Inversion of Aerosol Chemical Composition in the Beijing–Tianjin–Hebei Region Using a Machine Learning Algorithm

Abstract

Aerosols and their chemical composition exert an influence on the atmospheric environment, global climate, and human health. However, obtaining the chemical composition of aerosols with high spatial and temporal resolution remains a challenging issue. In this study, using the NR-PM1 collected in the Beijing area from 2012 to 2013, we found that the annual average concentration was 41.32 μg·m⁻³, with the largest percentage of organics accounting for 49.3% of NR-PM1, followed by nitrates, sulfates, and ammonium. We then established models of aerosol chemical composition based on a machine learning algorithm. By comparing the inversion accuracies of single models—namely MLR (Multivariable Linear Regression) model, SVR (Support Vector Regression) model, RF (Random Forest) model, KNN (K-Nearest Neighbor) model, and LightGBM (Light Gradient Boosting Machine)—with that of the combined model (CM) after selecting the optimal model, we found that although the accuracy of the KNN model was the highest among the other single models, the accuracy of the CM model was higher. By employing the CM model to the spatially and temporally matched AOD (aerosol optical depth) data and meteorological data of the Beijing–Tianjin–Hebei region, the spatial distribution of the annual average concentrations of the four components was obtained. The areas with higher concentrations are mainly situated in the southwest of Beijing, and the annual average concentrations of the four components in Beijing’s southwest are 28 μg·m⁻³, 7 μg·m⁻³, 8 μg·m⁻³, and 15 μg·m⁻³ for organics, sulfates, ammonium, and nitrates, respectively. This study not only provides new methodological ideas for obtaining aerosol chemical composition concentrations based on satellite remote sensing data but also provides a data foundation and theoretical support for the formulation of atmospheric pollution prevention and control policies.

1. Introduction

In recent years, air pollution has been a serious environmental problem in China [1,2,3,4,5]. Aerosols play an important role, not only seriously affecting the quality of the atmospheric environment [6,7,8,9,10], but also having a significant impact on climate change [11,12,13,14,15] and human health [16,17,18,19]. At present, the common methods used to obtain the chemical composition of aerosols include sampling analysis [20,21,22,23], electron microscope scanning [24,25], and model simulation [26,27]. Although these methods can accurately obtain information on the various chemical components of aerosols, they have problems, such as difficulty in real-time monitoring and poor spatial coverage, and it remains difficult to obtain aerosol concentration data with long time series and large area coverage. Through global or regional coverage, satellite remote sensing can make up for the spatial discontinuity of ground monitoring stations. However, there are still some remote sensing mechanisms that present difficulties in quantifying parameter inversion problems. In order to solve this problem, some scholars have tried to combine machine learning with satellite remote sensing data, which has powerful nonlinear fitting ability [28], to achieve the inversion of ground PM2.5 concentration [29,30], and thus make up for the spatial fault and information gap in ground-based monitoring data. Huttunen et al. compared and analyzed the advantages and disadvantages of the lookup table method, nonlinear regression method, and four machine learning methods, and found that the aerosol optical thickness (AOD) obtained by most machine learning methods has higher accuracy [31]. Lanzaco et al. used AOD data from AERONET ground-based observations and meteorological factors (temperature, humidity, wind speed, and wind direction) to train a neural network for correcting MODIS AOD, through which the monthly mean error of satellite AOD was reduced to 0.1–0.6 [28]. Sun et al. built a PM2.5 concentration prediction model based on a deep neural network and used the 1 km resolution AOD products and meteorological data (temperature, humidity, wind speed, direction, visibility, etc.) from the stationary satellite Himawari-8 and the model to estimate hourly surface PM2.5 concentration in the Beijing–Tianjin–Hebei region [30]. Using MODIS AOD data and meteorological data, Chen et al. constructed a random forest model for retrieving PM2.5 concentration and combined the model to estimate the PM2.5 concentration in China from 2005 to 2016. The R² and RMSE of the verified model were 0.83 and 28.1 µg·m⁻³ [32]. The technology of inverse representation of aerosol optical thickness and PM2.5 concentration based on machine learning and satellite remote sensing data is relatively mature. However, relatively few studies have been conducted on the inversion of aerosol chemical components based on machine learning.

