24h	17	17	15	9	…	9
1 January 2016	0	NO2	95	76	98	101	…	102
1 January 2016	0	NO2_24h	71	49	69	76	…	63
1 January 2016	0	O3	10	2	2	3	…	6
1 January 2016	0	O3_24h	29	72	20	22	…	87
1 January 2016	0	O3_8h	11	6	3	3	…	25
1 January 2016	0	CO	3.4	3.6	3.4	3.8	…	67
1 January 2016	0	CO_24h	1.771	1.5	1.542	1.758	...	1.093
...	...	...	...	...	...	...	...	…
1 January 2016	23	CO	4.5	4.2	4.5	5.6	…	1.212
1 January 2016	23	CO_24h	3.133	3.783	3.2	3.487	…	0.943

). Currently, air quality data for China since 13 May 2014, are available, including PM_2.5, PM₁₀, SO₂, O₃, CO, and AQI.

2.2.2. Satellite Remote-Sensing AOD Dataset

AOD reflects the extent to which suspended fine particles in the atmosphere absorb and scatter light, typically used to describe atmospheric transparency. In this study, the AOD data we obtained are sourced from the high spatial resolution MCD19A2 product provided by the MODIS sensor (https://lpdaac.usgs.gov/products/mcd19a2v061/, accessed on 28 March 2024). This product is derived using the multi-angle implementation of atmospheric correction (MAIAC) algorithm, which considers various influencing factors such as aerosol characteristics, surface reflectance, and observation geometry. It provides an important foundation for the study of air pollution and atmospheric aerosols [33]. Given the large volume of data, we utilized the Google Earth Engine (GEE) platform to mosaic and clip the 1 km resolution AOD data at a wavelength of 550 nanometers within the geographical extent of China from 2015 to 2021. Subsequently, we computed the annual average for each pixel to ensure the spatial integrity of the AOD data.

2.2.3. ERA5 Reanalysis of Meteorological Datasets

Changes in atmospheric conditions directly impact the generation, distribution, and concentration levels of particles in the atmosphere [34]. The meteorological data utilized in this study are sourced from the fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis for global climate and weather (ERA5) [35]. The ERA5 provides reanalyzed meteorological data for over 260 variables from 1979 to present. The temporal resolution of this dataset can reach hourly levels. The atmospheric data have a spatial resolution of 0.25° × 0.25°, while ocean waves data have a spatial resolution of 0.5° × 0.5°. In establishing the PM₁₀–AOD relationship model, we consider the influence of boundary layer height, wind speed, wind direction, humidity, pressure, temperature, and precipitation on PM₁₀ diffusion, transport, and deposition. Ultimately, we selected seven meteorological parameters provided by the ERA5 monthly averaged data on single levels from 2015 to 2021, including the 10 m u-component of wind (WU), 10 m v-component of wind (WV), 2 m temperature (T2M), surface pressure (SP), total precipitation (TP), boundary layer height (BLH), and relative humidity (RH).

2.2.4. Dust Emission Dataset

The dust emission (DE) data used in this study are sourced from the Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2), produced by NASA’s Global Modeling and Assimilation Office (GMAO). MERRA-2 generates global atmospheric reanalysis data using the GEOS-5 climate model and the Goddard Earth Observing System Data Assimilation System (GEOS-DAS), employing a global grid point statistical interpolation (GSI) analysis method [36]. It provides atmospheric analysis data with a spatial resolution of 0.5° × 0.625° for the years 1980–2021. The dust particle emission data include five categories with different aerodynamic diameters: 0.1–1, 1–1.8, 1.8–3, 3–6, and 6–10 μm. According to the definition of PM₁₀, these dust particles are important components of PM₁₀. Therefore, we obtained monthly dust emission data from 2015 to 2021 for five different diameters from the Aerosol Diagnostics dataset and superimposed them to obtain the total dust emission.

2.2.5. Population Dataset

In addition to natural sources, human activities also contribute significantly to PM₁₀ emissions. Thus, the spatial distribution of population is crucial for PM₁₀ research [37]. In this study, we utilized population distribution (POP) data at a 1 km spatial resolution on a yearly scale produced by WorldPop from 2015 to 2021. These data show the number of people within each grid at the pixel scale. WorldPop has received contributions from numerous collaborators, organizations, and governments worldwide. By integrating and leveraging a wealth of data products, WorldPop produces a variety of satellite-based thematic population-related products. WorldPop has developed peer-reviewed spatial statistical methods, harnessing the capabilities of machine learning, to transform and disaggregate population counts from the administrative unit levels to the pixel level, utilizing relationships with spatial covariate layers from satellites and other sources [38].

