Evaluating urban and nonurban PM_2.5 variability under clean air actions in China during 2010–2022 based on a new high-quality dataset

ABSTRACT

The air quality in China has changed due to the implementation of clean air actions since 2013. Evaluating the spatial pattern of PM_2.5 and the effectiveness of reducing anthropogenic emissions in urban and nonurban areas is crucial. Therefore, the China Long-term Air Pollutant dataset for PM_2.5 (CLAP_PM_2.5) was generated from 2010 to 2022 with a daily 0.1° resolution using the random forest model and integrating multiple data sources, including extensive in-situ PM_2.5 measurements, visibility, satellite retrievals, surface and upper-level meteorological data and other ancillary data. The CLAP_PM_2.5 dataset is more reliable and accurate than other public datasets. Analysis of CLAP_PM_2.5 from 2010 to 2022 reveals the decrease in positive urban-nonurban PM_2.5 differences and higher decreasing rates of PM_2.5 in most city clusters in eastern China. Furthermore, separating meteorological and emission contributions to the PM_2.5 variability by a meteorological normalization approach indicates that meteorological contribution gradually changed from unfavorable to PM_2.5 reduction during 2013–2017 to favorable to decline enhancement during 2018–2022, and in urban regions, meteorological contribution is higher than that in nonurban areas. Overall, the reduction in deweathered PM_2.5 concentrations highlights China's significant achievements in terms of comprehensive clean air actions.

KEYWORDS:

• PM_2.5
• urban and nonurban regions
• random forest model
• weather normalization
• clean air actions

1. Introduction

PM_2.5 pollution exerts notable impacts on the regional air quality, climate changes, human activities and health (Chan and Yao 2008; Che et al. 2007; Guo et al. 2018; Liao, Chang, and Yang 2015; Pope et al. 2002). Thus, the importance of air pollution for sustainable cities and communities is highlighted in Sustainable Development Goals (SDGs) of the United Nations. The PM_2.5 pollution levels in China are significantly higher than those in Europe and the United States (Apte et al. 2015; van Donkelaar et al. 2010), which contributes to the high risk of PM_2.5 exposure among Chinese residents (Lelieveld et al. 2015; Lim et al. 2012). To improve the air quality, China has implemented national clean air actions, i.e. the Air Pollution Prevention and Control Action Plan (2013–2017), the Blue Sky Protection Campaign (2018–2020) and the Opinions on Deepening the Battle Against Pollution Prevention and Control (2021–2035), which have led to significant achievements (Ding et al. 2019; Geng, Zheng, et al. 2021; Tao et al. 2020; Wang, Huang, et al. 2019).

Due to the limitations of long-term in-situ monitoring data in China, the retrieval and prediction of high-quality air pollutant datasets with high temporal and spatial resolutions on a long-term scale are crucial for air quality management, health protection and ecological conservation. Big Earth Data technologies, such as satellite remote sensing and geographic information systems, can regularly provide large-scale data, which facilitates global or national sustainable development research. As indicated in a large number of references, machine learning performs better in the retrieval of pollutants than interpolation techniques, semiempirical model, traditional statistical models, and chemical transport models (Bai et al. 2023; Candiani et al. 2013; Carnevale et al. 2011; Liu, Wang, Zhang, et al. 2023; Ma et al. 2022; Zhang et al. 2021) through the combination of multiple data sources. In particular, the Random Forest (RF) method has been widely used due to its high performance (Chen et al. 2018; Hu et al. 2017; Wei et al. 2019; Zhan et al. 2018). Additionally, the Light Gradient Boosting Machine (LightGBM) has achieved success in PM_2.5 retrieval in China (Zhong, Zhang, et al. 2021) and ozone estimation in East Asia (Kang et al. 2021).

Input datasets (explanatory variables) are crucial to improve the retrieval accuracy for PM_2.5. Because the aerosol optical depth (AOD) is highly correlated with PM_2.5, it is the most common variable for retrieving surface PM_2.5 (Chu et al. 2016; Gupta et al. 2006; Jing, Pan, and Sun 2023; Zeng et al. 2020). ⁣However, visibility, which is also closely related to PM_2.5 (Pui, Chen, and Zuo 2014), is rarely included in retrieval models involving machine learning approaches. Furthermore, almost all retrieval algorithms entail the use of only surface meteorological data as input data; however, previous studies indicated that upper-air meteorological factors greatly impact the variability in PM_2.5 concentrations (Cai et al. 2017; Li et al. 2020; Xiao, Miao, et al. 2020; Yang et al. 2023). In addition, the acquisition of as many ground-based in-situ PM_2.5 observations as possible is conducive to the improvement in the model performance. The Chinese air quality monitoring network (mainly installed in urban regions) was built in 2013, and covered 74 cities in the first year and all prefecture-level cities after 2015. The provided national-scale observations are widely used for PM_2.5 retrieval (Bai et al. 2022; Geng, Xiao, et al. 2021; Wei et al. 2021; Xiao et al. 2022). Moreover, as monitoring capabilities have improved, provincial-scale monitoring sites have been established in recent years, covering a larger geographical area and providing a greater source of data for model training. Notably, validation is necessary to evaluate the accuracy of models, and a comparison of the validation performance and predictive capability of different models has been conducted (Wei et al. 2019; Wei et al. 2020; Zhong, Zhang, et al. 2021). However, there is a lack of consistency in the abovementioned comparisons because they do not utilize the same external in-situ data (never used for model training) to assess different PM_2.5 retrieval datasets. Bai et al. (2022) used external data to validate the performance of three open access datasets for China. Unfortunately, the external data only referred to four sites. Moreover, the accuracy of PM_2.5 prediction prior to 2013 cannot be verified, because the Chinese air quality monitoring network was established in 2013.