The Beijing–Tianjin–Hebei (BTH) region is one of the three economic circles of China’s “capital economic circle”. It is not only the political, economic, and cultural center but also has important international influence. However, the rapid economic development and continuous expansion of the city have led to air pollution in the Beijing–Tianjin–Hebei region. Although the quality of its air environment has improved in recent years, it is still not optimistic. According to the national environmental quality status released by the Ministry of Ecology and Environment, in the past two years, cities such as Xingtai, Shijiazhuang, Handan, Baoding, and Tangshan in the Beijing–Tianjin–Hebei region are still in the bottom 20 of China’s key cities in terms of air quality. Therefore, retrieving the spatial distribution of aerosol chemical components in the Beijing–Tianjin-Hebei region through machine learning can not only clearly identify the source of air pollutants but also provide data support and theoretical guidance for government departments to formulate reasonable and effective air pollution prevention and control measures.

2. Data and Method

2.1. Observation Sites and Data Sources

The observational data used in this study were collected from Beijing station (39.97° N, 116.36° E), Xianghe Station (39.96° N, 116.95° E), and Xinglong station (40.40° N, 117.58° E), representing typical urban, suburb, and background environments over the Beijing–Tianjin–Hebei region (as shown in Figure 1). Among them, Beijing station is situated in the Institute of Atmospheric Physics, Chinese Academy of Sciences, approximately 15 m above the ground, with no significant point sources in the vicinity. The Xianghe observation station is located at the Xianghe Atmospheric Sounding Comprehensive Experiment Station of the Chinese Academy of Sciences, also 15 m above the ground. The station is surrounded by dense light industry and residential living areas, and there are no obvious local emissions or tall buildings nearby. Xinglong station is positioned in the Xinglong atmospheric background observation station, located south of Yanshan Mountain and north of the Great Wall, which is mainly encompassed by high mountains and less affected by human activities [33].

The aerosol chemical concentration data at the Beijing site were collected through aerosol mass spectrometers fabricated by the Aerodyne company. By employing an aerosol mass spectrometer to monitor the non-refractive (NR-PM1) portion of submicron aerosol PM1 (particle aerodynamic diameter ≤1 μm), the mass concentration of organics, nitrates, ammonium, sulfates, and nitrates can be derived [34]. The sampling period was 2012 to 2013 and the time resolution was 1 h. The aerosol chemical composition concentration data at Xinglong and Xianghe stations were acquired from particulate matter samples collected by the Wuhan Tianhong TH-150C medium flow sampler. Among them, the concentrations of organic carbon were measured using the thermal optical carbon analyzer developed by the Desert Research Institute of the United States, and water-soluble inorganic ions such as NH₄⁺, NO₃⁻ and SO₄²⁻ in PM2.5 were measured using ICS-90 ion chromatography. The time resolution of sampling is one week [33]. The proportions of various chemical components in PM1 and PM2.5 exhibit similarities over the BTH region [2,35]. Hence, data from the Beijing site of NR-PM1 was utilized to establish the models, and data from Xianghe and Xinglong of PM2.5 were used to examine the performance of the models.

Aerosol optical depth (AOD) is the integration of the aerosol extinction coefficient from the earth’s surface to the top of the atmosphere. MODIS AOD data is obtained from the LAADS website “https://ladsweb.modaps.eosdis.nasa.gov/ (accessed on 1 October 2024)” and the product of Terra C6 DB 10 km AOD data is selected. The temporal resolution is 1 day, and the spatial resolution is 0.1° × 0.1° [36]. The meteorological data used in this study are ERA-5 reanalysis data, with a spatial resolution of 0.25° × 0.25° and a temporal resolution of 1 h [37]. The meteorological variables used are Air Temperature at 2 m (2 Meter Temperature, T2M), Relative Humidity (RH), 10 Meter U Wind component (U10), 10 Meter V Wind component, (V10), Surface Pressure (SP) and Boundary Layer Height (BLH).

The different components of aerosols are classified according to their extinction properties into scattering and absorbing components. Light-scattering aerosols typically include sulfate, nitrate, ammonium salts, and most organic aerosols, while light-absorbing aerosols generally include black carbon, dust, and brown carbon [38]. The contributions of sulfate, ammonium, nitrate, and organic matter account for 70% of the total extinction [39]. The aerosol pollution in the Beijing–Tianjin–Hebei region primarily results from the rapid growth of secondary aerosols [40,41]. Therefore, our research focused on organics, ammonium, sulfates, and nitrates.

