Multi-Source Integration for Assessing Air Quality Dynamics in China: The Interplay of Anthropogenic Drivers, Meteorology, and Topography

Abstract

Air pollution remains a major public health and environmental challenge in China, driven by complex non-linear interactions among anthropogenic activities, meteorological conditions, and topographic features that go beyond simple linear relationships. This study presents a comprehensive spatio-temporal assessment of key air pollutants (CO, NO₂, SO₂, and PM_2.5) and their relationships with Total Column Ozone (TCO) across China’s provinces from 2019 to 2023. Multi-source high-resolution satellite data from Sentinel-5P/TROPOMI, the China High PM_2.5 dataset, MODIS, and ERA5-Land reanalysis were integrated. A tiered analytical framework was applied, combining linear Pearson correlations, non-linear Spearman rank correlations, and interpretable XGBoost machine learning with SHAP values. Results reveal a distinct seasonal “seesaw” pattern, with primary pollutants peaking during winter stagnation and TCO reaching maximum levels in late winter and spring. Non-linear analyses uncover critical threshold effects, including exponential increases in PM_2.5 and SO₂ when surface temperatures drop below 0 °C, very strong SO₂-TCO coupling (ρ = 0.93), and significant pollutant trapping in low-elevation regions (CO-elevation ρ = −0.82). These findings support the development of precision environmental policies with dynamic, temperature-threshold-based emission controls and topography-specific strategies to effectively mitigate air pollution in China.

1. Introduction

Air quality is a significant environmental issue, especially in urban areas, where industrial processes, vehicle exhaust, and high population density are major contributors to its deterioration [1,2,3,4]. Among the various pollutants, carbon monoxide (CO), sulfur dioxide (SO₂), nitrogen dioxide (NO₂), and PM_2.5 are considered primary indicators due to their hazardous impacts on both human health and the environment [5,6,7,8,9].

A complex interplay of natural and anthropogenic factors governs the dynamics of these pollutants. Meteorological conditions, such as surface temperature, wind speed, and direction, crucially influence their dispersion and transformation [10,11,12,13,14]. Furthermore, topographic features including elevation, slope, and aspect can significantly alter local airflow patterns, potentially trapping pollutants in basins or facilitating their dispersal [15,16]. Concurrently, socio-economic activities such as industrial production, construction, and transportation are major contributors to emissions, while ecological factors such as vegetation cover can act as natural sinks [17,18,19].

To analyze these multi-scale processes, satellite remote sensing has been increasingly employed to overcome the spatial limitations of ground-based monitoring. Instruments like the Tropospheric Monitoring Instrument (TROPOMI) onboard the Sentinel-5P satellite provide high-resolution data on trace gases (e.g., NO₂, SO₂, CO), enabling the detection of emission sources and transport patterns that sparse surface stations might miss [20,21]. Similarly, the Moderate Resolution Imaging Spectroradiometer (MODIS) provides critical observations of aerosol optical depth (AOD) and active fires, facilitating the tracking of particulate matter and biomass-burning events [22]. Studies by Alam et al. [23] and Huang et al. [24] have demonstrated the power of such data to reveal seasonal variability, long-term trends, and the efficacy of pollution control policies. The integration of satellite data with reanalysis products (e.g., ERA5-Land) and socio-economic datasets has thus become a fundamental approach in atmospheric research [25,26].

Despite these advancements, significant knowledge gaps persist regarding the drivers of air pollution in China. A substantial portion of existing literature exhibits methodological or spatial limitations that hinder comprehensive national-scale analysis. Early foundational studies were often geographically constrained to megacities such as Beijing and Shanghai, establishing significant correlations but failing to capture broader, heterogeneous patterns [27]. While subsequent research expanded to national scales, many studies remained heavily reliant on ground-station data and spatial interpolation techniques (e.g., Kriging), which can introduce significant bias in regions with sparse networks or complex terrain [28]. Analytically, the preponderance of this work employs conventional linear statistical methods to link pollutant concentrations with potential drivers [29,30,31,32,33]. These linear approaches often fail to capture the non-linear interactions inherent in atmospheric systems, such as climatic threshold effects, synergistic pollutant interactions, or saturation points in socio-economic impacts [34]. Although recent work has begun to explore non-linear relationships—particularly in health impact studies using models such as Distributed Lag Non-linear Models (DLNM) [35,36]—a critical gap remains. Few studies apply sophisticated, interpretable non-linear frameworks to disentangle the complex influences of topography, meteorology, and human activity on pollution concentrations themselves, leveraging the synoptic coverage provided by modern satellite data.

