Identification of Local and Transboundary Sources and Mechanisms of PM_2.5 and O₃ Pollution on the Tibetan Plateau: Implications for Sustainable Air Quality Governance

Abstract

Air pollution, particularly fine particulate matter (PM_2.5) and ozone (O₃) pollution, poses serious challenges to environmental quality and sustainable development. The Tibetan Plateau, often described as the “Third Pole,” functions as a key ecological shield for China and exerts wide-reaching influence on global climate systems, hydrological cycles, and cross-regional pollution transport. To better clarify the driving mechanisms of air pollution in this sensitive region, we propose an integrated MRG–HSW framework, which, for the first time, systematically couples statistical modeling and trajectory analysis by combining multivariate regression, residual-based screening, and HYSPLIT–WCWT trajectory analyses. Taking Qinghai Province as a case study, ERA5 and GDAS1 reanalysis products were coupled with in situ monitoring to identify the relative contributions of local emissions and long-range atmospheric transport. The results show that, in low-elevation zones, PM_2.5 levels are largely governed by local anthropogenic activities (R² = 0.631–0.803), whereas O₃ concentrations respond more strongly to meteorological variability (R² = 0.529–0.779). At higher elevations, however, local explanatory factors weaken, and long-range transport from the Hexi Corridor, Qaidam Basin, and even South Asia becomes the dominant influence for both pollutants. Additional sensitivity tests confirm that the framework performs robustly under diverse meteorological and seasonal conditions. Collectively, this work not only establishes a transferable methodology for source attribution in plateau environments but also underscores the pivotal role of the Tibetan Plateau in sustaining regional air quality and global environmental stability.

1. Introduction

In recent years, PM_2.5 and O₃ have been identified as two significant global air pollutants, presenting severe risks to human health and exerting substantial effects on ecological systems [1,2]. In response to acute air pollution issues, China initiated the Air Pollution Prevention and Control Action Plan in 2013, which aimed to enhance air quality via rigorous emission reduction strategies. Subsequent monitoring data revealed that, after the implementation of this plan, anthropogenic PM_2.5 emissions diminished by approximately 33%, which corresponded with a marked amelioration in levels of particulate pollution [3,4]. However, surface O₃ concentrations continued to rise, increasing by about 7.5% nationwide from 2015 to 2024 according to the China Ecological and Environmental Status Bulletin [5,6,7]. These findings highlight the ongoing challenge of jointly controlling PM_2.5 and O₃ pollution despite notable emission reductions [8].

In recent years, studies on the local attribution of PM_2.5 and O₃ pollution have shifted from empirical statistics toward multi-source integration and process-based interpretation. Early works mainly employed multiple linear regression (MLR) models to quantify the linear responses of pollutants to meteorological and emission factors. These models are simple and highly interpretable, enabling clear identification of key drivers and their contribution directions. For instance, national-scale trend analyses have shown that the decline in PM_2.5 concentrations was primarily driven by emission reductions, while O₃ variations were jointly influenced by meteorological conditions and photochemical processes [9]. At the urban scale, MLR has been used to reveal the differentiated impacts of temperature, humidity, wind speed, and boundary layer height on PM_2.5 and O₃, providing quantitative evidence for local air pollution control [10]. As understanding of nonlinear responses has deepened, the generalized additive model (GAM) has been introduced to overcome the linearity assumption of MLR. By using nonparametric smoothing functions, GAM captures nonlinear and threshold effects of meteorological variables, significantly improving explanatory power and fitting accuracy. Studies have identified a “threshold-enhancing” effect of temperature, solar radiation, and boundary layer height on O₃, while wind speed and precipitation exhibit “cleansing-dilution” patterns [11]. Compared with gradient boosting models (GBM) and other machine learning approaches, GAM shows greater advantages in causal interpretation and physical consistency [12]. In recent years, machine learning methods have been widely applied for multi-source data fusion and factor importance analysis. Models such as CatBoost, XGBoost, and random forest (RF) can identify major driving factors within complex multidimensional datasets and quantify their contributions, revealing nonlinear relationships between pollutants, meteorological, and socioeconomic factors [13,14]. Distributed lag nonlinear models (DLNM) further describe the lagged effects of meteorological variables [15]. However, despite their high accuracy, machine learning approaches often suffer from limited physical interpretability and strong dependence on sample size, particularly in data-scarce or topographically complex regions. To balance physical constraints and statistical flexibility, researchers have proposed hybrid statistical–physical frameworks that combine outputs from chemical transport models (CTMs) such as GEOS-Chem and WRF-Chem for bias correction and weighted integration, effectively reducing prediction errors [16]. Overall, local attribution studies are evolving from traditional linear statistics toward multi-source information fusion and mechanism-based modeling. Given the need for interpretability and robustness under data-scarce conditions, this study employs MLR and GAM to quantify both linear and nonlinear relationships between meteorological drivers and pollutant concentrations, ensuring physically consistent and mechanism-oriented analysis suitable for the Tibetan Plateau.