2.2.6. Vegetation Index Dataset

The growth status of vegetation has complex effects on air pollutant concentrations. MOD13A2 is one of the remote-sensing data products provided by the MODIS satellite sensor (https://lpdaac.usgs.gov/products/mod13a2v061/, accessed on 28 March 2024), primarily comprising parameters related to vegetation, such as the normalized difference vegetation index (NDVI). This product is highly available and widely applicable. With a spatial resolution of 1 km and a temporal resolution of 16 days, it provides detailed vegetation information for monitoring seasonal changes and long-term trends in vegetation worldwide. The NDVI data from MOD13A2 have extensive applications in various fields, providing important data support for understanding global vegetation status and ecological environmental changes. The NDVI obtained in this study is a core indicator from the MOD13A2 product, which measures the greenness and growth status of surface vegetation by calculating the difference in reflectance values between the near-infrared and red spectral bands. MOD13A2 has the same resolution as the AOD data, so we realized the mosaicking and annual synthesis calculation of the NDVI data on the GEE platform, and finally clipped them according to the administrative divisions of China.

2.2.7. Land-Cover Dataset

This study utilizes the global land-cover data product (LC) from MCD12Q1, developed through collaboration between the National Aeronautics and Space Administration (NASA) and the United States Geological Survey (USGS). This dataset provides global land-cover maps at a spatial resolution of 500 m annually from 2001 to 2021 [39]. Since the MCD12Q1 dataset is stored in multiple tiles based on row and column numbers, we conducted image mosaicking and extraction of images within the Chinese territory using the GEE platform. Then, we resampled them to a 1 km spatial resolution on GEE in order to maintain the spatio-temporal consistency of all data. The dataset offers five different land-cover classification schemes, and for this study, the International Geosphere–Biosphere Program (IGBP) classification was chosen. The IGBP classification system demonstrates relative consistency in its application across different regions and countries, enabling researchers from diverse backgrounds to utilize the same standards for land-cover classification and comparison. Moreover, the IGBP classification system is characterized by clear, simple categories, facilitating the ease of operation. These features make the classification of large-scale remote-sensing data relatively straightforward and practical for global-scale land-cover monitoring. It describes land surface conditions in 17 categories, including 11 natural vegetation types, 3 land development and mosaic categories, and 3 non-vegetation land type-defined categories. Detailed information on the data used in this study is summarized in

Table 2. Summary of data obtained for this study.

Variable	Unit	Spatial Resolution	Temporal Resolution	Data Source
PM10	μg/m3	In situ	Hourly	WAQIPT
AOD	-	1 km × 1 km	Daily	MCD19A2
BLH	m	0.25° × 0.25°	Monthly	ERA5
RH	%	0.25° × 0.25°	Monthly
SP	hPa	0.1° × 0.1°	Monthly
TP	mm	0.1° × 0.1°	Monthly
WU	m/s	0.1° × 0.1°	Monthly
WV	m/s	0.1° × 0.1°	Monthly
T2M	K	0.1° × 0.1°	Monthly
DE	kg/m2/s	0.5° × 0.625°	Monthly	MERRA-2
LC	-	500 m × 500 m	Yearly	MCD12Q1
NDVI	-	1 km × 1 km	16-Day	MOD13A2
POP	-	1 km × 1 km	Yearly	WorldPop

.

2.3. Methods

2.3.1. Extreme Randomized Trees

The extreme randomized trees model (ET) is a tree-based ensemble built according to a top-down procedure that can be well used in classification and regression problems. Its distinction from other tree ensemble models primarily lies in the randomized approach to selecting attributes and cut points during node partitioning. Additionally, it adjusts the degree of randomization to suit the particular problem at hand, thus offering adaptability tailored to specific contexts [40]. Furthermore, the tree ensemble model does not need to consider the covariance between multiple characteristic variables when performing regression [41], the information in multiple feature variables can be better mined to participate in the training of the model. In this study, we constructed an ET model using Python 3.9. To maintain consistency in the spatio-temporal resolution of the input data, we performed annual synthesis on the raw data and resampled them to 1 km using bilinear interpolation. We trained the retrieval model using a fused dataset comprising satellite remote-sensing data on AOD, meteorological parameters, NDVI, and population information. Ultimately, we established the relationship between the ground-level PM₁₀ stations and the AOD data.

2.3.2. Generalized Additive Model

GAM is a flexible statistical model used to explain complex nonlinear relationships between response variables and multiple predictor variables [42]. Unlike linear models, GAM allows for the modeling of each predictor variable’s relationship in a nonlinear manner while maintaining the overall interpretability of the model. In this study, to explore the impact of the land-use type area on PM₁₀ concentration, GAM is introduced to reveal the nonlinear response mechanism of PM₁₀ to land-cover type area. The general form of GAM can be expressed as

$$g\left(E\left(Y\right)\right)=\beta +{\sum }_{i=1}^n{f}_i\left(X_i\right)$$

(1)

where the $Y$ is the response variable for the probe, $E\left(Y\right)$ is the expected value of $Y$, and $g\left(\right)$ is a link function that can be used to relate $E\left(Y\right)$ to the linear combination through functions such as the identity function, logarithm; $\beta$ represents the intercept term, and $f_i\left(X_i\right)$ are smooth functions that respectively represent the nonlinear relationships associated with the predictor variable $X_i$. The core idea of GAM is to use smooth functions to flexibly fit the relationship between each predictor variable and the response variable. Smooth functions can be kernel functions, cubic natural splines, and cubic spline smoothing functions to adapt to different types of data structures. In this study, we implement GAM using the mgcv package in R 3.6.2.