To address these gaps, this study proposes an innovative synoptic approach that integrates multi-source satellite observations with interpretable machine learning. By synthesizing high-resolution data from the Sentinel-5P TROPOMI and MODIS instruments with ERA5-Land reanalysis fields, we overcome the spatial limitations of ground-based monitoring. The novelty of this research lies in its tiered analytical framework, which contrasts traditional linear statistics (Pearson correlation) and monotonic rankings (Spearman correlation) with non-linear techniques—specifically XGBoost regression and SHAP (Shapley Additive exPlanations) analysis.

Specifically, this study integrates data from 2019 to 2023 to accomplish three primary objectives:

• To characterize spatial patterns and seasonal trends of air quality using continuous satellite coverage to overcome the limitations of sparse ground station interpolation.
• To assess natural factors, specifically investigating how meteorological conditions (temperature, wind) and often-overlooked topographic features (elevation, slope) constrain pollutant dispersion.
• To investigate the correlation of air pollution intensity with anthropogenic drivers, namely socio-economic activities (population, industry, transport, vehicle density, construction output) and ecological vegetation cover, employing a tiered analytical framework that contrasts linear Pearson and monotonic Spearman correlations with interpretable machine learning (XGBoost) to capture complex threshold effects missed by traditional statistical models.

2. Materials and Methods

2.1. Overview of the Study Area

The People’s Republic of China, located in East Asia, spans approximately 9.6 million square kilometers, stretching from latitudes 18° N to 54° N and longitudes 73° E to 135° E as shown in Figure 1. The country features a notable topographic gradient that ranges from lowland fertile plains in the east, around 4 m above sea level in coastal provinces like Shanghai and Jiangsu, to highland plateaus and the Himalayas in the west, where elevations exceed 4700 m in Tibet and Qinghai. As altitude increases, the terrain becomes more rugged; eastern coastal areas are relatively flat, while the mountainous regions of Sichuan and Yunnan have steeper slopes. This variety leads to a complex topographic and climatic pattern.

For this research, the focus is on the Chinese mainland, excluding Taiwan Province and the Special Administrative Regions of Hong Kong and Macao, to account for variations in administrative reporting and data.

2.2. Basic Data

This study examines the spatial distribution of key air pollutants—carbon monoxide (CO), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), and particulate matter (PM_2.5)—as well as Total Column Ozone (TCO) as shown in

Table 1. Summary of Data Types, Sensors, and Spatial Resolutions.

Data Type	Type	Source	Spatial Resolution
Atmospheric Composition	CO	Sentinel-5P TROPOMI	7 km × 7 km
	NO2	Sentinel-5P TROPOMI	5.5 km × 3.5 km
	SO2	Sentinel-5P TROPOMI	7 km × 5.5 km
	TCO	Sentinel-5P TROPOMI	7 km × 3.5 km
	PM2.5	China High PM2.5 Dataset	1 km × 1 km
meteorological factors	Surface Temperature	ERA5-Land	0.1° (~9 km)
	wind speed	ERA5-Land	0.1° (~9 km)
	wind direction	ERA5-Land	0.1° (~9 km)
topographic factors	DEM	SRTM	30 m
	Slope	Derived from SRTM DEM	30
	Aspect	Derived from SRTM DEM	30
socio-economic	population	Statistical Yearbook/Administrative data	Administrative unit
	industry output	Statistical Yearbook/Administrative data	Administrative unit
	construction output	Statistical Yearbook/Administrative data	Administrative unit
	transport output	Statistical Yearbook/Administrative data	Administrative unit
	total number of vehicles	Statistical Yearbook/Administrative data	Administrative unit
Ecological Factors	vegetation area	MODIS MOD13A3	1 km

across Chinese provinces from 2019 to 2023. We explore their interactions with meteorological variables, topographic features, ecological indicators, and socio-economic factors, including population density and industrial output, to enhance our understanding of air quality dynamics.