2.2. Analysis Method

2.2.1. Data Preprocessing

Owing to the disparity of temporal and spatial resolution of observation data, namely MODIS AOD data and ERA-5 reanalysis data, it is necessary to align the data in time and space before establishing the model. Firstly, the meteorological variable grid data of T2M, RH, U10, V10, SP, and BLH corresponding to the Beijing station (39.97° N, 116.36° E) were extracted from ERA-5 data, and the meteorological data corresponding to the Beijing station was acquired by averaging within a rectangle of 50 km × 50 km. Secondly, in the process of establishing the model, taking the observation time of AMS as the benchmark, the AOD and meteorological data were matched.

2.2.2. Machine Learning Method

This study selected six of the most prevalently used algorithms, including: Multivariable Linear Regression (MLR), which features a short modeling duration, and excellent interpretability; Support Vector Machine (SVM), which is appropriate for regression prediction with a smaller number of samples and holds an advantage in handling sample data that is linearly inseparable [42]; Random Forest (RF), an ensemble learning algorithm based on Bagging and decision trees [43,44,45]; K-Nearest Neighbor (KNN), a commonly employed algorithm for classification and regression [46]; XGBoost (eXtreme Gradient Boosting), a representative algorithm based on boosting [47]; and LightGBM (Light Gradient Boosting Machine), an algorithm evolved from GBDT (Gradient Boosted Decision Tree) [48]. The construction of the combined model is founded on the prediction outcomes of all the above-mentioned single models, and the prediction results of all the single models are combined through the grey wolf weight optimization algorithm, ultimately obtaining the final prediction results of the prediction samples [49].

In this study, the criterion coefficient (R-Square, R²), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE) were selected as the evaluation indices of the model performance, and the calculation formula of each evaluation index was as follows:

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

(1)

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

(2)

$y_i$ is the observed value, ${\overline{y}}_i$ is the average value, and ${\widehat{y}}_i$ is the predicted value of the model or the inversion result.

3. Results and Discussion

3.1. Characteristics of Aerosol Chemical Composition Concentration in Beijing

As shown in Figure 2, the NR-PM1 in different seasons exhibits distinct characteristics. The average concentrations of NR-PM1 in spring, autumn, and winter were 41.98 μg·m⁻³, 44.93 μg·m⁻³, and 50.36 μg·m⁻³, respectively, which are significantly higher than the summer average of 22.32 μg·m⁻³. In addition, the average concentrations of organics, sulfates, ammonium, and nitrates in summer are also lower than those in other seasons. By comparing the concentrations of different chemical components, it can be revealed that the concentration and percentage of organics are the highest in different seasons. The average concentration of NR-PM1, organics, sulfates, ammonium, and nitrates in summer are all lower than those in other seasons. However, the percentage of sulfates in the summer average concentration is 23.68%, which is much higher than the percentage in other seasons. This is related to the fact that the temperature is higher in summer and the solar radiation is stronger, which is conducive to the transformation of sulfur dioxide into sulfates. The average concentration of nitrates in autumn is 10.14 μg·m⁻³, the highest among the four seasons. This is because the formation of nitrates is mainly based on photochemical reactions, but it is prone to decomposition at high temperatures. The radiation in autumn is stronger, and the temperature is moderate, which is conducive to the formation of nitrates and their existence as a compound in the atmosphere [23], so the nitrates concentration in autumn is higher. The concentration of organics and sulfates in winter are also higher than those in other seasons, and the percentage is the highest except for summer. This is mainly affected by the release of large amounts of sulfate precursors from fossil fuel combustion during winter heating in Beijing, as well as poor pollution dispersion conditions and lower boundary layer height, which will increase the concentration of pollutants [23].

3.2. Correlation Between AOD, Meteorological Factors and Aerosol Chemical Component

Statistically, the correlation coefficients between the concentrations of organics, sulfates, ammonium, nitrates, and AOD, as well as meteorological factors during the study period, were determined (

Table 1. Correlation coefficients (CC) and p values between aerosol chemical component and aerosol optical depth (AOD) as well as meteorological factors.