Meanwhile, research on external transport of air pollution has developed into a multi-level framework centered on trajectory analysis, concentration-weighted trajectory models (PSCF/CWT/WCWT), and CTMs. Receptor-based approaches using backward trajectories remain classical tools for identifying cross-regional transport pathways, typically employing the HYSPLIT model to compute 72 h trajectories and conduct clustering analyses that reveal dominant airflow channels [17]. Such clustering has been applied not only in urban environments but also in mountainous and plateau regions to distinguish between long-range and short-range seasonal transport [18,19]. The choice of clustering algorithm can influence pathway classification [20], while the integration of mobile and multi-site observations enhances spatial resolution [21]. For source region identification, weighted trajectory models link high concentration events with trajectory endpoints to highlight potential source areas [19]. The concentration-weighted trajectory (CWT) method has been widely applied to integrate satellite and ground-based observations for identifying major source regions [22], though its stability is sensitive to sampling resolution [23]. The weighted concentration-weighted trajectory (WCWT) further introduces concentration-weighted averaging, enabling finer quantifications of grid-level contributions and complementing clustering results; self-organizing maps (SOM) emphasize “path structures,” whereas WCWT focuses on “potential source regions” [21]. Chemical transport models such as WRF-Chem, CMAQ/CAMx, and GEOS-Chem can mechanistically evaluate emissions, transport, and chemical processes [24,25], though their application is often constrained by the need for detailed emission inventories and high computational demand, especially in data-limited regions like the Tibetan Plateau. Recent studies have combined numerical simulations with trajectory-based statistics to improve transport pathway identification [26,27], and hybrid approaches integrating “CTM + trajectory/receptor + machine learning” are emerging as efficient and interpretable tools for source attribution. Considering the sparse observation network and complex airflow over the Plateau, this study integrates HYSPLIT and WCWT analyses to trace transport pathways and identify dominant external source regions, providing a physically grounded yet computationally efficient approach to characterize transboundary influences.

Taken together, these developments demonstrate a clear methodological shift from single-technique analyses toward integrated frameworks that couple local and regional processes. Yet, most existing studies remain focused on lowland urban areas, with limited exploration of how local meteorology, emissions, and long-range transport interact under the complex topography and low-emission conditions of the Tibetan Plateau. To fill this gap, an integrated approach that links statistical interpretation with physical process representation is required to better distinguish local contributions from external transport influences.

To address these gaps, this study takes Qinghai Province as a representative region of the Tibetan Plateau and develops an integrated statistical–trajectory framework (MRG–HSW) that links local attribution with regional transport analysis. The goal is to better understand the mechanisms that drive PM_2.5 and O₃ pollution over the Tibetan Plateau. Specifically, the study aims to:

(1)

Develop an integrated MRG–HSW framework that couples statistical modeling with trajectory-based diagnostics;

(2)

Quantify the effects of local meteorological and emission-related factors on local pollutant concentrations;

(3)

Identify dominant source regions and transport pathways during high-pollution episodes;

(4)

Evaluate elevation-dependent mechanisms to support evidence-based and cooperative air-quality governance in complex plateau terrain.

Overall, this study contributes a systematic and scalable analytical framework that advances the understanding of pollution formation on the Tibetan Plateau and offers new scientific evidence for regional emission mitigation and transboundary air pollution management.

2. Materials and Methods

This study explores the formation mechanisms of PM_2.5 and O₃ pollution in Qinghai Province, located in the northeastern part of the Tibetan Plateau, which exhibits typical plateau characteristics and provides comprehensive air quality monitoring data. Therefore, it can serve as a representative region for studying the atmospheric environmental characteristics of the broader Tibetan Plateau. Qinghai Province, situated at the transition zone of the Plateau, shares similar geographical, climatic, and ecological conditions with the Plateau’s main area, and has high data continuity. Thus, selecting Qinghai as the study area effectively represents the pollution characteristics and formation mechanisms of the entire Tibetan Plateau. Analyses were conducted across eight prefecture-level regions, as illustrated in Figure 1.