2.2.1. Atmospheric Composition

Data for carbon monoxide (CO), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), and total carbon (TCO) were obtained from the Sentinel-5P satellite. Additionally, monthly mean data from the China High PM_2.5 datasets were downloaded for the period from January 2019 to December 2023. This approach was taken to minimize short-term variability and ensure consistency with other environmental datasets.

2.2.2. Meteorological Factors

Data on surface temperature, wind speed, and wind direction were collected from the ECMWF ERA5-Land, which provides high-quality reanalysis meteorological fields. Monthly averages were computed to align with the temporal resolution of the pollution and vegetation datasets.

2.2.3. Socio-Economic Factors

Provincial-level socio-economic indicators—such as population, industrial output, construction output, transportation output, and the total number of vehicles—were obtained from the National Bureau of Statistics of China (NBS). All socio-economic data were matched to the corresponding environmental data across the same study period.

2.2.4. Topographic Factors

Digital Elevation Model (DEM) data were sourced from the Shuttle Radar Topography Mission (SRTM) dataset, which is available through Google Earth Engine (GEE) at a spatial resolution of 30 m. Slope and aspect were derived directly within GEE using its built-in terrain analysis functions. All topographic layers were then resampled in GEE to match the spatial resolution of the air pollutant datasets, ensuring spatial consistency across all variables.

2.2.5. Ecological Factors

Ecological factors were represented using vegetation indices derived from the MODIS MOD13A3 product via Google Earth Engine (GEE). This product provides monthly Normalized Difference Vegetation Index (NDVI) data at a 1 km resolution, allowing for reliable, cloud-screened vegetation information from January 2019 to December 2023. All vegetation data were processed in GEE for compatibility with air pollutant layers. Data on carbon monoxide (CO), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), and total carbon (TCO) were sourced from the Sentinel-5P satellite, along with monthly mean data from the China High PM_2.5 datasets for the same period to ensure consistency.

2.3. Overall Technical Approach and Analysis Methods

The technical workflow prioritized spatial and temporal consistency across multi-source datasets through systematic standardization and processing as shown in Figure 2. All datasets, including satellite imagery, meteorological fields, and administrative boundaries, were projected to the WGS 1984 coordinate system (EPSG:4326). Reprojection was performed using a dual-platform approach: Google Earth Engine (GEE) automated the reprojection of Sentinel-5P, ERA5-Land, MODIS, and SRTM data, while Python (Google Colab) with rioxarray and geopandas handled the high-resolution China PM_2.5 dataset and provincial boundaries.

Following coordinate alignment, spatial resolutions were harmonized. Within GEE, gaseous pollutants (CO, NO₂, SO₂, TCO), meteorological variables, and topography layers were resampled to a uniform 7 km × 7 km grid to match the Sentinel-5P TROPOMI footprint. For PM_2.5, the native 1 km × 1 km China High PM_2.5 dataset was retained in Google Colab, and provincial means were derived via zonal statistics without resampling, preserving local accuracy while enabling aggregation.

2.3.1. Variable Derivation and Calculation

Several environmental indicators were derived from raw satellite bands and vector components to ensure physical meaningfulness:

• (a) Vegetation (NDVI) The Normalized Difference Vegetation Index (NDVI) was calculated using the Near-Infrared (NIR) and Red spectral bands from the MODIS dataset to assess vegetation health and density as shown in Equation (1).

$$\mathit{N}\mathit{D}\mathit{V}\mathit{I}={\displaystyle \frac{{\mathit{P}}_{\mathit{N}\mathit{I}\mathit{R}}-{\mathit{P}}_{\mathit{R}\mathit{E}\mathit{D}}}{{\mathit{P}}_{\mathit{N}\mathit{I}\mathit{R}}+{\mathit{P}}_{\mathit{R}\mathit{E}\mathit{D}}}}$$

(1)

where $P_{NIR}$and $P_{RED}$ represent the surface reflectance in the near-infrared and red bands, respectively.