Population migration and economic growth associated with urbanization can contribute to an increase in urban PM_2.5 concentrations (Han et al. 2014; Li and Huang 2020; Lu et al. 2019), and the more urbanization there is, the higher the PM_2.5 pollution level (Xing, Mao, and Duan 2022). Rural industrial development can also inevitably cause an increase in PM_2.5 exposure in nonurban areas. Due to clean air actions, many industrial sources were relocated from urban to nonurban areas. However, few monitoring sites are located in nonurban regions because the air quality improvement targets are lower (Bian et al. 2019; Yin et al. 2017). Therefore, the control of coal burning and industrial sources in nonurban areas is not as effective as it could be (Tian et al. 2022). Generally, the abovementioned factors can yield a gradual change in the variation characteristics of the PM_2.5 concentrations in urban and nonurban areas. It is necessary to evaluate urban-nonurban differences, which could assist policy-makers in designing effective emission reduction strategies for urban and nonurban areas, and is important for evaluating the effects of PM_2.5 on the regional air quality, climate, and human health. Some studies on this subject have been performed. For example: Xu et al. (2016) estimated urban and nonurban PM_2.5 concentrations in the North China Plain from 2011 to 2014; Lin et al. (2018) compared the decline rates of urban and rural PM_2.5 levels in eastern China from 2011 to 2015; Xiao, Geng, et al. (2020) assessed the interannual trend of the urban-rural PM_2.5 differences in China from 2000 to 2018; Gao et al. (2020) compared the urban-to-suburban and urban-to-background differences in PM_2.5 in four regions of China from 2015 to 2018; Li et al. (2022) observed the urban-suburban gap in PM_2.5 in Beijing from 2014 to 2019. Summarizing the above studies, it can be found that the PM_2.5 concentrations in urban areas are generally higher than those in nonurban areas, the decrease rate of PM_2.5 in cities is significantly higher than that in rural areas due to the implementation of clean air policies, and the concentration difference between these two areas decreases year by year. However, there are limitations in the aforementioned studies. First, PM_2.5 was characterized by regional pollution; therefore, a regional-scale assessment is more appropriate. Second, gridded data should better reflect urban-nonurban differences in a region than limited in-situ site observations. Third, these studies did not include the most recent data, particularly after 2020, which is crucial for the government to promptly adjust control strategies. Finally, PM_2.5 levels were determined by the combination of anthropogenic emissions and meteorological conditions (An et al. 2019; Zhai et al. 2019); therefore, separating the contribution of meteorology and emission reduction to the PM_2.5 variability in urban and nonurban regions is crucial for evaluating the effectiveness of clean air actions, which can facilitate the formulation of appropriate control measures.

In this study, we constructed the China Long-term Air Pollutant dataset for PM_2.5 (CLAP_PM_2.5) during 2010–2022 with a daily resolution of 0.1° from multisource data and assessed and compared the CLAP_PM_2.5 dataset to other public PM_2.5 datasets, involving as many external datasets as possible. Based on the CLAP_PM_2.5 dataset, the spatiotemporal variation and urban-nonurban differences in PM_2.5 in China and key regions were analyzed, and the contribution of meteorology and emission sources to the PM_2.5 concentrations in different regions were quantified. Our findings could provide important support for air quality management and human health risk studies in China.

2. Data

2.1. Multisource data

Hourly in-situ PM_2.5 ground measurements in China were collected from the Ministry of Ecology and Environment of China (MEE) during 2013–2022 and the provincial environmental monitoring centers during 2019–2021 (Figure 1(a)). The number of MEE sites is 802 and 1296 (average value) in 2013 and 2014–2022, respectively, and the average number of provincial environmental monitoring sites is 3110 during 2019–2021. Daily averaged PM_2.5 concentrations for each site were calculated with at least 20 h of measurements every day. PM_2.5 observations were averaged within the 0.1° grid.

Figure 1. (a) Spatial distributions of PM_2.5 monitoring sites in Chinese mainland from the Ministry of Ecology and Environment of China (MEE, average 1296 sites during 2014–2022) and provincial environmental monitoring centers (average 3110 sites during 2019–2021). (b) Spatial locations of the seven representative study areas (JJJ-JingJinJi, YRD-Yangtze River Delta, PRD-Pearl River Delta, CS-Chongqing and Sichuan, SH-Shandong and Henan, SS-Shaanxi and Shanxi and XJ-Xinjiang), the scope of the UCS (Urumqi-Changji-Shihezi) city clusters and the Taklamakan Desert in XJ Province, and land cover in XJ.

Note: The number of MEE sites is 802 in 2013.

Input data from multiple sources were collected in this study to generate a new high-quality PM_2.5 dataset. The data includes surface visibility, satellite aerosol optical depth (Bai et al. 2022), surface and upper-air meteorology from ERA5 reanalysis data including 2 m temperature (TEM), relative humidity (RH), boundary layer height (BLH), evaporation (E), 10 m u component of wind (U), 10 m v component of wind (V), surface pressure (SP) and total precipitation (TP), as well as geopotential height (GEO850/500), relative humidity (RH850/500), temperature (TEM850/500), u component of wind (U850/500), and v component of wind (V850/500) at 850 and 500 hPa, satellite tropospheric NO₂ columns and ground-level NO₂ (Wei et al. 2022; 2023), as well as geographic (Normalized Difference Vegetation Index (NDVI), land cover, elevation) and social (population) factors. The detailed information of aforementioned input data is summarized in Table 1. The input data are at different spatial scales, thereby to ensure spatial consistency of the data, all data were interpolated to the 0.1° grid. In addition, the time variable DOY (the day of the year) was also introduced into the model as an input variable.

Table 1. Summary of the multisource datasets used in this study to retrieve high-quality PM_2.5 datasets.