2.3.3. Evaluation Indicators

To evaluate the prediction accuracy of the retrieval model, this study uses the 10-fold cross-validation method. The dataset was first randomly divided into 10 equal datasets, 9 of which were used as training datasets to parameterize the model, and the rest were used as test datasets to evaluate the model. This validation process was repeated 10 times, and the results of the 10 validations were averaged to represent the overall model performance [43]. Meanwhile, the PM₁₀ concentrations predicted by the model were validated using the PM₁₀ concentrations observed at the actual ground stations. In this study, we the choose the coefficient of determination (R²), the mean absolute error (MAE), and the root mean square error (RMSE) to evaluate the model. The detailed logical structure of this study is illustrated in the flowchart (Figure 2).

$$R^2=1-\frac{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2}{{\sum }_{i=1}^n{(y_i-\overline{y})}^2}$$

(2)

$$MAE=\frac{1}{n}({\sum }_{i=1}^n\left|{\widehat{y}}_i-y_i\right|)$$

(3)

$$RMSE=\sqrt{\frac{1}{n}{\sum }_{i=1}^n{({\widehat{y}}_i-y_i)}^2}$$

(4)

3. Results

3.1. Spatial and Temporal Variations in Ground-Level PM₁₀ Concentrations

We constructed an ET model for the retrieval of annual ground-level PM₁₀ concentrations in China from 2015 to 2021. Ultimately, the 10-fold cross-validation results showed that R², MAE, and RMSE were 0.878, 5.742 μg/m³, and 8.826 μg/m³, respectively, indicating that the model has a good fitting accuracy. Additionally, the small difference between R² values of the training set and the testing set suggests that the model did not experience overfitting during the training process. Figure 3 displays the estimation accuracy at each ground monitoring station from 2015 to 2021. It can be observed that 47.01% of the stations have an R² exceeding 0.8, with the majority of these stations concentrated in the eastern region where there are more monitoring stations. The proportion of stations with an R² exceeding 0.6 reaches 74.79%. In regions with a dense station distribution, the model is able to capture features more comprehensively, resulting in a higher prediction accuracy. Conversely, in western regions where stations are sparse, the limited features available in the dataset during model training may lead to relatively larger errors. The proportion of stations with an MAE and RMSE within 5 μg/m³ is 52.11% and 43.29%, respectively, and over 90% of the stations have an MAE within 10 μg/m³. The proportion of stations with an MAE and RMSE exceeding 25 μg/m³ is only 0.98% and 1.24%, respectively. Overall, the spatial distribution of stations accuracy is similar to that of R², with very few stations exhibiting large errors in regions such as Xinjiang, Tibet, and Qinghai. In previous studies, random forest (RF), geographically weighted regression (GWR), land-use regression (LUR), and space-time extremely randomized trees (STET) models have been applied to retrieve ground-level PM₁₀ concentrations, with R² values of 0.74, 0.81, 0.64, and 0.86, respectively [44,45,46,47]. The ET model demonstrates a better accuracy compared to previous research. This indicates that the constructed inversion model can effectively perform the retrieval of ground-level PM₁₀ concentrations in China.

Figure 4 illustrates the spatiotemporal distribution of annual average ground-level PM₁₀ concentrations in China from 2015 to 2021, with a spatial resolution of 1 km × 1 km, estimated based on the ET model. PM₁₀ exhibits significant spatial clustering features in Figure 4, with the highest values occurring in the desert areas of Xinjiang, western Inner Mongolia, and the northwest of Qinghai and Gansu. Not only do these regions have annual average PM₁₀ concentrations exceeding 100 μg/m³, but the spatial extent of high-concentration areas is also substantial. Additionally, relatively high PM₁₀ concentrations are observed in the Beijing–Tianjin–Hebei urban agglomeration and northern Henan Province, where the annual average PM₁₀ concentrations range from 100 μg/m³ to 125 μg/m³. In other regions with high PM₁₀ concentrations, they are mainly concentrated in economically relatively developed cities and their surrounding areas. Overall, the PM₁₀ pollution in the north is more severe than in the south, with annual average PM₁₀ concentrations generally below 75 μg/m³ in the south. Northern Sichuan, Yunnan, Guizhou, and coastal cities in the south have lower annual average PM₁₀ concentrations, all below 50 μg/m³. From the perspective of land-cover types, areas with more forest tend to have lower PM₁₀ concentrations, while areas covered by deserts tend to have higher PM₁₀ concentrations.