		ORG	SO42−	NH4+	NO3–
AOD	CC p	0.70 0.0002	0.79 0.0000	0.82 0.0000	0.76 0.0000
T2M	CC p	0.13 0.02	0.21 0.003	0.24 0.001	0.24 0.001
RH	CC p	0.41 0.02	0.57 0.0008	0.50 0.0006	0.45 0.0001
U10	CC p	−0.27 0.09	−0.20 0.9	−0.25 0.08	−0.26 0.08
V10	CC p	0.31 0.05	0.28 0.03	0.33 0.0002	0.32 0.0003
SP	CC p	−0.14 0.004	−0.13 0.04	−0.13 0.001	−0.13 0.003
BLH	CC p	−0.38 0.008	−0.29 0.005	−0.35 0.004	−0.36 0.04

). The concentrations of organics, sulfates, ammonium, nitrates, and AOD in the Beijing area demonstrated a linear correlation. The correlation between meteorological factors and the four chemical components of aerosol was high, especially T2M, RH, U10, V10, SP, and BLH. U10, SP, and BLH were negatively correlated with the concentrations of organics, sulfates, ammonium, and nitrates, while T2M, RH, and V10 were positively correlated with the concentrations of organics, sulfates, and ammonium. When there is a strong correlation between the input variables in the machine learning algorithm, the performance and stability of the constructed model will deteriorate [47]. We further analyzed the correlation between meteorological factors and AOD in Beijing and discovered that, except for the correlation between BLH and U10, which was above 0.5 and strong, the absolute values of the correlation coefficients among other features were all below 0.5, indicating that the linear correlation between other variables was weak. Considering the stability of the model, U10 was selected for exclusion when the model was established because, compared with the correlation coefficient between BLH and aerosol chemical composition concentration, U10 is a relatively less important feature. In summary, AOD, T2M, RH, V10, SP, and BLH can be used as input features for constructing models.

3.3. Establishment of Single Model and Comparison of Results for Different Chemical Components

During the construction of each model, the initial values of the parameters are set in accordance with the algorithmic characteristics and parameter-tuning experience of different models. GridSearch provided by Scikit-learn is adopted for hyperparameter optimization [50]. The optimal parameters of each model are obtained based on the evaluation indicators, and the respective machine learning models are constructed using the optimal parameters and the training set. The scope of optimization and the optimal model parameter combination determined by repeated experiments for each model are shown in

Table 2. Scope of optimization and parameter selection for different models.

Parameter	ORG	SO42−	NH4+	NO3–	Scope of Optimization
SVR
kernel	rbf	rbf	rbf	rbf	[‘linear’, ‘poly’, ‘rbf’, ‘sigmoid’]
gamma	6.31	0.52	2.10	2.09	[1 × 10−4, 1]
C	0.73	1.31	0.72	0.74	[1 × 10−2, 1 × 102]
RF
n_estimators	80	50	70	70	[50, 1000]
max_depth	10	11	10	12	[5, 50]
min_samples_split	9	3	2	3	[2, 20]
min_samples_leaf	1	2	2	1	[1, 10]
max_features	5	5	4	4	[0.1, 1.0]
KNN
n_neighbors	3	5	3	3	[1, 30]
weights	distance	distance	distance	distance	[‘uniform’, ‘distance’]
algorithm	auto	auto	auto	auto	[‘auto’, ‘ball_tree’, ‘kd_tree’, ‘brute’]
XGBoost
objective	reg:squarederror	reg:squarederror	reg:squarederror	reg:squarederror	[‘reg:squarederror’, ‘reg:linear’, ‘reg:pseudohubererror’, ‘reg:logistic’]
learning_rate	0.01	0.01	0.01	0.01	[0.01, 0.3]
n_estimators	350	340	355	390	[50, 1000]
max_depth	5	5	5	6	[3, 10]
min_child_weight	1	1	5	5	[1, 10]
subsample	0.6	0.9	0.7	0.7	[0.5, 1.0]
colsample_bytree	0.9	0.9	0.7	0.9	[0.5, 1.0]
gamma	0	0	0	0	[0, 5]
LightGBM
boosting_type	gdbt	gdbt	gdbt	gdbt	[‘gbdt’, ‘dart’, ‘goss’, ‘rf’]
objective	regression	regression	regression	regression	[‘regression’, ‘huber’, ‘fair’, ‘poisson’, ‘quantile’, ‘mape’, ‘gamma’, ‘tweedie’]
learning_rate	0.01	0.01	0.01	0.01	[0.01, 0.3]
n_estimators	490	160	435	480	[50, 1000]
max_depth	8	8	8	8	[−1, 15]
num_leaves	19	18	20	28	[20, 200]
max_bin	125	55	175	65	[25, 500]
feature_fraction	0.7	0.6	0.6	0.6	[0.5, 1.0]
bagging_fraction	0.7	0.8	0.9	0.8	[0.5, 1.0]
bagging_freq	40	50	10	40	[0, 100]
lambda_l1	0.1	0	0.1	0.001	[0, 1]
lambda_l2	0.001	0	1	0.5	[0, 1]

.