Data category	Variable	Spatial resolution	Temporal resolution	Time range	Data source
Particulate Matter	PM2.5	in situ	hourly	2013–2022	MEE (‘https://air.cnemc.cn:18007/’)
					Province-level air quality publishing platforms in China
Visibility	VIS	in situ	hourly	2010–2022	the China Meteorological Administration in the Meteorological Information Combine Analysis and Process System (MICAPS); Chinese Ecosystem Research Network (CERN)
AOD	AOD	1 km	daily	2010–2021	LGHAP_AOD: the daily gap-free AOD (‘https://zenodo.org/record/5652257’)
Surface Meteorology	TEM	0.1°	hourly	2010–2022	ERA5 (https://climate.copernicus.eu/climate-reanalysis)
	RH
	BLH
	E
	U
	V
	SP
	TP
Upper-air Meteorology	GEO850	0.1°	hourly	2010–2022	ERA5 (https://climate.copernicus.eu/climate-reanalysis)
	RH850
	TEM850
	U850
	V850
	GEO500
	RH500
	TEM500
	U500
	V500
Tropospheric NO2 columns	NO2	0.125°	monthly	2010–2012, 2021	OMI (‘https://www.temis.nl/airpollution/no2col/no2regioomimonth_qa.php’)
Ground-level NO2		1 km		2019–2020	ChinaHighNO2 (‘https://zenodo.org/record/6622056’)
		10 km		2013–2018	ChinaHighNO2 (‘https://zenodo.org/record/7574177’)
NDVI	NDVI	1 km	monthly	2010–2022	MOD13A3 (‘https://ladsweb.modaps.eosdis.nasa.gov/search/’)
Land cover	LC	30 m	annual	2010–2021	CLCD (‘https://zenodo.org/record/5816591’ \l ‘.Y-4bEsisGvM’)
Elevation	DEM	1 km	–	–	SRTM (‘https://www.resdc.cn/data.aspx?DATAID=123’)
Population	POP	1 km	annual	2010–2022	LandScan (‘https://landscan.ornl.gov/’)

Note: VIS-visibility (m), TEM-2 m temperature (K), RH-relative humidity (%), BLH-boundary layer height (m), E-evaporation (m of water equivalent), U-10 m u component of wind (m/s), V-10 m v component of wind (m/s), SP-surface pressure (Pa), TP-total precipitation (m), and GEO850/500-geopotential height (m²/s²), RH850/500-relative humidity (%), TEM850/500-temperature (K), U850/500-u component of wind (m/s), V850/500-v component of wind (m/s) at 850 and 500 hPa.

2.2. Open access PM_2.5 datasets

In this study, we collected three open access PM_2.5 datasets including TAP (Tracking Air Pollution in China, http://tapdata.org.cn/) (Geng, Xiao, et al. 2021; Xiao et al. 2022), CHAP (China High Air Pollutants, https://weijing-rs.github.io/product.html) (Wei et al. 2020; 2021) and LGHAP (Long-term Gap-free High-resolution Air Pollutants, https://zenodo.org/record/5652265) (Bai et al. 2022).

2.3. External datasets for PM_2.5 validation

We obtained valuable external PM_2.5 datasets from the Campaign for Atmospheric Aerosol Research network of China (CARE-China) (Xin et al. 2015), Beijing-Tianjin-Hebei Atmospheric Environment Monitoring Network (Xin et al. 2012), and Hubei Ecological Environment Monitoring Center. Information on the time range, regional coverage, and the number of sites of the external datasets are given in Table S1. Here, the ‘external datasets’ means the datasets have not been involved in the PM_2.5 datasets retrievals (including CLAP_PM_2.5, TAP, CHAP and LGHAP).

3. Methods

3.1. Model development

Two machine learning algorithms, i.e. the Random Forest (RF) and Light Gradient Boosting Machine (LightGBM) were used to develop PM_2.5 retrieval models. Both approaches are ensemble algorithms for obtaining a strongly supervised model from multiple weakly supervised models, and weakly supervised models refer to the decision tree. ⁣In general, the RF model involves a bagging algorithm that constructs multiple independent decision trees and enables them to jointly decide on the final prediction (Breiman 1996; 2001), whereas the LightGBM model is a boosting algorithm, which incrementally improves the model performance by iteratively training multiple decision trees (Ke et al. 2017). In feature selection, the RF model randomly considers only a fraction of the features, while the LightGBM model prioritizes the features that are the most valuable for the given task. Therefore, the RF and LightGBM models possess distinct abilities in terms of avoiding overfitting and substantially enhancing the model accuracy, respectively. Overall, both algorithms have their advantages, and the selection of the optimal algorithm requires evaluating the performance of all the models through 10-fold cross-validation. With the use of the ‘ranger’ R package (Wright and Ziegler 2017), the parameters for the RF models are as follows: num.trees (a forest of 600 trees) = 600; mtry (the number of variables that may be split at each node) = 5. The main hyperparameters for the LightGBM models in Python are as follows: learning_rate = 0.1; n_estimators = 600; max_depth = 7; num_leaves = 95; bagging_fraction = 0.8; feature_fraction = 0.80. The optimal parameter settings for the RF and LightGBM models were achieved by obtaining the maximum coefficient of determination and the smaller root mean square error based on the samples. The results of these two metrics can indicate the model performance, as shown in Table 2.

Table 2. Validation performance (10-fold cross-validation) of daily PM_2.5 retrievals (μg/m³).

Retrieval Year	Sample-based				Site-based				Number of samples
	RF		LightGBM		RF		LightGBM
	R^2	RMSE (μg/m^3)	R^2	RMSE (μg/m^3)	R^2	RMSE (μg/m^3)	R^2	RMSE (μg/m^3)
2013	0.88	18.63	0.90	18.17	0.86	19.90	0.83	25.64	92286
2014	0.86	15.31	0.88	15.49	0.83	16.49	0.79	22.22	171567
2015	0.84	14.23	0.85	14.75	0.80	15.50	0.76	20.94	283069
2016	0.86	13.04	0.87	13.33	0.83	14.21	0.80	18.41	284042
2017	0.87	11.89	0.88	12.17	0.85	12.84	0.82	16.77	287659
2018	0.88	9.95	0.89	10.26	0.86	10.76	0.82	14.31	289323
2019	0.84	11.82	0.83	12.56	0.82	12.44	0.78	15.98	802109
2020	0.83	11.66	0.82	12.31	0.81	12.18	0.78	15.52	941742
2021	0.84	10.23	0.84	10.88	0.83	10.56	0.79	13.79	723649
2022	0.90	6.95	0.90	7.27	0.88	7.44	0.84	9.95	311092

Based on multisource input data, PM_2.5 retrieval models from 2013 to 2022 were developed year-by-year by the abovementioned two machine learning algorithms. Then, the final algorithm with the better model performance was selected to construct the CLAP_PM_2.5 dataset. Notably, historical (2010–2012) PM_2.5 concentrations are scarce; therefore, a hindcast model with the optimal algorithm was established for 2013–2015, which was used to predict PM_2.5 concentrations throughout the historical period in China.