• (b) Meteorological Factors (Wind and Temperature) Wind parameters were derived from the zonal (u) and meridional (v) components of the 10 m wind vector provided by the ERA5-Land dataset. Wind Speed (WS) was calculated as the magnitude of the wind vector as shown in Equation (2).

$$\mathit{W}\mathit{s}=\sqrt{{\mathit{u}}^2+{\mathit{v}}^2}$$

(2)

Wind Direction (WD) was calculated to represent the meteorological direction (the direction from which the wind is blowing, measured clockwise from North). This was derived using the four-quadrant inverse tangent function (atan2) as shown in Equation (3).

$$\mathit{W}\mathit{D}=\left({\displaystyle \frac{\mathbf{180}}{\mathit{\pi }}} .\mathit{a}\mathit{t}\mathit{a}\mathit{n}\mathbf{2}\left(-\mathit{u},-\mathit{v}\right)+360\right)(\mathit{m}\mathit{o}\mathit{d} \mathbf{360})$$

(3)

where u represents the eastward component and v represents the northward component. The modulo operation ensures the final direction remains within the 0–360° range.

• Additionally, surface temperature data were converted from Kelvin (K) to degrees Celsius (C) for consistency with standard reporting units as shown in Equation (4).

$${\mathit{T}}_{\mathit{c}}={\mathit{T}}_{\mathit{k}}-\mathbf{273.15}$$

(4)

• (c) Topographic Slope: Slope was calculated from the SRTM Digital Elevation Model (DEM) based on the rate of change in elevation in the horizontal (x) and vertical (y) directions as shown in Equation (5).

$$\mathit{S}\mathit{l}\mathit{o}\mathit{p}\mathit{e}=\mathbf{arctan}(\sqrt{({{\displaystyle \frac{\partial \mathit{z}}{\partial \mathit{y}}})}^2+({{\displaystyle \frac{\partial \mathit{z}}{\partial \mathit{x}}})}^2})\times {\displaystyle \frac{180}{\mathit{\pi }}}$$

(5)

Aspect identifies the downslope direction of the maximum rate of change in value from each cell to its neighbors. It was calculated using the inverse tangent function of the gradients as shown in Equation (6).

$$\mathit{A}\mathit{s}\mathit{p}\mathit{e}\mathit{c}\mathit{t}=\left({\displaystyle \frac{\mathbf{180}}{\mathit{\pi }}} .\mathit{a}\mathit{t}\mathit{a}\mathit{n}\mathbf{2} \left({\displaystyle \frac{\partial \mathit{z}}{\partial \mathit{x}}},{\displaystyle \frac{\partial \mathit{z}}{\partial \mathit{y}}}\right)\right)$$

(6)

2.3.2. Spatial and Temporal Aggregation

The calculation of mean values was performed using a multi-platform strategy that handled the specific data formats and computational requirements of each variable. Within the Google Earth Engine (GEE) platform, monthly mean values for gaseous pollutants (CO, NO₂, SO₂), Total Column Ozone (TCO), surface temperature, wind speed, wind direction, NDVI, and topographic factors were computed using cloud-based parallel processing to ensure temporal consistency across the five-year study period. For the PM_2.5 data, a specialized Python workflow was implemented in Google Colab. This process used the xarray and netCDF4 libraries to manage high-resolution NetCDF files, and the rasterstats library to compute zonal statistics. This allowed for the precise extraction of monthly mean values for each province by masking the raster data with the corresponding administrative shapefiles.