Figure 3 and

Table 3. Linear regression functions between simulated results of 6 single models and observations for four chemical components in Beijing.

	KNN	LGBM	MLR	RF	SVR	XGBoost
oganics	y = 0.76x + 4.82	y = 0.71x + 5.45	y = 0.49x + 8.75	y = 0.68x + 6.08	y = 0.62x + 6.38	y = 0.75x + 4.80
	(R2 = 0.75, p = 0.00)	(R2 = 0.75, p = 0.00)	(R2 = 0.53, p = 0.00)	(R2 = 0.74, p = 0.00)	(R2 = 0.68, p = 0.00)	(R2 = 0.75, p = 0.00)
surfates	y = 0.70x + 1.81	y = 0.59x + 3.36	y = 0.63x + 2.39	y = 0.67x + 2.15	y = 0.67x + 2.15	y = 0.73x + 1.93
	(R2 = 0.70, p = 0.00)	(R2 = 0.61, p = 0.00)	(R2 = 0.56, p = 0.00)	(R2 = 0.60, p = 0.00)	(R2 = 0.50, p = 0.00)	(R2 = 0.68, p = 0.00)
ammonium	y = 0.86x + 0.61	y = 0.71x + 1.30	y = 0.68x + 1.42	y = 0.74x + 1.19	y = 0.79x + 1.38	y = 0.75x + 1.25
	(R2 = 0.84, p = 0.00)	(R2 = 0.83, p = 0.00)	(R2 = 0.66, p = 0.00)	(R2 = 0.81, p = 0.00)	(R2 = 0.75, p = 0.00)	(R2 = 0.83, p = 0.00)
nitrates	y = 0.82x + 1.54	y = 0.74x + 2.07	y = 0.59x + 3.00	y = 0.75x + 2.20	y = 0.74x + 2.86	y = 0.81x + 1.77
	(R2 = 0.85, p = 0.00)	(R2 = 0.77, p = 0.00)	(R2 = 0.70, p = 0.00)	(R2 = 0.79, p = 0.00)	(R2 = 0.74, p = 0.00)	(R2 = 0.77, p = 0.00)

show that the performance of the model of organics and sulfates is worse than that of ammonium and nitrates. In addition, in terms of the inversion accuracy of six models of each chemical component of aerosol, the KNN model has the highest inversion accuracy with higher R², followed by the XGBoost model, RF model, and LightGBM model. The MLR model and SVR model have poor inversion accuracy among the six models. The red dots in Figure 3 represent the outliers, and the data are for 27 November 2012. From 16 November 2012 to 1 December 2012, Beijing experienced a prolonged period of air pollution. The maximum PM2.5 concentration reached 300 μg·m⁻³ in 1 December. The explosive increase of sulfates and organics was the primary reason for this pollution event. The simulation results indicate that the model has relatively low accuracy in simulating the concentration of organics and sulfates when dealing with severe pollution incidents.

3.4. Performance of Single Models in Xianghe and Xinglong

The satellite remote sensing AOD data and ERA-5 data corresponding to the latitude and longitude of Xianghe and Xinglong were extracted. Six machine learning models were used to estimate the concentrations of organics, sulfates, ammonium, and nitrates at Xianghe and Xinglong, and the MAE of the models was compared and analyzed by the observed data at Xianghe and Xinglong in

Table 4. MAE and R² between the simulated results of 6 single models and observations for four chemical components in Xianghe and Xinglong.