3.2. Model validation and comparison

The performance of the RF and LightGBM retrieval models was evaluated using two cross-validation (CV) strategies, i.e. the sample-based and site-based 10-fold CV methods. For the sample-based 10-fold CV scheme, all samples were randomly divided equally into 10 groups, and sequentially nine groups were used as the training data and the remaining group as the testing data. The abovementioned training and testing process were repeated 10 times. This approach was commonly utilized for evaluating PM_2.5 retrieval models (Chen et al. 2019; Li et al. 2017; Yao et al. 2019). Similarly, for the site-based 10-fold CV scheme, the data groups were separated by their locations. The site-based approach is more accurate and reliable because it can capture the model's spatial prediction ability (for estimating unknown site values). However, this approach has been rarely utilized in previous studies. In evaluating the hindcast model for historical predictions and testing the quality of the newly constructed dataset, the inclusion of external datasets played a pivotal role, facilitating the comparison between the CLAP_PM_2.5 dataset and three other open access datasets. The statistic metrics, such as coefficient of determination (R²) and root mean square error (RMSE), were calculated to evaluate the performance of the models at the daily scale.

3.3. Meteorological normalization technique

The machine learning-based meteorological normalization approach (such as random forest model), first proposed by this reference (Grange et al. 2018), was used to eliminate the influence of meteorology on pollutants, which can separate the impact of meteorological conditions and emission source on air pollution. According to previous studies (Grange and Carslaw 2019; Guo et al. 2022; Shi et al. 2021; Vu et al. 2019; Zhang et al. 2023; 2024), this study conducted the random forest-based meteorological normalization through the R package ‘rmweather’ to evaluate the meteorological influence and effectiveness of emission reductions for long-term PM_2.5 concentrations during 2010–2022 in key regions in China.

The daily retrieval gridded PM_2.5 data were used as dependent variables. The predictors included time variables (Unix time, week, weekday, month, DOY) and all meteorological variables in Table 1, including TEM, RH, BLH, E, U, V, SP, TP, GEO850, RH850, TEM850, U850, V850, GEO500, RH500, TEM500, U500, V500. Each of key regions was modeled separately, and 70% of the original data were used for model training, whereas the rest (30%) was used for model testing. For each meteorological normalization, time variables were not normalized, and we only normalized the meteorological data (randomly resample 1000 times for each day). Then, at a particular day, 1000 predicted concentrations were averaged as the deweathered concentrations, which can reflect the impact of the change of anthropogenic emission sources. Notably, the meteorological variables were resampled by randomly selecting from the two weeks before and after the selected date (not limited to the present year), which can maintain the seasonal variability characteristic of pollutants as suggested by previous studies (Guo et al. 2022; Vu et al. 2019). The model performance of the testing dataset by the RF model exhibites that the averaged RMSE and R² of PM_2.5 are 12.57 μg/m³ and 0.91, respectively, which indicates that our models have a good prediction capability for daily concentrations.

3.4. Urban and nonurban area classification

According to this reference (Wang et al. 2022), we classified Chinese mainland (hereafter referred to as China) into urban and nonurban areas based on the population density. The China Statistical Yearbook 2010 reported that China's urban population density was 2,200 people per square kilometer. Consequently, we categorized areas with a population density of ≥ 2,200/km² as urban, while considering areas with lower population density as nonurban.

4. Results and discussion

4.1. High-quality CLAP_PM_2.5 dataset

4.1.1. Evaluation of model performance

The year-by-year performance of the RF and LightGBM retrieval models for estimating daily PM_2.5 concentrations during 2010–2022 is presented in Table 2. For the sample-based CV, PM_2.5 estimations by the RF and LightGBM models showed strong agreement with PM_2.5 measurements, with 10-fold CV R² (RMSE) values ranging from 0.83 to 0.90 (6.95–18.63 μg/m³) and 0.82–0.90 (7.27–18.17 μg/m³), respectively. However, the site-based retrieval performance of the two algorithms showed slight difference. The site-based 10-fold CV R² (0.81–0.88) and RMSE (7.44–19.90 μg/m³) by RF were generally better than those by LightGBM (R²: 0.76–0.84; RMSE: 9.95–25.64 μg/m³), which demonstrated that RF achieved a stronger estimated ability for PM_2.5 estimation at sites which were never used for training. Furthermore, the spatial distribution of 10-fold CV results of the RF retrieval models exhibited that more than 80% of the grids had R² values > 0.7 during 2013–2022 (Figure S1). Thus, the performance of the RF model is generally more stable in the time and space series. Overall, previous studies also indicated that the retrieval model generated by RF was better and more robust compared with other methods (such as orthogonal regression, generalized additive model, support vector machine, decision tree regression, etc.) (Enebish et al. 2021; Li et al. 2018; Lyu et al. 2022), which is presumably attributed to the better generalization ability of the RF. Thereby the RF algorithm was selected for the subsequent construction of the long-term PM_2.5 dataset in this study.

Overall, the performance of the retrieval models obtained by the RF model was comparable with that determined in previous studies for estimating long-term PM_2.5 concentrations. For the sample-based CV scheme, the CV R² values for the TAP-10 km and TAP-1 km daily datasets range from 0.80 to 0.88 (2013∼2020) (Geng, Xiao, et al. 2021) and 0.80–0.84 (2015∼2020) (Xiao et al. 2022), respectively; the CV R² values of the daily PM_2.5 estimates for CHAP-1 km range from 0.86 to 0.90 from 2013 to 2018 (Wei et al. 2021); and the prediction model of LGHAP-1 km shows a sample-based CV R² value of 0.9 from 2000 to 2020 (Bai et al. 2022). However, site-based CV R² (spatial CV) results are rare. TAP-10 km exhibits a spatial 5-fold CV R² value ranging from 0.69 to 0.83 (Geng, Xiao, et al. 2021), and the spatial 10-fold CV R² value reaches 0.81 for estimating PM_2.5 concentrations in northeastern China (Meng et al. 2021).