2.1. Meteorological and Pollutant Data

The investigation covers the period from 2020 to 2024, encompassing air pollutant concentrations and meteorological data, including driver fields.

Hourly air pollutant data were collected from eight cities via CNEMC. Missing data points were imputed following a hierarchical strategy: gaps of 1–3 h were linearly interpolated, 4–6 h gaps were filled using a centered 5 h moving average, and 7–24 h gaps were replaced by the mean of the same hour from the preceding and following three years. The dataset was rigorously cleansed before any analytical procedures were performed. Subsequently, the daily mean concentrations of PM_2.5, PM₁₀, SO₂, NO₂, and CO were calculated. For O₃, the maximum daily 8 h average (MDA8) was derived from the processed hourly data. Considering the underestimation and uncertainties in emission inventories over the Tibetan Plateau, ground-based pollutant concentrations were used as proxies for local anthropogenic emissions. This validated approach helps improve the reliability of source attribution [28,29,30,31,32]. Meteorological data were obtained from the ERA5 reanalysis dataset (ECMWF). Hourly ERA5 data were matched to each monitoring site using the nearest-neighbor method. Daily variables-relative humidity (rh), boundary layer height (blh), surface solar radiation (ssrd), 2 m air temperature (t₂m), total precipitation (tp), and wind speed (wind)-were calculated by averaging hourly values (for t2m, rh, blh, wind) and summing accumulative ones (for tp, ssrd). This ensured temporal and spatial consistency with the pollutant data. These ERA5 variables were used only as predictors in the MLR model. Although potential site-level biases may exist, previous studies have shown that ERA5 provides reliable temporal consistency over the Tibetan Plateau [33,34,35]. Meteorological fields for trajectory simulations were derived from the GDAS1 reanalysis dataset released by NCEP. For each monitoring site, 72 h backward trajectories were simulated using the HYSPLIT model at multiple altitudes (500 m, 1000 m, 1500 m) and four commencement times per day (00, 06, 12, and 18 UTC). To maintain temporal consistency, trajectories were aligned by start time and matched with the daily pollutant values of the same day. The trajectories thus generated formed the foundation for subsequent analyses.

Table 1. Main data sources and variables used in this study.

Data Type	Variable Name	Resolution	Data Source
Air quality	PM2.5, PM10, SO2, NO2, CO, O3	Hourly/Daily (O3: MDA8)	CNEMC http://www.cnemc.cn/
Meteorology	rh, blh, ssrd, t2m, tp, wind	Hourly/Daily	ERA5 https://cds.climate.copernicus.eu/
Meteorological driver	GDAS1 fields	6-hourly	GDAS1 (NCEP) https://www.ready.noaa.gov/gdas1.php (accessed on 1 July 2025)

presents a comprehensive summary of all datasets, variables, and their sources used in this study.

2.2. Integrated Methodology: The MRG-HSW Attribution Framework

To systematically elucidate the local drivers and external influences on PM_2.5 and O₃ concentrations over the Tibetan Plateau, this study introduces the MRG–HSW integrated attribution framework, which comprises two primary components—the local attribution module (MRG) and the external transport diagnostic module (HSW). A schematic overview is provided in Figure 2, while detailed descriptions are presented in Section 2.2.1 and Section 2.2.2.

2.2.1. Local Attribution Block: MRG

M (Multiple linear regression, MLR): The MLR model is favored for its simplicity and interpretability and is extensively utilized in environmental research, offering clearer physical insights compared with the black-box nature of most machine learning and deep learning models [36,37]. Variations in PM_2.5 and O₃ concentrations are primarily influenced by local meteorological conditions and co-emitted pollutants [38,39,40,41,42], with potential modifications stemming from external transport. This research constructed three distinct types of models: meteorological-factor models, pollutant factor models, and combined models. These models are employed to compare the goodness of fit and statistical significance among various combinations of factors, thereby determining the relative contributions of each. The performance of these models was assessed using adjusted R² (to gauge explanatory power), t-values (to rank the influence strength of variables), and Variance Inflation Factors (VIF, all < 5) to evaluate multicollinearity.