From 2015 to 2021, there was an overall downward trend in PM₁₀ concentration in China. We conducted linear fitting for each pixel of PM₁₀ over these seven years to clarify spatial change trends and performed significance tests on the trends (Figure 5) to assess the statistical significance of the sample data, ensuring that the observed trend changes are significant. It can be seen that in most areas of central and eastern China, the p-values are less than 0.05, indicating significance, while in western and northeastern China, the significance is less pronounced. Figure 6 shows the results of the trend fitting, and to highlight areas with significant changes in China, we selected regions where the trend change was significant at a 95% confidence level (p < 0.05) for plotting. In most areas passing the significance test, PM₁₀ concentration showed a significant downward trend, especially in northern Xinjiang, western and eastern Inner Mongolia, and major urban agglomerations in the Beijing–Tianjin–Hebei region and Shandong Province, where the decrease exceeded 3 μg/m³/year. Areas with significant increases in PM₁₀ concentration were mainly in desert areas surrounding Xinjiang, as well as central Qinghai, the Qilian Mountains in Gansu, western Sichuan, and the Yunnan–Guizhou Plateau. However, the PM₁₀ concentrations in these areas still remain relatively low.

3.2. Land-Cover Change from 2015 to 2021

Figure 7 shows the overall spatial distribution of land-cover types in China in 2015 and 2021. Due to various factors such as geographical environment, climate conditions, and human activities, land cover exhibits diversity and complexity, with different types of land cover intermingling. Forests are predominantly concentrated in the eastern and southwestern regions, particularly in the northeast, North China, the Yangtze River Basin, southwestern, and northwestern parts. The forests in the northeast are mainly composed of coniferous trees, while those in the southern regions consist mainly of broad-leaved trees. Grasslands are mainly distributed in the northern and northwestern regions of China, such as Inner Mongolia, Xinjiang. Cropland is widely distributed in the eastern plain areas and along the valleys of rivers and lakes. Urban areas are scattered and mainly concentrated in the eastern coastal areas and around some important economic centers. Barren lands are primarily located in the northwestern region, such as the Taklamakan Desert and the Kumtage Desert. In recent decades, with the rapid economic development in China, significant changes have occurred in the spatial pattern of land cover, leading to a series of environmental issues [48].

Studying the transition of land-cover types is conducive to better analyzing and understanding the spatial structure and quantity changes in land surfaces, providing critical information for land management, environmental protection, climate change, and sustainable development. The land-cover transition matrix effectively extracts the area conversion between different types of land cover. In this study, we computed the land-cover type transition matrices for 2015 and 2021 (

Table 3. Land-cover transition matrix from 2015 to 2021 in China (km²).

2015	2021
	Forest	Grassland	Wetland and Water	Cropland	Urban	Barren Land	Total
Forest	994,618.79	73,068.24	252.96	786.78	4.65	0.93	1,068,732.35
Grassland	170,610.36	4,075,128.25	6224.49	107,611.23	4886.22	28,286.88	4,392,747.43
Wetland and Water	416.64	3816.72	152,886.82	415.71	69.75	789.57	158,395.21
Cropland	3579.57	93,090.21	833.28	1,432,184.53	3450.3	33.48	1,533,171.37
Urban	26.97	729.12	21.39	440.82	145,883.93	1.86	147,104.09
Barren land	196.23	57,954.81	2463.57	232.51	46.52	2,233,789.52	2,294,683.16
Total	1,169,448.56	4,303,787.35	162,682.51	1,541,671.58	154,341.37	2,262,902.24	9,594,833.61

). From 2015 to 2021, significant conversions occurred among different land-cover types across China. Specifically, forests mainly received conversions from grasslands, with an area conversion reaching more than 170,000 km², possibly associated with afforestation projects and related activities. Grasslands had the largest areas converted both into and out of, as a substantial amount of grassland was converted into forests, while significant areas were also converted into croplands, closely related to China’s demand for food. Additionally, during the rainy season or snowmelt period, grasslands may be submerged by river or lake flooding, transforming them into wetlands and water. Conversely, through drainage and marshland drying, wetlands may revert back to grasslands [49]. The conversion of urban areas over the seven years under study was the smallest among all land-cover type conversions, as the proportion of urban areas globally is extremely small. It is noteworthy that there were also considerable conversions between grasslands and barren lands, but overall the area converted from barren land to grassland was larger.

The results of the land-cover type transition matrix provide a quantitative description of land-cover changes from 2015 to 2021. To better understand the spatial dynamics of different land-cover types, spatial change maps of forests, grasslands, croplands, wetland and water, as well as urban areas across China from 2015 to 2021, were also plotted (Figure 8). Both grasslands and forests witnessed large-scale conversions across various provinces in China. China has been focusing on afforestation and forest conservation efforts in regions surrounding the Beijing–Tianjin–Hebei urban agglomeration and the desert areas in the northwest, leading to the conversion of barren land into grasslands and forests. Areas converted into croplands are primarily located in the North China Plain and the Northeast China Plain, which are traditional agricultural areas critical for China’s food security. The Northeast China Plain was historically one of the main agricultural regions, but recent economic restructuring and urbanization have led to spatial shifts in cropland distribution. Land improvement and the rational use of cropland are issues of great concern for the Chinese government, resulting in significant changes in cropland spatial distribution over the seven years under study [50]. Wetlands and water primarily originate from grassland conversions, appearing in river and lake basins. Urban expansion is driven by economic development, with urban areas continuously expanding outward from major cities [51]. In many regions of northwest China and Inner Mongolia, extensive grassland degradation has occurred due to arid climates, fragile ecological environments, and unsustainable land management practices.