		MLR	SVR	RF	KNN	XGBoost	LightGBM
				Xianghe
organics	MAE R2	18.26 0.4	19.05 0.52	17.53 0.59	18.87 0.56	19.72 0.51	17.35 0.61
sulfates	MAE R2	10.93 0.39	10.46 0.47	10.40 0.46	10.15 0.50	10.41 0.48	9.43 0.55
ammonium	MAE R2	7.43 049	6.15 0.70	5.30 0.70	5.65 0.60	6.36 0.54	5.18 0.72
nitrates	MAE R2	14.16 0.55	15.08 0.54	12.57 0.61	12.63 0.61	15.13 0.58	14.07 0.73
				Xinglong
organics	MAE R2	25.05 0.34	24.73 0.37	24.23 0.41	24.50 0.38	24.98 0.35	24.19 0.52
sulfates	MAE R2	13.83 0.29	13.24 0.32	12.97 0.37	12.68 0.33	13.49 0.31	11.93 0.42
ammonium	MAE R2	7.21 0.35	6.86 0.40	6.60 0.42	6.55 0.51	6.89 0.40	5.39 0.62
nitrates	MAE R2	17.41 0.39	17.82 0.37	16.55 0.46	17.05 0.45	17.76 0.35	17.65 0.37

. For Xianghe Station, the MAE between the estimated concentrations of organics, sulfates, ammonium, and nitrates by the LightGBM model and the measured values is the smallest, with MAE values of 17.53 μg·m⁻³, 9.43 μg·m⁻³, 5.18 μg·m⁻³, and 14.07 μg·m⁻³, respectively. Meanwhile, R² between the estimated concentrations by the LightGBM model and the measured values is the highest, with values of 0.61, 0.55, 0.72, and 0.73 for organics, sulfates, ammonium, and nitrates. The MAE between the estimated concentrations of organics, sulfates, ammonium, and nitrates by the MLR model and the measured values is larger. The MAE values are 18.26 μg·m⁻³, 10.93 μg·m⁻³, 7.43 μg·m⁻³, and 14.16 μg·m⁻³, respectively. Overall, the MAE between the estimated concentrations of organics, sulfates, ammonium, and nitrates by the MLR model, SVR model, and XGBoost model and the measured values is larger, while the MAE between the estimated concentrations by the RF model, KNN model, and LightGBM model and the measured values is smaller. For Xinglong Station, single models show similar performance with Xianghe station, the MAE between the estimated concentrations of organics, sulfates, ammonium, and nitrates by the LightGBM model and the measured values is the smallest, with the values being 24.19 μg·m⁻³, 11.93 μg·m⁻³, 5.08 μg·m⁻³, and 17.65 μg·m⁻³, respectively. The R² between the estimated concentrations by the LightGBM model and the measured values is the highest in addition to nitrates. The MAE between the MLR model-estimated concentrations of organic matter, sulfate, ammonium, and nitrate and the measured values is relatively large. The MAE values are 25.05 μg·m⁻³, 13.83 μg·m⁻³, 7.21 μg·m⁻³, and 17.41 μg·m⁻³, respectively.

3.5. Establishment of the Combined Model and Examination of the Performance

The combined model has two basic requirements of high precision and a large difference from the internal single model. According to the above research conclusions, the RF, KNN, XGBoost, and LightGBM models of each component meet the above two basic requirements. Therefore, this paper builds a combined model based on RF, KNN, XGBoost, and LightGBM, and names the combined model CM (Combined Model). In the process of constructing the combined model, the grey wolf weight optimization algorithm was used to optimize the weight of the single model inside the CM model, and then the CM model was constructed [49]. Figure 4 presents the flowchart for establishing the combined model.

The concentrations of organics, sulfates, ammonium, and nitrates at the Beijing, Xianghe, and Xinglong stations were estimated by the CM model, and the error of the model was compared with the measured values at the Beijing, Xianghe, and Xinglong stations.

Table 5. MAE and R² between the simulated results of the combined model and observations for four chemical components in Beijing, Xianghe, and Xinglong.

Stations		Organics	Sulfates	Ammonium	Nitrates
Beijing	MAE R2	13.08 0.84	7.87 0.69	4.65 0.86	9.93 0.85
Xianghe	MAE R2	17.23 0.63	10.17 0.54	5.47 0.72	13.13 0.73
Xinglong	MAE R2	24.05 0.52	12.58 0.40	5.71 0.62	16.64 0.40

shows the MAE and R² between the estimated concentrations of organics, sulfates, ammonium, and nitrates and the measured values at the Beijing, Xianghe, and Xinglong stations based on the CM model. By comparing the mean absolute error (MAE) and R² of Beijing, Xianghe, and Xinglong, it is found that the magnitude of the error is in the order: Beijing < Xianghe < Xinglong, which is consistent with the result of the single model.