The assessment through external datasets provided a higher credibility and a better understanding for further validation and comparison of different PM_2.5 retrieval datasets (Table 3). In the hindcast years (2010–2012), the test results showed that the R² (RMSE) values were 0.32, 0.47, 0.47, and 0.49 (40.45, 30.10, 30.23, and 25.67 μg/m³) for the TAP, CHAP, LGHAP, and CLAP_PM_2.5 datasets, respectively. A higher performance was observed in eastern China for the four datasets. The poor performance of the predictions from 2010 to 2012 is consistent with that of previous hindcast models (Geng, Xiao, et al. 2021; Gui et al. 2020; Wei et al. 2020), and it is a notable challenge in accurately constructing historical pollutant datasets. In the retrieval years (2013–2018), external testing indicated a higher performance of the four datasets with increasing R² and declining RMSE values because the observation data in these years were employed for model building. Overall, the CLAP_PM_2.5 dataset exhibited a relatively higher quality, although the other datasets performed better at some sites (Figure S2). Notably, the R² values at CERN sites (i.e. background sites) were much lower than those at the JJJ and Hubei sites (Figure S2). Here, it is necessary to emphasize the advantage of external testing, which not only provides validation of the quality of the constructed dataset but also provides a greater understanding of the spatial differences between the CLAP_PM_2.5 and open access datasets. More importantly, the accuracy of PM_2.5 retrievals during historical periods (prior to 2013) with fewer observations was evaluated, which has not been reported in previous references.

Table 3. Comparison of testing results for the PM_2.5 daily values between the CLAP_PM_2.5 dataset and other public datasets through external datasets.

Year	Area	Sample Size	TAP		CHAP		LGHAP		CLAP_PM2.5
			R^2	RMSE	R^2	RMSE	R^2	RMSE	R^2	RMSE
Hindcast (2010–2012)	Whole	8972	0.32	40.45	0.47	30.10	0.47	30.23	0.49	25.67
Hindcast (2010–2012)	Eastern	5376	0.39	40.09	0.50	30.38	0.50	30.65	0.53	26.56
Retrieval (2013–2018)	Whole	72311	0.55	22.28	0.72	15.06	0.71	17.00	0.74	14.81

Note: The unit for RMSE is μg/m³.

4.1.2. Importance of input variables

For the modeling of RF, the contribution of different input variables was analyzed as shown in Figure 2. VIS, AOD and RH had the greatest influence on PM_2.5 variation. Notably, the importance of VIS was predominant which was in agreement with previous studies (Gui et al. 2020; Liu, Bi, and Ma 2017; Wang et al. 2023; Zhong, Zhang, et al. 2021). The inclusion of every type variable can effectively improve the ability to estimate PM_2.5 concentrations, with coefficient of determination improving from 0.67 to 0.84 (Figure 2(b)). AOD has been widely used as the main factor for PM_2.5 retrieval because of its strong correlation with PM_2.5 (Gupta et al. 2006), and the inclusion of AOD increased R² by 6.2% on average through massive references (Bai et al. 2023), and 6.0% (from 0.67 to 0.71) in this study. In comparison, VIS played a more prominent role, with R² increasing from 0.67 to 0.74 (by 10.4%). Thus, this revealed that visibility provided a great possibility to generate more accurate PM_2.5 estimations. It has been confirmed that the meteorological parameters can improve the accuracy of PM_2.5 estimations (Chelani 2019; Liu, Bi, and Ma 2017; Xu, Huang, and Guo 2021). Here, the surface and upper-air meteorological variables elevated R² to 0.81 and 0.83 (by 9.5% and 2.5%), respectively. The inclusion of the NO₂ variable also contributed marginally to the improvement of model estimation.

Figure 2. (a) Importance of the top ten variables for Random Forest model to estimate the daily PM_2.5 concentrations in 2019 in Chinese mainland. (b) Increases of coefficient of determination (R²) and incremental feature selection.

4.2. Spatiotemporal distribution of PM_2.5 in China

4.2.1. Variation characteristics of PM_2.5 in China

To better illustrate the spatiotemporal variations of PM_2.5 in China, we divided key regions with high pollution level and different time periods based on the clean air actions. Six city clusters in China were selected as representative study regions (Figure 1(b)): Jingjingji (JJJ), Yangtze River Delta (YRD), Pearl River Delta (PRD), Chongqing and Sichuan (CS), Shandong and Henan (SH) and Shaanxi and Shanxi (SS). These regions were characterized by high density populations, massive anthropogenic emission sources and heavy pollution. Three periods, i.e. Stage I (2013–2017), Stage II (2018–2020), and Stage III (2021–2022) were separated according to the implement of clean air actions in China.

Figure 3 presents the spatial distribution of annual average PM_2.5 concentrations during 2010–2022 in China. In general, the spatial pattern of PM_2.5 for each year remained similar, with severe PM_2.5 pollution in the eastern, central and northwestern Chinese regions, especially the highest in the JJJ (53.6 μg/m³) and SH (59.3 μg/m³) regions. The extent and intensity of PM_2.5 pollution were gradually shrinking and weakening from 2013 to 2022, which could be attributed to the clean air actions issued by the Chinese government. However, approximately 20% of the land area still indicated annual average concentrations exceeding the national air quality standards (35 μg/m³) by 2022. This finding served as a clear indication that China, as a developing country, must continue to invest substantial efforts to attain a more favorable atmospheric environment.

Figure 3. Spatial distributions of annual average PM_2.5 concentrations in the Chinese mainland from 2010 to 2022 based on the CLAP_PM_2.5 dataset estimated by Random Forest algorithm.