3.3. Nonlinear Relationship between Land-Cover Change and PM₁₀ Concentrations Change

In order to investigate the relationship between land-cover types and PM₁₀ concentrations, we conducted statistical analysis on the average PM₁₀ concentrations for different land-cover types, as shown in

Table 4. Average PM₁₀ concentrations for different land-cover types (unit: μg/m³).

	2015	2016	2017	2018	2019	2020	2021
Forest	37.3	37.25	36.24	38.19	38.34	36.86	35.96
Grassland	55.81	56.02	54.17	54.66	54.47	52.91	53.58
Wetland and Water	59.2	60.65	57.86	59.45	60.7	59.62	59.41
Cropland	61.07	63.71	60.32	62.28	62.53	61.52	60.79
Urban	61.76	61.93	61.62	61.54	61.21	60.25	61.19
Barren land	100.33	98.97	100.01	101.18	98.42	97.04	98.34

. From 2015 to 2021, the interannual variation range of PM₁₀ concentrations at various land-cover types was within 5 μg/m³. It can be observed from the table that the atmospheric quality is the best for forest cover locations, with values within 40 μg/m³ throughout the seven years, significantly lower than other land-cover types. Grassland areas exhibit the second lowest PM₁₀ concentration among the other land-cover types. Wetland and water, croplands, and urban areas have similar PM₁₀ concentrations, with wetland and water being slightly lower than croplands and urban areas. The PM₁₀ concentration in urban areas remains relatively stable, except for 2020, where it stayed between 61 and 60 μg/m³, approximately 1 μg/m³ lower compared to other years. This may be closely related to the sudden outbreak of the coronavirus disease 2019 (COVID-19) pandemic at the end of 2019. The lockdown policy implemented by China resulted in a reduction in traffic and industrial activities. Traffic restrictions reduced vehicle exhaust emissions, while industrial shutdowns reduced industrial emissions, leading to a decrease in PM₁₀ concentration in some urban areas. With the arrival of the post-pandemic era, economic and social activities gradually resumed, and industrial and transportation activities gradually returned to normal levels, leading to a resurgence of air pollutant emissions in some areas [52,53]. Barren land, with its exposed and loose surface, is highly susceptible to wind erosion, which easily generates large amounts of particulate matter into the atmosphere, resulting in air pollution. Moreover, desertification is predominant in barren lands, and sandstorms are one of the main sources of PM₁₀, making barren lands exhibit the highest PM₁₀ concentration among all land-cover types.

The differences in PM₁₀ concentration among different land-cover types are significant. To further explore the relationship between land-cover types and PM₁₀ concentration, we plotted the proportions of different graded PM₁₀ concentrations under different land-cover types. Figure 9 illustrates significant differences in air quality among different land-cover types. The PM₁₀ concentrations in forested areas are all within 75 μg/m³, with over 90% of forested areas having PM₁₀ concentrations below 50 μg/m³. More than half of grassland areas have PM₁₀ concentrations below 50 μg/m³. However, wetland and water, croplands, and urban areas have PM₁₀ concentrations exceeding 50 μg/m³, with over half of these areas surpassing this threshold. Currently, China’s primary standard for PM₁₀ concentration is 40 μg/m³, indicating that there is still a PM₁₀ exposure risk in many areas with frequent human activities, such as croplands and urban areas. Additionally, barren lands are consistently affected by sandstorms, with 47.8% of the regions experiencing PM₁₀ concentrations exceeding 100 μg/m³.

Table 5. The average PM₁₀ concentration change caused by land-cover change (unit: μg/m³).

2015	2021	ΔPM10	2015	2021	ΔPM10
Forest	Forest	−7.21	Cropland	Forest	−9.40
	Glassland	−5.33		Glassland	−7.96
	Wetland and Water	−2.51		Wetland and Water	−10.65
	Cropland	−0.69		Cropland	−2.42
	Urban	−5.11		Urban	−9.67
	Barren land	0.95		Barren land	−1.05
Grassland	Forest	−6.82	Urban	Forest	−5.94
	Grassland	−1.04		Grassland	−9.67
	Wetland and Water	−7.79		Wetland and Water	−14.23
	Cropland	−1.12		Cropland	−3.49
	Urban	−11.48		Urban	−5.76
	Barren land	−0.42		Barren land	−8.09
Wetland and Water	Forest	−5.86	Barren land	Forest	−17.34
	Grassland	−9.22		Grassland	1.12
	Wetland and Water	−8.04		Wetland and Water	−5.11
	Cropland	−2.85		Cropland	−2.34
	Urban	−9.82		Urban	−6.72
	Barren land	0.79		Barren land	3.39