The MAE between the concentrations of organics, sulfates, ammonium, and nitrates estimated by the CM model and the measured values at the Beijing site was 13.08 μg·m⁻³, 7.87 μg·m⁻³, 4.65 μg·m⁻³, and 9.93 μg·m⁻³, respectively. The MAE between the concentrations of organics, sulfates, ammonium, and nitrates estimated by the CM model and the measured values in the Beijing station is lower than that of the single MLR model, SVR model, RF model, KNN model, and XGBoost model, and slightly higher than that of the LightGBM model.

The MAE between the concentrations of organics, sulfates, ammonium, and nitrates estimated by the CM model and the measured values at the Xianghe site was 17.23 μg·m⁻³, 10.17 μg·m⁻³, 5.47 μg·m⁻³, and 11.61 μg·m⁻³, respectively. Compared with the single model, it was found that the MAE between the concentration of organics, sulfates, and nitrates estimated by the CM model and the measured value was lower; the MAE between the estimated concentration of ammonium and the measured value was lower than that of the MLR model, SVR model, KNN model, and XGBoost model, and slightly higher than that of the LightGBM model and RF model.

The MAE between the concentrations of organics, sulfates, ammonium, and nitrates estimated by the CM model and the measured values at the Xinglong site was 24.05 μg·m⁻³, 12.58 μg·m⁻³, 5.71 μg·m⁻³, and 16.64 μg·m⁻³, respectively. Compared with the single model, it is found that the MAE between the organics concentration and the measured value of the Xinglong site estimated by the CM model is lower than that of most single models, while the MAE between the estimated sulfate and ammonium concentration and the measured value is lower than that of other single models except the LightGBM model. The MAE between the estimated nitrate concentration and the measured value was lower than that of the other single models except the RF model.

In summary, except for the LightGBM model and RF model, the average absolute error MAE between the concentrations of organics, sulfates, ammonium, and nitrates at the Beijing, Xianghe, and Xinglong sites estimated by the CM model and the measured values is smaller. The MAE magnitude of the mean absolute error in Beijing, Xianghe, and Xinglong is, in turn, Beijing < Xianghe < Xinglong, which is consistent with the six single models.

Beijing represents the urban areas of the BTH region, Xianghe represents the suburban areas, and Xinglong represents the rural areas. Since the model was established in Beijing, its accuracy will decline when applied to different sites, especially in Xinglong, which is farther away and has a lower air pollution level. The R² between the combined model simulated results and observations for organics decreased from 0.84 in Beijing to 0.52 in Xinglong. We evaluate the accuracy level of our model by comparing it with other studies. Zhang et al. estimated the atmospheric columnar organics mass concentration from remote sensing measurements of aerosol spectral refractive indices using bimodal parameters. The R² between the remote sensing results of organics in Beijing and the ground observation values was 0.35 [51]; Xie et al. established a comprehensive aerosol composition model to quantify black carbon (BC), brown carbon (BrC), mineral dust (DU), particulate organic matters, ammonium sulfate like (AS), sea salt, and aerosol water uptake. The R² between the remote sensing results of AS in the Beijing suburbs and the ground observed sulfates is 0.67, and the R² between the remote sensing results of AS and the ground observed nitrates is 0.71 [52]; Van Beelen et al. derived aerosol water and chemical composition using a modeling approach that combines individual measurements of remotely sensed aerosol properties (e.g., optical thickness, single-scattering albedo, refractive index, and size distribution) from an AERONET (Aerosol Robotic Network) Sun-sky radiometer with radiosonde measurements of relative humidity. The R² between the remote sensing results of organic matter in rural areas and the ground observation values is 0.42, and the R² between the remote sensing results of inorganic salts and the ground observation values is 0.35 [53].

3.6. Chemical Composition Concentration Estimation Based on Combination Model and Spatial Distribution Analysis

In order to estimate the annual and seasonal mean concentrations of aerosol chemical components in the Beijing–Tianjin–Hebei region, this study took 2012 as an example and estimated the daily concentrations of organics, sulfates, ammonium, and nitrates by using the CM model. By averaging the estimated daily concentrations of organics, sulfates, ammonium, and nitrates, the average annual and seasonal concentrations of organics, sulfates, ammonium, and nitrates in the Beijing–Tianjin–Hebei region in 2012 were finally obtained.