The visualization and mapping phase used a hybrid approach that leveraged the cartographic precision of ArcGIS Pro 3.5.2 and the analytical flexibility of Python. ArcGIS Pro, in conjunction with Python scripts, was used to generate high-quality cartographic maps illustrating the spatial distribution of all pollutants, including PM_2.5, at both monthly and seasonal scales. These maps, which also incorporated meteorological and environmental variables across China’s provinces, were instrumental in identifying regional trends and pollution “hotspots.” Additionally, Python’s specialized libraries were utilized to produce analytical maps and complex statistical charts to effectively visualize the temporal evolution and dynamic fluctuations of the studied factors from 2019 to 2023.

2.3.3. Statistical Correlation Analysis

To identify the primary drivers of air pollution, processed environmental and socio-economic datasets were integrated using the Pandas library. The analysis followed a multi-stage approach: first, Pearson correlation coefficients (r) were calculated to evaluate linear relationships, followed by Spearman’s rank correlation (ρ) to assess monotonic trends and handle non-normal distributions. To capture complex, non-linear interactions, XGBoost regression models were developed for each pollutant. Finally, model results were interpreted using SHAP (Shapley Additive Explanations) values, which quantified the specific contribution and importance of each environmental and socio-economic factor to variations in air quality.

2.3.4. Machine Learning Configuration and Interpretability

To quantify non-linear interactions, we developed distinct XGBoost regression models for each pollutant using 13 meteorological, topographic, and socio-economic features. To prevent overfitting on the spatio-temporal dataset, we restricted model complexity using conservative hyperparameters (adept = 4, learning_rate = 0.03, and n_estimators = 300) and validated performance on a 10% independent test set held out from the 90% training data.

Robustness was confirmed via out-of-sample on the test set, ensuring the learned patterns were not artifacts of overfitting. SHAP (Shapley Additive exPlanations) analysis was then applied to these validated models to resolve their “black box” nature and quantify non-linear thresholds.

3. Results

3.1. Spatial and Seasonal Patterns of Air Pollutants and Environmental Drivers

The mean CO distribution across China reveals a transparent east–west gradient, with elevated concentrations in the eastern and southeastern regions and minimal values in the western highlands, as shown in Figure 3. This pattern corresponds with land surface temperature anomalies, showing positive values in the east and negative anomalies over the Tibetan Plateau. Seasonally, temperature and CO exhibit opposite trends: peak temperatures occur in July (Summer), especially in southern China, as shown in Figures S1 and S2, while CO concentrations peak in January over northern industrial regions and reach their lowest levels in July.

NO₂ and SO₂ mean concentrations are significantly higher in eastern and southeastern China, as shown in Figure 4. Both pollutants display pronounced seasonal variability, peaking in December, particularly across northern industrial areas. High concentrations are concentrated in provinces such as Hebei, Shanxi, and Jilin, as shown in Figures S2 and S3, before declining to minimum levels in July. Similarly, reaches its maximum values in northern port regions, including Tianjin and Shandong, in December and decreases to its lowest levels in August, especially over the western highlands of Tibet and Yunnan.

PM_2.5 and TCO mean concentrations are substantially higher in northern and central China, particularly over the North China Plain, and lowest in southern regions and western highlands, as shown in Figure 5. In contrast, mean Total Column Ozone (TCO) exhibits apparent latitudinal variability and a delayed seasonal peak relative to primary pollutants.

While concentrations peak in January, with provinces such as Henan and Shandong showing the highest values (Figures S2 and S4), levels decline to minima in July and August. Conversely, TCO reaches its maximum in late winter to early spring (February–March), peaking in northeastern provinces including Heilongjiang and Jilin, before decreasing to minimum values in late summer and early autumn, as shown in Figures S8 and S9.

The mean wind speed over China is higher in the northern and western regions, including the Mongolian and Tibetan Plateaus, and lowest across the central and southeastern plains, as shown in Figure 6. Wind direction shows substantial spatial variability driven by monsoonal circulation and topographic effects. Seasonally, wind speed peaks in late winter, as shown in Figure S6, with December recording the highest values in regions such as Hainan and Inner Mongolia (Figures S5 and S7), before declining to a minimum in August, particularly in interior provinces like Sichuan. Wind direction exhibits an apparent seasonal reversal, with prevailing northwesterly flows in northern cities during December shifting to southeasterly directions in July.