The model represents the target pollutant concentration as a linear combination of an intercept and regression terms for meteorological and pollutant factors.

$$Y=\alpha +{\displaystyle \sum _{m\in \mathcal{M}}{\beta }_m{X}_m+{\displaystyle \sum _{p\in \mathcal{P}}{\gamma }_p{Z}_p}+\epsilon }$$

(1)

In this formulation, $Y$ denotes the observed concentration of the target pollutant (PM_2.5 or O₃); $\alpha$ is the intercept term of the regression model; $\mathcal{M}$ denotes the set of meteorological factors (rh, blh, ssrd, t2m, tp, wind); $\epsilon$ denotes the residual term, calculated as ${\epsilon }_t=Y_t-{\widehat{Y}}_t$, (${\widehat{Y}}_t$ being the model-fitted concentration). Positive residuals ($\epsilon >0$) were interpreted as potential external inputs, whereas negative ones ($\epsilon <0$) indicated local removal or model uncertainty, and $\mathcal{P}$ represents the set of local pollutant factors (SO₂, NO₂, CO, PM₁₀). In the models for PM_2.5 or O₃, the pollutant itself is excluded from the predictor set. To distinguish the relative effects of different predictor groups, three model configurations were defined as follows: Thus, three model configurations are defined:

• Meteorological factors model (Equation (1)): Excludes pollutant factors $\mathcal{P}=\varnothing$;
• Pollutant factors model (Equation (1)): Excludes meteorological factors are excluded $\mathcal{M}=\varnothing$;
• Combined factors model (Equation (1)): Includes both meteorological and pollutant factors $\mathcal{M},\mathcal{P}$;

R (Residual-based screening): We utilized the residuals from the MLR model as indicators of external transport. Specifically, days with high pollution were initially identified based on the Grade I limits of the Ambient Air Quality Standards (GB 3095-2012) [43], which are daily mean PM_2.5 concentrations ≥ 35 μg/m³ or O₃ MDA8 concentrations ≥ 100 μg/m³. For each high-pollution day, we calculated the positive residual relative to the MLR-fitted value. The 90th percentile of the residual distribution was selected as the threshold. Threshold justification (small-sample discretization): In small-n settings, quantiles map to order statistics. Since upper-tail residuals have ties, different quantiles in the range of [0.85,0.95] result in identical thresholds and same flagged-day set. Therefore, the 90th percentile represents an optimal compromise. A day was classified as a “high-pollution high-residual day” only if both the concentration exceeded the established standard and the residual surpassed the threshold. The residuals signify the portion of the observed concentrations unexplained by local factors. This “dual-threshold” approach effectively isolates extreme cases where external influences are likely predominant, offering a more robust assessment of the contributions from cross-regional transport. In addition, we quantified the external-transport intensity ($I_{ext}$) based on normalized regression residuals $(y-\widehat{y})/y$, representing the proportion of pollutant concentrations unexplained by local factors and thus potentially attributable to external inflow. Under the MRG–HSW design, this residual screening serves as a gatekeeping rule: only flagged “high-pollution–high-residual” days are routed to the HSW branch (HYSPLIT-SOM-WCWT); otherwise, attribution is concluded within the MRG block.

G (Generalized Additive Model, GAM): At sites exhibiting high goodness-of-fit, this study enhances the MLR by incorporating a GAM to capture the nonlinearities and threshold effects of key factors through smoothing functions, thereby augmenting the linear model’s explanation of local driving mechanisms. The GAM offers flexibility in representing nonlinearities and interactions among variables, and has been widely applied in studies of atmospheric pollution and meteorological impacts [44,45]. Its basic structure is similar to that of the MLR, which is not reiterated here; instead, the Results section will present the main nonlinear response characteristics of the essential factors.

Partial dependence was calculated to assess each variable’s independent effect on pollutant concentration. For each variable, other predictors were fixed at their mean values, and the model response was averaged over the variable’s range to obtain its marginal contribution.

2.2.2. External Transport Block: HSW

H (HYSPLIT model): Following the gatekeeping rule in Section 2.2.1, the HYSPLIT model executed 72 h backward trajectory simulations only on flagged high-pollution–high-residual days, using GDAS1 reanalysis data. The simulations initiated from the geographic coordinates of each monitoring site, with starting times set at 00, 06, 12, and 18 UTC. For PM_2.5, which exhibits rapid deposition and pronounced vertical distribution characteristics [46,47], trajectories were initiated at three different altitudes-500, 1000, and 1500 m. This configuration generated twelve trajectories per site per day. Conversely, for O₃, which demonstrates a lesser degree of deposition [48,49], only a single altitude of 500 m was employed, resulting in four trajectories per site per day. All trajectory data were standardized in format, encapsulating latitude, longitude, altitude, and arrival time to facilitate the production of uniform datasets for further clustering and WCWT analyses.