We input the gridded MODIS AOD data and meteorological data into the CM model to obtain the spatial distribution of the mass concentration of aerosol chemical components. Figure 5 shows the spatial distribution of the average annual mass concentration of organics, sulfates, ammonium, and nitrates. From the spatial perspective, these four chemical components in the Beijing–Tianjin–Hebei region exhibit similar distribution patterns, with higher values in the southwest but low values in the northwest. This is attributed to the more economically developed southwestern area of the BTH region, which features more intensive industrial activity, higher population density, and greater pollutant emissions from extensive human activities. Figure 5a shows the spatial distribution of the average annual mass concentration of organics. The organic matter concentration in the southwestern region of Beijing–Tianjin–Hebei exceeded 28 µg·m⁻³, with some areas even exceeding 33 µg·m⁻³. Figure 5b shows the spatial distribution of the annual mass concentration of sulfates. The average sulfates concentration in the south part of the BTH region is significantly higher than that in the north part, with the average annual concentration above 7 µg·m⁻³. Figure 5c shows the spatial distribution of the average annual mass concentration of ammonium. The average annual concentration of ammonium in the southwest of the BTH region is above 8 µg·m⁻³ and reached more than 10 µg·m⁻³ in some areas. Figure 5d shows the spatial distribution of the annual mass concentration of nitrates. In the whole study area, the highest average annual nitrates concentration was mainly located in the southwest part, and the average annual nitrates concentration was above 15 µg·m⁻³ and exceeded 18 µg·m⁻³ in some areas. Overall, the concentration of nitrates is higher than that of sulfates. Sulfates primarily originate from stationary sources, whereas nitrates are predominantly derived from traffic sources. Therefore, the ratio of sulfates to nitrates is often used to indicate the relative contributions of stationary and traffic sources [23]. When this ratio exceeds 1, it suggests that traffic sources contribute more to the total PM concentration in the BTH region.

As observed from Figure 6, the spatial distribution of the concentrations of the four aerosol chemical components exhibits distinct seasonal variations. The average concentrations of organics and sulfates in winter were significantly higher than those in the other three seasons. The areas with higher average concentrations in winter were mainly distributed in the south area of the BTH region, where the concentrations of organics and sulfates were above 30 µg·m⁻³ and 15 µg·m⁻³, respectively. The main reason for this phenomenon is that winter heating requires burning a large amount of fossil energy such as coal, which leads to excessive emissions of pollutants. The mass concentrations of the four components were all relatively low in summer. The spatial mean of organics concentration is about 20 µg·m⁻³, which is associated with the high humidity and precipitation in summer. The spatial distribution of nitrates in the Beijing–Tianjin–Hebei region was significantly higher in autumn and spring than in summer. Nitrates are a typical product of photochemical reactions. In autumn, strong radiation facilitates the formation of nitrate, and moderate temperatures promote its retention in the atmosphere [23].

4. Conclusions

Through the analysis of the chemical composition of aerosols in the Beijing area from 2012 to 2013, it was found that the annual average concentration of NR-PM1 was 41.32 µg·m⁻³, the annual average concentration of organics was 20.37 µg/m³, accounting for 49.3% of NR-PM1, the largest among the four components, followed by nitrates, sulfates, and ammonium. From the seasonal change, the Beijing area was most polluted in autumn and winter, followed by spring, and the lightest in summer. Through the comparison of different single models and combination models, it was found that four single models—RF, KNN, XGBoost, and LightGBM—performed relatively well, but the CM model had higher accuracy with smaller MAE and higher R². Therefore, the spatial distribution of aerosol chemical composition concentration obtained by using the CM model is the most reasonable. Based on the CM model, the estimated concentrations of the four components in the Beijing, Xinglong, and Xianghe areas are basically consistent with the measured values. The spatial distribution of the annual average concentrations of organics, sulfates, ammonium, and nitrates overall shows a south-high-north-low pattern. The annual average concentrations of the four components in the southwestern part of Beijing are the highest, ranging from more than 28 µg·m⁻³ for organics to more than 15 µg·m⁻³ for nitrates. From the spatial distribution of the annual average concentrations of organics, sulfates, ammonium, and nitrates in the four seasons, the concentrations of organics and sulfates are relatively consistent in all seasons, while the distribution characteristics of ammonium and nitrates are similar. The spatial distribution of the annual average concentrations of the four components overall shows a south-high-north-low pattern, which is similar to the spatial distribution of the annual average concentrations of organics, sulfates, ammonium, and nitrates.