The mean Normalized Difference Vegetation Index (NDVI) over China indicates higher vegetation density and greener cover in the southern and southeastern regions, driven by subtropical climates, abundant rainfall, and dense forests, as shown in Figure 7. In contrast, the lowest NDVI values are observed in the northwestern deserts, the Tibetan Plateau, and arid northern areas, reflecting sparse vegetation and barren landscapes resulting from harsh climatic and topographic conditions.

3.2. Monthly and Seasonal Correlation Analysis of Drivers

Pearson and Spearman analyses (Tables S1 and S2, and Figures S8–S13) identify surface temperature as the primary regulator of air quality, exhibiting robust negative correlations with SO₂ ($r=-0.63,\rho =-0.65$) and PM_2.5 ($r=-0.55$) annually. This confirms a seasonal pattern where cold temperatures drive heating emissions and atmospheric stagnation. Wind speed consistently acts as a dispersion factor, showing a negative correlation with PM_2.5 across all seasons. Regarding inter-pollutant coupling, combustion markers NO₂ and $CO$ display the strongest annual association ($r=0.74$), reflecting shared industrial and vehicular sources. Notably, spring data reveals a unique meteorological synchronization between surface SO₂ and stratospheric TCO ($r=0.79$). This strong correlation is likely not due to direct chemical interactions, but rather driven by large-scale synoptic weather systems that influence the entire atmospheric column. The same meteorological conditions that favor surface pollutant accumulation (such as stable boundary layers) coincide with the seasonal peak in stratospheric ozone transport, resulting in simultaneous highs for both variables.

3.3. Annual Correlation Analysis of Drivers

Pearson and Spearman analyses (Table S3; Figures S14 and S15) reveal surface temperature as a complex driver, showing robust negative correlations with SO₂ (r = −0.65, ρ = −0.46) and PM_2.5 (r = −0.26, ρ = −0.63). Wind speed acts as a purification force with stronger negative association to PM_2.5 (r = −0.25, ρ = −0.43). Vegetation coverage exhibits non-linear mitigation of PM_2.5 (r = −0.45, ρ = −0.60). Topography (DEM) controls pollution deposition with strong negative correlation to CO (r = −0.9, ρ = −0.82). Socioeconomically, vehicle ownership couples with CO levels (ρ = 0.55). Inter-pollutant synergy is profound: SO₂-TCO strongest non-linearly (r = 0.51, ρ = 0.93) from synoptic weather; NO₂-CO peaks (ρ = 0.83), confirming combustion cohesion.

3.4. XGBoost and SHAP Analysis of Drivers

Based on the XGBoost model and SHAP (Shapley Additive exPlanations) analysis, the non-linear contributions of meteorological, topographic, and socioeconomic drivers were quantified, as shown in Figure 8. To validate the reliability of these non-linear insights, the model’s performance was evaluated using the coefficient of determination (R²), indicating high predictive accuracy for most pollutants, as shown in Table S4.

Temperature emerges as the most critical determinant for PM_2.5 and ${So}_2$, exhibiting a clear non-linear threshold effect where lower temperatures (below 10 °C) drastically increase output predicted values of pollutant variables values, corresponding to winter heating emissions. Specifically, the partial dependence plots reveal a sharp “elbow” pattern for PM_2.5 where SHAP values rise steeply as temperatures drop below 0 °C, confirming the non-linear impact of cold-stagnation events as shown in Figure 9 and Figure S16. For NO_2, Aspect (wind-facing direction) and Temperature are the top predictors, with vegetation also showing a strong mitigating effect where higher green cover leads to negative SHAP values as shown in Figure S17. In contrast, Elevation is the dominant factor for CO, showing a sharp decline in influence as altitude increases, confirming that carbon monoxide accumulates primarily in low-lying basins as shown in Figure S18. Furthermore, TCO displays a strong seasonal non-linearity, with SHAP values peaking in the warm months (May-August) and dropping significantly in winter, driven by photochemical reaction rates as shown in Figure 10. Regarding socioeconomic impacts, industrial output shows a saturation effect for NO₂ where pollution impacts rise rapidly with initial activity but plateau at extremely high output levels, suggesting a dispersion limit, while vehicle ownership shows a logarithmic-like increase in its contribution to CO levels.