S (SOM clustering): To delineate distinct atmospheric transport pathways, the gathered backward trajectories underwent clustering via a self-organizing map (SOM) neural network. This technique is renowned for its efficacy in dimensionality reduction and classification, and it has been extensively utilized in the analysis of atmospheric trajectories [50,51]. The optimal number of clusters was determined through an evaluation of candidate groups (ranging from k = 2 to 10) using the Davies–Bouldin Index (DBI), with the smallest DBI indicating the most effective clustering solution. Following this, the Kruskal–Wallis H test was applied to assess the statistical significance of differences in pollutant concentrations among the identified clusters. This nonparametric method is particularly appropriate given the non-normal distribution of the data employed in this analysis.

W (Weighted Concentration Weighted Trajectory): The WCWT method was implemented to quantitatively ascertain potential source regions of pollution. In this methodology, concentrations of pollutants at the endpoints of trajectories were used as weights to compute a weighted mean concentration over a grid of 0.5° × 0.5°, thus reflecting the relative impact of different regions on the pollutant levels at the receptor site [52]. The mathematical expression for this calculation is provided below:

$$C_{ij}={\displaystyle \frac{{\displaystyle {\sum }_{m=1}^M{C}_m\cdot n_{m,ij}}}{{\displaystyle {\sum }_{m=1}^M{n}_{m,ij}}}}$$

(2)

where $C_{ij}$ represents the weighted mean concentration in grid cell $(i,j)$, $C_m$ denotes the concentration of pollutants associated with the m-th trajectory, $n_{m,ij}$ is the number of trajectory endpoints within the grid cell, and M is the total number of trajectories. To validate the robustness of these findings, a permutation significance test was conducted by randomly reshuffling concentration values, thus identifying grid cells with statistically significant contributions and evaluating the reliability of the spatial distribution patterns inferred from the data.

Similar to previous trajectory studies, no additional uncertainty tests were conducted, because the trajectory–WCWT framework already accounts for sampling variability through averaging and weighting [53,54].

All analyses were performed in Python 3.10, and backward trajectories were simulated using HYSPLIT 5.3 (NOAA ARL, College Park, MD, USA). All computations ran on Windows 10.

3. Results and Discussion

3.1. Variations of PM_2.5 and O₃

Xining was chosen to represent the overall pattern observed across Qinghai Province (Figure 3).

Overall, O₃ peaks in summer and drops in winter, showing a clear afternoon maximum around 15:00–16:00 LT, driven by strong sunlight and higher boundary layers that favor photochemical production. In contrast, PM_2.5 reaches its highest level in winter because of enhanced emissions and poor dispersion, while summer levels are the lowest. Its diurnal cycle shows morning and evening peaks, mainly due to traffic and stable nighttime conditions.

3.2. Local Attribution and Regression Model Performance

This section presents the results of MLR and GAM analyses to quantify the influence of meteorological and local emission-related factors on PM_2.5 and O₃ across Qinghai Province.

3.2.1. Linear Attribution with MLR

For each pollutant (PM_2.5 and O₃), three regression configurations were established as described in Section 2.2: a meteorological model (list (a) under Equation (1), rh, blh, ssrd, t₂m, tp, wind), a pollutant model (list (b) under Equation (1), SO₂, NO₂, CO, PM₁₀), and a combined model (list (c) under Equation (1)) including both groups. These configurations allow comparison of explanatory power and significance among different predictors.

Among all prefectures, Xining showed the best model performance with closely matched R² and adjusted R² values, indicating good robustness. It was therefore selected as the representative site, while results for other regions are summarized in

Table A1. Coefficient of determination (R²) of O₃ and PM_2.5 under different modeling scenarios in eight cities of Qinghai Province.