Figure 4 shows the trends of annual average PM_2.5 estimations during 2010–2022 in China and six representative study regions. In China, the PM_2.5 concentrations exhibited a significant decreasing trend with a peak in 2013. The decreasing rates were 1.8, 2.8, 2.7, 1.8, 2.3, 3.0 and 1.9 μg/m³/year in the China, JJJ, YRD, PRD, CS, SH and SS regions, respectively. The greatest decline occurred at Stage Ⅰ, and the PM_2.5 concentrations exhibited a consistent decline, decreasing from 53.2 μg/m³ in 2013 to 35.6 μg/m³ in 2017, corresponding to a reduction of 33.1%. Among the six representative regions, the CS region had the largest reduction (34.6%), followed by the JJJ region (30.3%). At Stages II and III, the PM_2.5 values continued to decline with a diminishing magnitude of the reduction. Moreover, at Stage III, the PM_2.5 concentrations rebounded in China and the YRD, PRD, CS and SS regions, and the greatest increase reached 5.3% from 2021 to 2022 in the CS region. However, the PM_2.5 concentrations in the JJJ and SH regions exhibited sustained reductions of 10.8% and 2.6%, respectively, due to the notable effort of joint pollution prevention and control. In addition, the high PM_2.5 pollution level in Xinjiang (XJ) Province requires attention and will be thoroughly analyzed in the subsequent chapters.

Figure 4. Annual trends of PM_2.5 concentrations estimated from the RF models from 2010 to 2022 in Chinese mainland, the JJJ (JingJinJi), YRD (Yangtze River Delta), PRD (Pearl River Delta), CS (Chongqing and Sichuan), SH (Shandong and Henan) and SS (Shaanxi and Shanxi) regions. The ** represents the trend is significant at the 99.9% (P value < 0.001) confidence levels. The unit for slope is μg/m³/yr.

4.2.2. Evolution of PM_2.5 in Xinjiang

PM_2.5 pollution in XJ was always heavy (Figure 3) with the annual average concentration of 43.4 μg/m³, and the annual (monthly) PM_2.5 values decreased slowly at a rate of 0.92 (0.08) μg/m³ per year over the 2010–2022 period (Figure 5(a)). Because of the unique topography and land cover (large desert area) in XJ, the UCS (Urumchi-Changji-Shihezi) city clusters (85.7°E–89.1°E, 45.5°N–42.6°N) and Taklamakan Desert (77.6°E–88.4°E, 41.2°N–36.8°N) were selected and analyzed separately. There is a significant difference in emission sources between these two regions. The Taklamakan Desert is primarily a source of dust emissions, while the UCS region is predominantly characterized by anthropogenic emissions (Yao et al. 2023). Additionally, it should be mentioned that the accuracy of the PM_2.5 estimations in XJ can be ensured to the greatest extent based on the high spatial validation R² values (Figure S1).

Figure 5. Trends of monthly averaged PM_2.5 concentrations in (a) XJ, (b) Taklamakan Desert, and the (c) UCS city clusters. The variations of monthly averaged wind speed in Taklamakan Desert and deweathered PM_2.5 concentrations in the UCS region are shown in (b) and (c), respectively.

The annual average PM_2.5 concentration in Taklamakan reached 61.9 μg/m³ and increased insignificantly at a rate of 0.04 μg/m³ per month (0.64 μg/m³/yr) during the 13 years (Figure 5(b)). The highest PM_2.5 concentration in Taklamakan occurred in spring, especially in March (88.3 μg/m³), which can be attributed to the dust caused by strong wind (Chai et al. 2017; Liu et al. 2020). The PM_2.5 pollution was heavier in the southwestern part of Taklamakan (Figure 3) due to cyclonic circulation and aerosol transport from the upstream desert (Yao et al. 2023). The Taklamakan Desert is a major source of dust storms, therefore, the massive dust aerosols generated in Taklamakan Desert can be transported and affect seriously air quality in its surrounding regions and central and eastern China (Manktelow et al. 2010; Sun et al. 2010).

There was a significant seasonal variation of PM_2.5 in the UCS, with higher (71.2 μg/m³) and lower (22.5 μg/m³) values in winter and summer, respectively (Figure 5(c)). The heaviest pollution in winter is attributed to the mixing effect of massive anthropogenic emission sources under stagnant meteorological (Wang, Zhang, et al. 2019). PM_2.5 levels exhibited a slight decline over the 13-year period (0.13 μg/m³ per month and 1.50 μg/m³/yr). Moreover, during the three periods influenced by the clean air actions, the reduction of PM_2.5 was 36.4%, 9.5% and 0.7%, respectively, which was lower than those in the other six representative regions. The trend of the deweathered PM_2.5 concentrations was lower than the PM_2.5 estimation trends, which was attributed to favorable meteorological conditions for PM_2.5 elimination after 2018. Generally, air pollution control was effective in this region; however, meteorology increased the yearly variability of the actual values. The exceptionally adverse meteorological conditions during the winter of 2013, which was the most polluted period, significantly exacerbated the levels of pollution.

4.3. Long-term changes in urban and nonurban PM_2.5 under the clean air actions

4.3.1. Changes in urban-nonurban PM_2.5 differences

Due to the high level of human activity in urban areas (Harrison et al. 2021), the intensity of pollutant emissions is generally higher in urban areas than in nonurban areas (Zheng et al. 2021). During the period from 2010 to 2022, the average PM_2.5 concentration in urban areas exceeded that in nonurban areas by 12.9 μg/m³ on average, whereas the urban-nonurban differences in the different key regions were varied (Figure 6). The largest urban-nonurban difference occurred in the CS (21.0 μg/m³) and SS (14.7 μg/m³) regions, which is related to the faster urbanization in terms of the population scale and the industrial manufacturing industry (Zhu, Wang, and Zhang 2019). The urban-nonurban differences in PM_2.5 were the smallest in the YRD (an absolute difference of 1.6 μg/m³) and PRD (2.2 μg/m³) regions; however, the urban PM_2.5 concentrations were lower than those in nonurban areas in the YRD (Figures 6 and S3, respectively) region, and this situation primarily occurred in the autumn and winter months (Figure S4c), which is consistent with the site-based results (Gao et al. 2020). The urban-nonurban difference in the SH region was moderate (8.2 μg/m³). In addition, as shown in Figure S4, the largest urban-nonurban differences of in PM_2.5 mainly occurred in winter, and the smallest differences were observed in summer. This phenomenon may be due to the reduction in local emission sources, the higher proportion of secondary PM_2.5 in the background areas and the highly favorable weather conditions for pollution diffusion in summer (Gao et al. 2020; Zhang et al. 2022).