shows the concentration of variation in PM₁₀ when the land-cover transition occurred. Where there was no change in land-cover type, PM₁₀ concentrations showed a decreasing trend, except for barren land. It can be seen that among all types of transitions, only three transitions show an increase trend in PM₁₀ concentration. Two of these transitions result in an increase in PM₁₀ concentration due to conversion to barren land, while the other land-cover type transitions show a decreasing trend. There is a significant decrease in PM₁₀ concentration when various land-cover types are converted to forest, especially when barren land is converted to forest, with a decrease of −17.34 μg/m³; compared to the PM₁₀ concentration baseline of barren land, the PM₁₀ concentration decreased by 20.73 μg/m³. Additionally, when forest is converted to other land-cover types, the change in PM₁₀ is minimal. But compared to the baseline of forest, the decrease in the PM₁₀ concentration is much smaller, indicating that the land-cover type switch decreased the suppression of PM₁₀. When various land-cover types are converted to urban and wetland and water bodies, there is also a significant decrease trend in PM₁₀ concentration, ranging from −5.11 μg/m³ to −11.48 μg/m³ and −2.51 μg/m³ to −14.23 μg/m³, respectively. Conversely, there is relatively little change in the PM₁₀ concentration when converted to cropland.

It is difficult to quantitatively assess the complex nonlinear response mechanism of PM₁₀ concentration to land-cover types by simple statistical analysis. In this study, GAM models were constructed to calculate the nonlinear response characteristics of the PM₁₀ concentration to the proportion of land-cover types in 2015 and 2021. The overall interpretability of the GAM models was 75.6% in 2015 and 76.7% in 2021. Among all land-cover types, only wetland and water did not pass the significance test, while the other five types of land cover passed the significance test at p < 0.01 (

Table 6. GAM fitting results of PM₁₀ concentrations and land-cover types.

2015		edf	Ref.df	F	p-Value
	Forest	5.551	6.41	57.11	<2 × 10−16 **
	Grassland	7.915	8.257	3.7	0.000486 **
	Wetland and Water	3.969	4.451	1.53	0.172
	Cropland	2.263	3.493	31.08	<2 × 10−16 **
	Urban	5.244	7.088	13.73	<2 × 10−16 **
	Barren land	1.518	1.752	41.72	<2 × 10−16 **
2020		edf	Ref.df	F	p-value
	Forest	5.797	7.114	55.13	<2 × 10−16 **
	Grassland	6.222	7.003	3.899	0.000408 **
	Wetland and Water	3.804	4.746	2.057	0.0641
	Cropland	3.424	4.692	22.22	<2 × 10−16 **
	Urban	4.574	6.568	7.844	<2 × 10−16 **
	Barren land	1.546	1.842	115.6	<2 × 10−16 **

Note: ** p-value < 0.01.

). This might be due to the unique spatial distribution of wetland and water, often found within other land-cover types, resulting in relatively small sample sizes and areas. The results in the table indicate that forests were the dominant factor in the variation in PM₁₀ concentration in 2015, while in 2021, the dominant factor shifted to barren land, which may be closely related to the occurrence of a severe dust storm in China in March 2021 [54]. In nonlinear fitting, the F-statistic is an indicator that assesses the overall significance of the GAM model and can help to determine whether the model is appropriate and whether the explanatory variables have a significant effect on the response variable. This also resulted in a significant difference in the F-statistic for barren land between these two years. Overall, forests and barren land remain the primary factors driving the changes in PM₁₀ concentration.

The trend of PM₁₀ concentration changes with variations in the proportion of different land-cover types remains unclear. We plotted the fitted smoothed curves for 2015 and 2021 using the R language with the average PM₁₀ concentration and the proportion of each land-cover type area as the dependent variable and independent variable, respectively (Figure 10 and Figure 11). The proportion of the land-cover type area at all sample points is indicated at the x-axis. It can be observed that in both years, the smooth curve trends between various land-cover types and PM₁₀ were generally similar. This may be due to the limited time span covered by the dataset and the fact that land-cover types in many areas show a relatively stable state during this time, which leads to the similarity of the nonlinear patterns. Forests exhibited a significant inhibitory effect on PM₁₀ concentration changes. Before the proportion of forest area exceeds 60%, the PM₁₀ concentration decreases rapidly as the proportion increases. When it exceeds 60%, it still maintains a certain inhibitory effect, but the effect on the change in the PM₁₀ concentration is not as significant. The inhibitory effect of grasslands is relatively moderate. Changes in wetland and water are also relatively gentle. An increase in cropland land leads to an increase in PM₁₀ concentration. When the proportion of cropland reaches around 40%, the rate of increase in the PM₁₀ concentration is the fastest. As the proportion of cropland continues to increase, the effect on PM₁₀ concentrations levels off. The relationship between urban and PM₁₀ concentration shows an initial increase followed by a decrease as the urban area proportion rises, with the PM₁₀ concentration reaching its peak when the urban area proportion reaches about 10%. The curve also reveals that during urban expansion, the rate of increase in the PM₁₀ concentration exceeds the rate of decrease. The pattern is similar to the environmental Kuznets curve. The environmental Kuznets curve describes the relationship between economic growth and environmental pollution. The curve has an inverted U-shape, representing the rise in environmental pollution with GDP growth in the early stages of economic development and the improvement in environmental quality in the later stages as people become more aware of environmental protection and government environmental policies are implemented. In both years, the degrees of freedom of the fitted curve for barren land were less than two, indicating that the relationship between barren land and PM₁₀ concentration is very close to linear. In desert areas, which are the most important natural sources of PM₁₀ within barren land types, there is a significant trend of increasing PM₁₀ concentration.