City	R2 (O3–Meteorology)	R2 (O3–Pollutants)	R2 (O3–All Factors)	R2 (PM2.5–Meteorology)	R2 (PM2.5–Pollutants)	R2 (PM2.5–All Factors)
Haibei	0.623	0.054	0.633	0.096	0.803	0.822
Huangna	0.529	0.293	0.561	0.19	0.786	0.805
Hainan	0.628	0.205	0.635	0.113	0.631	0.648
Guoluo	0.339	0.025	0.348	0.068	0.346	0.36
Yushu	0.36	0.27	0.42	0.394	0.741	0.773
Haixi	0.517	0.17	0.535	0.024	0.786	0.789
Haidong	0.779	0.364	0.799	0.282	0.681	0.728
Xining	0.71	0.365	0.751	0.333	0.801	0.825

.

In Xining, the observed and modeled daily concentrations of O₃ and PM_2.5 during 2020–2024 were analyzed using three MLR configurations (Figure 4). Results for the remaining cities are provided in Appendix A Table A1. Those considering only meteorological factors, only pollutant factors, and a combination of both factors. The analysis revealed that the model considering only pollutant factors accounted for 80.1% of the variability in PM_2.5 concentrations but only 36.5% for O₃. Conversely, the meteorological-factor model explained 71.0% of the variability in O₃ concentrations but a mere 33.3% for PM_2.5. When the models integrated both meteorological and pollutant factors, there was a significant enhancement in explanatory power, with R² values increasing to 82.5% for PM_2.5 and 75.1% for O₃. These results underscore the importance of considering both meteorological conditions and local emissions concurrently to significantly augment the explanatory capacity for both pollutants. However, the predominant driving factors differ markedly: O₃ is primarily influenced by meteorological conditions, whereas PM_2.5 is more strongly linked to local emissions. This pattern aligns with established atmospheric chemical processes, where O₃ formation is significantly dependent on photochemical reactions and is sensitive to factors such as temperature, radiation, and boundary layer dynamics. In contrast, the generation and accumulation of PM_2.5 are mainly driven by primary emissions and local secondary transformations [55,56,57,58,59]. Accordingly, strategies for O₃ control should focus on regional collaborative management and regulatory measures under adverse meteorological conditions to mitigate O₃ formation and accumulation. Simultaneously, PM_2.5 mitigation should prioritize comprehensive reductions in emissions from local pollution sources, with which it is closely associated.

Residual diagnostics supported the reliability of the regression models. The Jarque–Bera values were between 10² and 10⁴ (p < 0.001), showing a slight departure from normality. The Durbin–Watson values were close to 1.0, indicating mild serial correlation that is common in daily observations. These results suggest that the residual patterns are reasonable and do not affect the interpretation of the regression results.

For both O₃ and PM_2.5, linear models tended to underestimate observed concentrations during high-pollution episodes. The bias was more evident for O₃, where all model types failed to reproduce peak values. For PM_2.5, the pollutant- and combined-factor models performed slightly better, though underestimation still occurred. These results indicate that linear models cannot fully capture extreme events, implying that nonlinear processes and regional transport likely contribute to such peaks [60,61,62,63].

At the Xining site, t-statistics from the MLR models show that for O₃, air temperature had the strongest effect, followed by solar radiation and boundary layer height. Wind speed and precipitation had negative effects, while relative humidity was not significant when pollutant factors were included. For PM_2.5, PM₁₀ showed the highest t-value, with NO₂ and CO also contributing positively. SO₂ had little influence. The addition of meteorological factors strengthened the role of O₃, highlighting the strong impact of weather on PM_2.5-O₃ interactions. These nonlinear relationships justify further analysis with the GAM model (Figure 5).

The air quality monitoring sites in Qinghai Province display marked differences in elevation due to topographic and geographic variations. These elevation differences not only affect regional meteorological conditions and pollutant transport mechanisms but also contribute to the spatial heterogeneity in the goodness of fit (R²) of the multiple linear regression models for both O₃ and PM_2.5.

The monitoring sites in Qinghai Province present substantial elevation disparities, as shown in

Table 2. City names, site codes, and elevations (m) of air quality monitoring stations included in this study.

City	Haibei	Huangna	Hainan	Guoluo	Yushu	Haixi	Xining	Haidong
Site code	2671A	2672A	2673A	2674A	2675A	2676A	3055A	3129A
Elev. (m)	3072	2471	2818	3718	3689	2942	2204	2072

, with elevations ranging from 2072 m in Haidong to 2204 m in Xining, and reaching up to 3689 m in Yushu and 3718 m in Guoluo.