Figure 6. Annual trends of the difference between the urban and nonurban PM_2.5 concentrations from 2010 to 2022 in China and the six representative regions. The * and ** represent the absolute difference trends that are significant at the 95% (P value < 0.05) and 99.9% (P value < 0.001) confidence levels, respectively. The unit for slope is μg/m³/yr.

Narrowing differences in urban and nonurban PM_2.5 levels during 2010–2022 in China and especially in the JJJ, SH, CS, PRD and SS regions are shown in Figure 6, which depends in part on the difference between the urban and nonurban PM_2.5 reductions (Figure S3). The regions with the fastest and slowest narrowing differences were the JJJ (1.6 μg/m³/yr) and PRD (0.3 μg/m³/yr) regions, respectively. However, the urban-nonurban PM_2.5 difference exhibited a slightly increasing trend in the YRD region, at a rate of 0.2 μg/m³/yr, and the difference changed from positive to negative after 2013. In terms of the seasonal scale, the most notable narrowing urban-nonurban PM_2.5 differences occurred in winter in five key regions, except the YRD region. Increasing differences dominated in autumn and winter in the YRD region, and the nonurban PM_2.5 concentrations were higher than those in urban areas since 2014 (Figure S4c). During the three periods influenced by clean air actions, the urban-nonurban PM_2.5 difference experienced the largest decrease (18.5%∼67.9%) at Stage Ⅰ, except for the YRD region, and the second-largest decline (5.2%∼14.7%) at Stage II, except for the YRD and SS regions. At Stage Ⅲ, the urban-nonurban PM_2.5 differences were the smallest, with values of ±1 μg/m³, remaining nearly unchanged.

4.3.2. Trends in deweathered PM_2.5 concentrations

Figure 7 depicts the changes in the urban and nonurban deweathered PM_2.5 concentrations in the six representative regions during 2010–2022. Consistent with the trend of reconstructed PM_2.5 concentrations, the deweathered PM_2.5 concentrations exhibited a smaller decline from 2010 to 2022 with the rates of 3.5, 1.9, 1.5, 2.5, 2.1, and 1.7 μg/m³/yr (urban regions) and 2.0, 2.0, 1.3, 1.7, 2.1, and 1.3 μg/m³/yr (nonurban regions) for the JJJ, YRD, PRD, CS, SH, SS regions, respectively. The contribution of the emission declining to PM_2.5 reductions was 82.4% (74.1%), 73.1% (86.6%), 69.9% (78.4%), 66.6% (82.2%), 59.7% (76.2%) and 79.5% (87.6%) in urban (nonurban) regions for the JJJ, YRD, PRD, CS, SH, SS regions, respectively, therefore, emissions played the critical role for the PM_2.5 reduction in all regions, reflecting China's significant achievements in comprehensive air pollution control in recent years. The fastest and slowest declines in the urban (nonurban) deweathered PM_2.5 occurred in the JJJ and PRD (SH and PRD, SS) regions, respectively. Overall, the deweathered PM_2.5 concentrations declined faster in urban areas than those in nonurban areas, which may be attributed to more effectiveness of emission reduction in urban regions. In conjunction with the current development of rural revitalization and the transfer of industries (Liu, Wang, Liu, et al. 2023), our finding suggested that nonurban areas need to adjust control strategy.

Figure 7. Annual trends of the urban and nonurban deweathered PM_2.5 concentrations in the six representative regions, namely, (a) JJJ, (b) YRD, (c) PRD, (d) CS, (e) SH, and (f) SS; (g) annual trends of the difference between the urban and nonurban deweathered PM_2.5 concentrations for 2010–2022 period. The * and ** represent trends that are significant at the 95% (P value < 0.05) and 99.9% (P value < 0.001) confidence levels, respectively. The trends in (g) are absolute difference trends. The unit for slope is μg/m³/yr.

Different from the reconstructed PM_2.5 concentrations, the curves of deweathered PM_2.5 concentrations were smoother and the average annual deweathered PM_2.5 reduction level were lower (Table S2), which is consistent with the previous study (Zhong, Tao, et al. 2021). The deweathered PM_2.5 concentrations did not appear a significant increase in 2013 in all regions and a rebound in 2022 in the YRD, PRD, CS and SS regions, which demonstrated that emission reductions were constantly effective for PM_2.5 decline, while adverse meteorology offset this effectiveness. Similar to the urban-nonurban differences in the reconstructed PM_2.5 concentrations, the deweathered PM_2.5 concentration in urban areas, particularly in the JJJ, CS, SH and SS regions, exceeded that in nonurban areas. However, in the YRD region, the deweathered PM_2.5 concentrations in urban areas were lower than those in nonurban areas during the 2010–2018 period and remained almost the same as the nonurban values after 2019 (Figure 7(g)). Severe PM_2.5 pollution mainly occurred in winter, therefore, strengthening anthropogenic emission control in this season was key to mitigate air pollution (Gao et al. 2023). Compared to other seasons, the deweathered concentration in winter showed the highest decreasing rate in the JJJ and CS regions, with a larger reduction rate in urban areas (Figure S5). But a slightly faster reduction rate in nonurban areas appearred in the YRD region. Notably, a different trend is observed in the SH region, showing a lower decrease in the deweathered PM_2.5 during winter (Figure S5e) due to relatively higher deweathered concentrations between 2018 and 2022. This suggested that the control of anthropogenic emissions during winter should be strengthened.