4. Discussion

This study focuses on China, analyzing the spatiotemporal variations in PM₁₀, land cover, and the nonlinear relationship between land-cover changes and PM₁₀ concentration changes from 2015 to 2021. There is a strong link between the spatial distribution of PM₁₀ and atmospheric conditions [55]. High temperatures increase convective activity in the atmosphere, which may make PM₁₀ more susceptible to being suspended in the air. High temperatures may also lead to chemical reactions between gases and particulate matter that can generate new PM₁₀ or accelerate the release of PM₁₀. Wind speed and wind direction can determine how much particulate matter is diluted in the atmosphere and the direction in which it is dispersed. Water vapor in the atmosphere can wash particulate matter from the atmosphere and settle it to the ground, thus reducing PM₁₀ concentrations. Overall, in regions such as Xinjiang, western Inner Mongolia, and the northwest of Qinghai and Gansu, the underlying surface is covered by barren land, much of which comprises deserts. Deserts are often characterized by high temperatures and dryness, which leads to increased convective activity in the atmosphere and strong winds, which provide important conditions for the occurrence of dust storms and enrich the sources of PM₁₀ in the atmosphere [56]. Consequently, these regions are subject to severe PM₁₀ pollution throughout the year. In addition, the causes of the high PM₁₀ present in the Beijing–Tianjin–Hebei urban agglomeration and northern Henan Province area are not exactly the same as in the barren land area. It is noteworthy that the main land-cover type in this area is urban. Due to the combination of the barometric pressure system, topography, and seasonal winds, strong northwesterly winds are formed, which bring dust and sand from far away to the Beijing–Tianjin–Hebei region. In addition to this, human activities such as industrial emissions, construction activities, and road dust are also important reasons for the increase in PM₁₀ in the air [57]. However, over a seven-year period, China has seen very significant reductions in PM₁₀ concentrations within many heavily polluted areas. This indicates that the series of environmental protection measures implemented by the Chinese government over the years have played an important role. Examples include the implementation of a series of air pollution prevention and control action plans, adjustment of industrial structure, promotion of clean energy development, and establishment of ecological environment protection policies such as the Three-North Shelter Forest Program [58]. In many of these implemented policies, changes in land-cover types are utilized to alter atmospheric conditions. Measures such as desertification control and the creation of artificial forest belts transform land from barren to grassland or woodland. Additionally, adjustments in industrial structure lead to changes in the spatial pattern and scale of urban areas, thereby effectively controlling pollution source emissions. In the southern region of China, the climate is relatively humid and precipitation is high, which cleanses the atmosphere of particulate matter and allows it to settle to the ground [59]. It is also often influenced by the southeastern monsoon, which carries pollutants from inland areas over the ocean, making the southern region relatively less polluted. In addition to climatic conditions, its lower degree of industrialization, lower population density, and large area of vegetation cover are also important factors in the overall better air quality in the south than in the north.