For both pollutants, R² values decreased with increasing altitude (Figure 6). Low-altitude sites, such as Xining and Haidong, exhibited higher R² values. In contrast, high-altitude sites, like Yushu and Guoluo, showed lower performance. This can be attributed to complex meteorological conditions at higher elevations. These include low air pressure, strong radiation, and high winds. Additionally, stronger external transport influences reduce the explanatory power of local factors [64,65]. The local-only model produced lower R² values, indicating that local emissions alone cannot fully explain the variability of these pollutants. Correlation analysis further confirmed that model performance for both O₃ and PM_2.5 declined with elevation. The relationship for O₃ was stronger (r = −0.91, p = 0.002), while for PM_2.5, a moderate negative correlation was found (r = −0.51, p = 0.20). These results reflect the influence of local emission variability.

3.2.2. Nonlinear Supplementation with GAM

To overcome the constraints of linear models in adequately capturing nonlinear and threshold effects, GAM were implemented within the framework of MLR in regions demonstrating relatively high goodness of fit (Figure 7). This was done to delineate the nonlinear dynamics influencing the concentrations of O₃ and PM_2.5.

Across all sites, GAM models improved R² by 10.8% for O₃ and 17.4% for PM_2.5 compared with MLR, showing stronger explanatory power. The improvement was about 5–15% in Yushu and Haixi and over 30% for PM_2.5 in Guoluo. However, O₃ models in high-altitude areas such as Yushu and Guoluo still performed poorly, suggesting that local factors alone cannot explain pollution there and that external transport likely plays a greater role. Based on these results, partial dependence plots were created for key variables at Xining to explore their nonlinear effects.

Air temperature (t2m), solar radiation (ssrd), and boundary layer height (blh) were identified as key meteorological factors influencing O₃. GAM analysis (Figure 8) shows that O₃ increases steadily with temperature, with a sharp rise above about 10 °C, confirming the strong temperature sensitivity of photochemical formation, consistent with findings in North China and the Yangtze River Delta [66,67]. Solar radiation and boundary layer height also show positive effects, enhancing O₃ production and vertical mixing. In contrast, precipitation and wind speed exhibit threshold effects: when they exceed certain levels, O₃ decreases markedly due to dilution and scavenging [68,69]. The nonlinear effect of relative humidity is weak, indicating limited independent influence and stronger interactions with other factors.

Figure 9 shows the partial dependence of PM_2.5 on major pollutant factors. PM₁₀ has the strongest positive effect, increasing steadily across all concentrations and sharply at higher levels, suggesting common sources such as dust and coal combustion, consistent with previous studies [70]. CO also shows a clear positive relationship, with PM_2.5 rising rapidly when CO exceeds about 1.5 mg/m³, reflecting the combined influence of traffic and combustion emissions [71]. NO₂ has little effect at low concentrations but increases PM_2.5 sharply above about 45 µg/m³, indicating its role in secondary particle formation [72,73]. In contrast, SO₂ shows only weak variation, implying a minor contribution to PM_2.5 in this region.

Compared with MLR, the GAM model better captures the nonlinear responses of key factors, reducing the limitations of linear methods. However, in remote plateau regions such as Guoluo and Yushu, its explanatory power remains low, even with GAM, indicating that local factors alone cannot explain pollutant variations. This underscores the importance of external transport in high-altitude pollution episodes and provides a basis for subsequent trajectory analysis. Therefore, at lower altitudes, where the explanatory power of local factors is higher, controlling PM_2.5 should emphasize emission reductions from industrial and transportation sources, whereas O₃ management should strengthen regional coordination and meteorological forecasting.

3.3. External Transport Analysis

Building on the aforementioned premises, this section employs the HYSPLIT model to perform 72 h backward trajectory simulations for high-pollution and high-residual days. The simulations are combined with clustering analysis to identify key transport pathways and potential source regions. Quantitatively, the calculated external-transport intensity ($I_{ext}$) values ranged from 0.16–0.31 in low-altitude regions and 0.27–0.44 in high-altitude regions (Guoluo, Yushu). This indicates that external transport contributes approximately 20–40% of the observed pollution, with a particularly strong influence at higher elevations where local explanatory power is weakest.