4. Discussion

4.1. Spatial and Temporal Patterns of Air Quality Dynamics

Air quality in China exhibits pronounced spatial heterogeneity, with a distinct east–west gradient aligned with topographic and socioeconomic divides. The eastern and southeastern regions, particularly the North China Plain (Hebei, Henan, Shandong) and coastal hubs (Jiangsu, Shanghai), exhibit elevated levels of primary pollutants (CO, NO₂, SO₂, PM_2.5) due to high population density, industrial agglomeration, and vehicular emissions. In contrast, western highlands (e.g., Tibetan Plateau) maintain robust background air quality owing to limited anthropogenic activity. These patterns are modulated by topography, with low-lying basins (e.g., in Jiangsu, Hebei, Sichuan) acting as pollutant traps, as evidenced by CO accumulation in elevation-constrained zones. These regions are often hemmed in by rugged, high-level terrain, which facilitates local drainage flows and regional synoptic stagnation, further preventing the horizontal and vertical dispersion of pollutants.

Seasonally, air quality displays a “seesaw” pattern: primary pollutants (PM_2.5, SO₂, CO) peak in winter (December–January) due to heating emissions and stagnation under thermal inversions and low boundary layers. Conversely, total column ozone (TCO) peaks in late winter–spring, driven by Brewer-Dobson Circulation transporting stratospheric ozone, rather than surface photochemistry. These rhythms underscore the interplay of anthropogenic cycles and meteorological phenology in regional pollution dynamics.

4.2. Relationships Between Key Drivers and Air Quality

Our integrated analytical framework combining Pearson and Spearman correlations with XGBoost-SHAP machine learning reveals that while anthropogenic activities (e.g., industrial and urban emissions) establish the baseline pollution load, natural drivers such as temperature, topography, and vegetation critically amplify pollution spikes. Temperature exhibits a pronounced nonlinearity: although linear correlations indicate a general inverse relationship with pollutants, machine learning identifies a sharp threshold below 0 °C, where PM_2.5 and SO₂ concentrations increase exponentially. This reflects a feedback loop wherein cold stagnation and intensified heating emissions synergistically elevate pollution.

Wind speed acts as a consistent dispersion factor, showing significant negative correlations with PM_2.5 (r = −0.25, ρ = −0.43) and SO₂ (r = −0.65, ρ = −0.46), with a more substantial cleansing effect on SO₂. Topography further modulates pollution persistence, evidenced by a strong negative correlation between elevation and CO (ρ = −0.82, r = −0.90), confirming that low-lying basins serve as pollutant traps due to restricted airflow. Notably, the reduced CO levels observed in high-altitude regions such as the Tibetan Plateau are partially attributable to atmospheric thinning rather than cleanliness alone.

Vegetation cover (NDVI) functions as a natural sink, with steadily decreasing pollutant levels (NO₂, PM_2.5, CO) as greenness increases, as indicated by SHAP dependence plots—an effect more continuous than threshold-driven. Although industrial output is a dominant source of emissions, its relationship with air quality is not strictly linear. At the highest levels of production, pollution impacts tend to level off or saturate. This trend suggests that mature industrial zones may be decoupling production from emissions through improved technology adoption and higher efficiency standards, preventing pollution from rising infinitely. Additionally, a strong non-linear coupling between SO₂ and total column ozone (TCO, ρ = 0.93) suggests that synoptic weather systems concurrently influence both surface pollutants and stratospheric ozone, rather than shared emission pathways.