The results of the nonlinear relationship between land-cover change and PM₁₀ concentration change provide insights for the future development of rational air pollution control policies and land-related policies. Trees, as important dust sinks, have the function of filtering and trapping dust, obstructing the dispersion of particulate matter in the atmosphere through their multi-layered structure, thus producing a certain purification effect [60]. Additionally, vegetated surfaces effectively reduce ground dust and mitigate the risk of resuspension of dust [61,62]. The protection of forested areas and the promotion of sheltered forest projects can have a direct and significant reduction effect on PM₁₀ concentrations. Although grasslands are limited by their canopy structure and cannot achieve the same level of effectiveness as forests, they still possess some air purification capabilities. Furthermore, in many regions of China, grasslands have suffered severe degradation due to climate change, population growth, economic development, overgrazing, and unregulated development. As a result, they have lost their ability to intercept particulate matter and may even become sources of pollution [63]. Therefore, there is also a need to prioritize the restoration of grasslands to prevent further degradation. As one of the world’s major agricultural countries, China has a large rural population base. The occurrence of high concentrations at cropland locations is closely related to the intensity of agricultural activities. During the sowing and harvesting seasons of agricultural production, the soil structure is disturbed, and large amounts of dust and sticky substances are exposed. Driven by the wind, these substances are easily suspended in the air, becoming potential sources of PM₁₀ [64,65]. In addition, agricultural activities involve a series of other factors such as mechanical operations, straw burning, and the use of pesticides and fertilizers, all of which may contribute to PM₁₀ or its precursors as significant dust sources, indicating the potential impact of agricultural activities on air quality [66]. This explains the response of PM₁₀ concentration to changes in cropland area share in the results of the nonlinear relationship well. Usually, moisture in the air tends to absorb and settle particulate matter in the air. The moist environment helps to slow down the suspension time of particulate matter in the air, reducing its concentration. The air humidity is usually higher in the vicinity of wetlands and water bodies, which also creates better conditions for precipitation and increases the likelihood of precipitation. Therefore, areas of high relative humidity and precipitation can have a significant impact on PM₁₀ concentrations [67,68]. However, the nonlinear relationship curve between wetlands and water bodies did not exhibit a significant negative correlation. This may be due to the unique shape and geographical location of inland water bodies; they are susceptible to the influence of nearby particulate matter emissions, especially in areas adjacent to high particulate matter emission sources such as industrial zones and areas with dense traffic pollution. The air quality around water bodies may be significantly affected by the surrounding environment, resulting in high levels of pollution [69], thus making their impact on PM₁₀ concentration largely dependent on the surrounding environment. In developed countries, there is often a negative correlation between urban development and air pollution, due to the fact that in developed countries, there is more capital investment in environmental management and more sophisticated urban planning projects [70]. But China is still in a state of unbalanced and inadequate development, with many regions having relatively low levels of economic development and light environmental pollution. However, as urbanization accelerates, environmental pollution increases from low to high. When the proportion of urban area reaches around 10%, PM₁₀ concentration also reaches its peak. As the level of economic development reaches a certain level, environmental pollution decreases from high to low, and PM₁₀ concentration begins to decrease, ultimately leading to a gradual flattening of the curve as the environmental quality improves [71]. China has invested a great deal of human, material, and financial resources in combating sand and dust storms over the years, and it has indeed achieved remarkable results in terms of the trend of PM₁₀ concentration and land-cover transfer, and it should adhere to the implementation of the relevant policies in the future so as to bring the PM₁₀ concentration up to the qualified standard. In conclusion, the average level of PM₁₀ has a significant positive correlation with the percentage of barren land, cropland, and urban area, but a negative correlation with the proportion of wetland and water, forest, and grassland area. This is consistent with the findings of previous researchers in urban agglomerations in China [17,72].

5. Conclusions

In this study, we established the relationship between ground-based PM₁₀ monitoring stations in China and AOD data obtained from satellite remote sensing. By utilizing the ET model, we achieved the retrieval of PM₁₀ concentration across China, ultimately generating spatio-temporal distribution maps of the annual average PM₁₀ concentration from 2015 to 2021 at a spatial resolution of 1 km. The results indicate that the ET model demonstrates a better estimation accuracy compared to many classical machine-learning models, with final R², MAE, and RMSE values of 0.878, 5.742 μg/m³, and 8.826 μg/m³, respectively. High values primarily appear in the desert areas of northwest China and the Beijing–Tianjin–Hebei urban agglomeration. Overall, northern regions experience more severe PM₁₀ pollution compared to southern regions. Between 2015 and 2021, significant decreases in PM₁₀ concentration were observed across most regions of China, particularly in northern Xinjiang, western Inner Mongolia, and major urban agglomerations in the eastern region.

There exists a complex nonlinear relationship between land cover and air pollutants. Currently, the response mechanism of PM₁₀ concentration to land-cover types is not well understood. Based on the generated PM₁₀ concentration data and combined with land-cover data, we used the GAM model to investigate the response of PM₁₀ concentration to land-cover types. The PM₁₀ concentration is mainly influenced by forests and barren land, with the PM₁₀ concentration rapidly decreasing with an increasing forest area and increasing at a slower rate with an increasing barren land area. Therefore, efforts directed towards forest conservation and afforestation, as well as desertification control, are crucial for reducing air pollution significantly. Additionally, attention should be given to emission management in agricultural activities and the urbanization process.

In this study, there are also some limitations. For example, in areas with sparse station data in western China, the model estimation results still show some underestimation. In future research, we intend to incorporate more geographical auxiliary factors to improve the accuracy of the model. Additionally, the process influencing the PM₁₀ concentration is highly complex, and its diffusion, migration, and transformation are not solely affected by changes in land-cover types. Factors such as meteorological conditions and indicators of human activity intensity, such as road network density and the proportion of the three major industries in the region, also play a significant role in influencing PM₁₀ concentration changes, which requires further in-depth investigation.