In general, the clean air actions have made a strong contribution to the reduction of PM_2.5. As shown in Table S2b, the urban and nonurban deweathered PM_2.5 concentrations in the six representative regions from 2013 to 2022 decreased by 26.2%∼41.7% and 27.5%∼38.2%, respectively. The largest decreases in the urban(nonurban) deweathered PM_2.5 at Stage Ⅰ, Stage Ⅱ, Stage Ⅲ, and Stage Ⅰ∼Ⅲ were observed in the CS(CS), JJJ(SS), JJJ(JJJ) and JJJ(CS) regions, respectively. The clean air actions implemented during 2013–2017 (Stage I) were more effective in mitigating air pollution, both in urban and nonurban areas. This can be attributed to the implementation of numerous air pollution mitigation measures by the government during this period (Zhang et al. 2019). The effectiveness of emission reduction at Stage Ⅲ (2021–2022) was weakest, and there was even a slight rebound (0.7%) in the deweathered PM_2.5 concentrations in nonurban areas of the CS region. This suggests that reducing PM_2.5 concentrations could be a challenging task, particularly during periods when PM_2.5 levels are relatively low. The largest reduction in urban-nonurban differences of the deweathered PM_2.5 occurred in the JJJ region, suggesting that more stringent policy enforcement efforts have carried out in this region, especially in urban regions.

4.3.3. Impact of meteorology on PM_2.5 levels

Figure 8 shows the meteorological impact on the urban and nonurban PM_2.5 concentrations in the six representative regions from 2010 to 2022, which was reflected through the difference between the observed and deweathered PM_2.5 concentration. In general, meteorology had an adverse impact on PM_2.5 pollution during 2013–2017 (Stage I), but meteorology changed to a favorable impact on amplifying decline after 2018, especially during 2021–2022. This result suggests that the improvement in air quality during 2018–2022 has benefited from the joint efforts of meteorology and emission control. Overall, the impact of meteorological conditions on urban and nonurban PM_2.5 was similar, and there was a slightly higher positive or negative contribution for PM_2.5 reduction in urban regions among all key regions, which suggested urban areas were more affected by meteorology. The magnitude of meteorological contributions on PM_2.5 pollution in different seasons was different. At Stage I, there was a stronger adverse contribution to PM_2.5 pollution in winter. However, at Stage II and III, better favorable contributions were observed, particularly in summer, which amplified the decline in PM_2.5 levels (Figure S6).

Figure 8. Meteorological impacts on the urban and nonurban annual average PM_2.5 concentrations in the (a) JJJ, (b) YRD, (c) PRD, (d) CS, (e) SH and (f) SS regions from 2010 to 2022 and (g) the three clean air action periods in the aforementioned regions. The meteorological impact can be calculated follows: (observed-deweathered)/ observed × 100%. The positive and negative values indicate meteorological conditions are unfavorable and favorable to PM_2.5 decline, respectively.

It's worth noting that the highest positive meteorological contributions in amplifying PM_2.5 deterioration was in 2013, which contributed to 19.3%, 15.1%, 9.6%, 13.9%, 20.7%, and 18.1% (19.0%, 12.7%, 8.7%, 15.7%, 14.4%, and 16.4%) to severe PM_2.5 concentrations especially in winter in urban (nonurban) areas for the JJJ, YRD, PRD, CS, SH and SS regions, respectively (Figures 8 and S6). This is the main cause for a sudden spike in PM_2.5 concentrations and larger urban-nonurban PM_2.5 difference in this year (Figures 4 and 6). Massive studies have indicated that extreme adverse meteorological conditions led to severe PM_2.5 pollution during the winter of 2013 (Cai et al. 2017; Wang et al. 2014). Furthermore, the differences in decreasing percentage between observed and deweathered PM_2.5 during the three clean air actions were all significant (Table S2), indicating that meteorological conditions were positive contributions to the reduction of PM_2.5 in each stage of the clean air actions, which is similar with previous study (Guo et al. 2022).

5. Conclusions

In this study, we constructed a high-quality and high-resolution daily PM_2.5 datasets named the China Long-term Air Pollutant dataset for PM_2.5 (CLAP_PM_2.5) from 2010 to 2022 over China through multiple data source. Then a clearer understanding of the spatiotemporal pattern of PM_2.5 in China, as well as the effectiveness of the clean air actions and meteorological influence on PM_2.5 variability in urban and nonurban areas within key city clusters were explored based on the CLAP_PM_2.5 dataset. The main findings are as follows:

1. The CLAP_PM_2.5 dataset demonstrates a satisfactory performance and is comparable to the TAP, CHAP, and LGHAP datasets. However, accurately predicting PM_2.5 remains a significant challenge. Moreover, compared to the AOD, visibility plays a more crucial role in the retrieval of PM_2.5.

2. The positive urban-nonurban differences in PM_2.5 in the key high-pollution regions (except the YRD region) exhibited a decreasing trend, with the highest (lowest) decline in the JJJ (PRD) region.

3. Heavy PM_2.5 pollution in China is concentrated in eastern China and Xinjiang, which is caused by anthropogenic and dust sources, respectively. Anthropogenic emission abatement was the main diver in reducing PM_2.5 pollution in city clusters from 2010 to 2022, with a higher effectiveness in the JJJ region, urban areas, and the first stage (2013–2017), which reveals clean air actions are very effective. Meteorology exerted an adverse impact on PM_2.5 pollution from 2013 to 2017 but a favorable impact on enhancing the decline from 2018 to 2022, while urban areas were more strongly affected. Notably, the slower decline or the rebound in PM_2.5 concentrations after 2021 indicates that pollution control in China will remain a challenging task even at relatively low PM_2.5 levels.

Acknowledgments

This work was partially supported by the Special Project on National Science and Technology Basic Resources Investigation of China (No. 2021FY100702), Field Station Basic Research Project of the Chinese Academy of Sciences (Grant KFJ-SW-YW043-3), and the Strategic Priority Research Program of the Chinese Academy of Sciences (XDA19020303). All the authors thank for the above funding.

Data availability statement

The data that support the findings of this study are openly available in the Open Science Framework data repository.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

This work was partially supported by the Special Project on National Science and Technology Basic Resources Investigation of China (No. 2021FY100702), Field Station Basic Research Project of the Chinese Academy of Sciences (grant KFJ-SW-YW043-3), and the Strategic Priority Research Program of the Chinese Academy of Sciences (XDA19020303).