4.3. Comparison with National and Regional Studies

Our findings on spatio-temporal factors influencing air pollution in China validate and extend prior estimates in this domain, aligning with the documented decline in primary pollutants (PM_2.5 and SO₂) and rise in total column ozone (TCO) reported by Liu et al. [28] and Zhang et al. [27]. While these studies identified meteorological stagnation as a key winter driver using linear regression or panel models, our XGBoost and SHAP analyses reveal complex non-linear relationships, including a thermal threshold of 0 °C below which PM_2.5 and SO₂ concentrations increase in a nuance undetectable by linear approaches. This threshold elucidates the “cold-dry” conditions associated with elevated respiratory disease risks, as observed in recent epidemiological reports, such as influenza outbreaks in Shanghai and Huai’an, reported by Si et al. [32] and Wang et al. [30].

Regarding sources, winter peaks stem from stagnation and heating emissions. Still, autumn-winter elevation of pollutant concentration also involves agricultural straw open-field burning (ASOB), as noted by Huang et al. [29], contributing to a “seasonal seesaw” pattern via synoptic systems that trap pollutants and influence stratospheric ozone. This non-linear coupling between contaminants (e.g., SO₂ and O₃) and meteorology contrasts with traditional linear correlations [27,28]. It aligns with synergistic health impacts, including HFRS incidence Kuang et al. and tuberculosis Sun et al. Overall, non-linear methods better capture thresholds (<0 °C) and holistic interactions, thereby improving air quality and health assessments [30,31,32,33].

4.4. Limitations of This Study and Future Directions

This study provides a multi-scale, multi-source analysis of air quality drivers across China, but is subject to several limitations.

First, the use of monthly aggregated data, while suitable for identifying seasonal and climatological patterns, obscures short-term, high-intensity pollution episodes (e.g., 1–3-day haze events). Such acute events are critical for health impact assessments and are often driven by sub-daily meteorological variations that monthly averages cannot capture.

Second, although high-resolution inputs were employed (e.g., 1 km PM_2.5 data), spatial aggregation to provincial means masks significant intra-regional variability. This approach may underestimate exposure in densely populated urban centers while overrepresenting pollution levels in rural areas, limiting the precision of localized exposure assessments.

For future research, we recommend moving beyond broad administrative units by adopting county-level or pixel-based analyses to preserve spatial heterogeneity and provide higher granularity within each province. Direct validation against in situ ground measurements should be prioritized to quantify the accuracy of satellite-derived products. Furthermore, incorporating precipitation data is essential, as wet deposition plays a key role in removing particulate matter and soluble gases, a process not fully addressed here. Finally, integrating chemical transport models (CTMs) with interpretable machine learning (XAI) could further disentangle the complex causal pathways underlying air quality dynamics.

5. Conclusions

This study presents an extensive spatio-temporal analysis of air quality across China from 2019 to 2023 by integrating multi-source satellite observations (Sentinel-5P, MODIS), ERA5-Land meteorological data, and provincial socio-economic datasets. By advancing beyond linear statistics to incorporate Spearman rank correlations and XGBoost-SHAP machine learning, we elucidate the non-linear interactions between anthropogenic drivers and natural modulating factors.

Key Findings

• Distinct seasonal variability, with primary pollutants (CO, NO₂, SO₂, PM_2.5), peaks in winter due to combustion emissions and atmospheric stagnation, while total column ozone (TCO) peaks in late winter–spring via stratospheric transport.
• Temperature exhibits a non-linear threshold effect: PM_2.5 and SO₂ increase exponentially below 0 °C. Topography strongly modulates pollution, with low-lying basins acting as pollutant traps (e.g., ρ = −0.82 between elevation and CO).
• Socio-economic activity drives baseline emissions but shows saturation at high industrial output, whereas vegetation (NDVI) serves as a non-linear sink for NO₂ and CO.
• Potent pollutant–meteorology couplings, such as between SO₂ and TCO (ρ = 0.93), highlight systemic atmospheric linkages beyond shared emission sources.

These results underscore that while anthropogenic sources set emission baselines, natural filters (wind, vegetation) and traps (topography, thermal inversions) critically shape exposure intensity. We therefore advocate for precision environmental policies that target specific thermal thresholds (e.g., sub-zero periods) and topographically sensitive regions, rather than uniform emission caps, to more effectively enhance air quality mitigation.