3.3.1. Trajectory Simulation, Clustering, and Source Pathway Analysis for Low-R² Sites/Days

A total of 92, 22, and 165 high-pollution and high-residual days were identified for O₃ (Guoluo), PM_2.5 (Guoluo), and O₃ (Yushu), respectively, generating 1028 trajectories for O₃ and 264 for PM_2.5. The optimal cluster numbers, determined by the Davies–Bouldin Index, were k = 2 for O₃ in Guoluo, k = 3 for O₃ in Yushu, and k = 2–3 for PM_2.5 at different altitudes. These trajectories were then grouped using SOM clustering, and representative transport pathways with their relative proportions were extracted for subsequent analysis (Figure 10 and Figure 11).

In Guoluo, air masses mainly arrived from the southwest and west, indicating both local inflow and long-range transport across regions. In Yushu, transport was dominated by short-range circulation from the southwest, with occasional inflow from the northwest and west, reflecting combined local and regional influences.

In Guoluo, PM_2.5 transport showed multiple pathways that varied with altitude. At lower levels (500 m), air masses mainly arrived from the southwest and west, reflecting local and near-regional influences. At 1000 m and above, trajectories increasingly originated from the west and northwest, indicating enhanced long-range transport. Overall, higher altitudes were dominated by cross-regional inputs, whereas local sources played a greater role near the surface.

To assess the validity of the clustering results, the Kruskal–Wallis H test was employed to compare pollutant concentrations across different trajectory categories. The findings indicated statistically significant differences in concentrations among categories (p < 0.001), confirming that the clustering technique effectively differentiated between transport pathways associated with varying levels of pollution.

3.3.2. Spatial Identification of Potential Source Regions by WCWT

Building on the trajectory findings from high-pollution, high-residual days discussed in Section 3.2.1, WCWT calculations were performed for each trajectory category to delineate the spatial distribution of potential source regions and their respective contributions to different transport pathways. The results, detailed in Figure 12, Figure 13, Figure 14, Figure 15 and Figure 16, illustrate the distribution patterns for each category.

It should be noted that in certain regions or pollutants, the high-value areas identified by WCWT may not exactly coincide with the dominant transport directions obtained from SOM clustering. This discrepancy arises from their methodological focus: SOM emphasizes airflow pathways, while WCWT identifies potential source regions where pollutants may accumulate along the trajectories. The two approaches are therefore complementary in interpreting transport dynamics.

At 500 m in Guoluo, O₃ transport was mainly driven by local and nearby regional sources from the southwest and west. High-contribution areas within and north of Guoluo reached over 200 μg/m³, while concentrations decreased eastward and southeastward with increasing transport distance. Cluster A showed a maximum WCWT value of 226.88 μg/m³ and an average of 145.33 μg/m³, indicating strong local and medium-range influence. In contrast, Cluster B exhibited lower overall contributions, with most areas below 140 μg/m³ and moderate values of 160–200 μg/m³ along the western and southern edges of the Plateau. The maximum and mean values were 196.25 μg/m³ and 142.77 μg/m³, suggesting long-range transport with progressive dilution of pollutants.

At 500 m in Yushu, WCWT analysis of O₃ trajectories during high-pollution, high-residual days shows distinct spatial patterns (Figure 13). Cluster A exhibits a high-contribution zone across southern and southeastern Yushu, where concentrations exceed 200 μg/m³, with a maximum of 168.08 μg/m³ and an average of 136.41 μg/m³. This distribution, aligned northwest–southeast along the southern Tibetan Plateau, indicates combined effects of nearby sources and transboundary transport. Cluster B shows generally low contributions, mostly below 120 μg/m³, with a few moderate areas (120–160 μg/m³) along the southern and western plateau margins. The maximum concentration is 160.0 μg/m³, suggesting long-range transport with pollutant dilution. Cluster C presents higher contributions to the east of Yushu and southeastern Qinghai, expanding westward, with a maximum of 189.88 μg/m³ and an average of 138.16 μg/m³. This pattern reflects the influence of both short- and medium-range transport on O₃ accumulation.

A comparison of the WCWT results for O₃ between Guoluo and Yushu shows clear differences in the location and strength of high-contribution areas. In Guoluo, high values are mainly concentrated in the southwest, reflecting both local and medium-range transport. This zone extends eastward into eastern Qinghai and the western Sichuan Plateau, where contributions quickly weaken. In contrast, Yushu’s high-value areas are linked to the southern border and the Qinghai–Tibet boundary, showing stronger and more continuous contributions across short- and medium-range transport. For Cluster B, Guoluo shows low and scattered values, suggesting stronger dilution during long-range transport, while Yushu keeps moderate contributions over southwestern Sichuan, indicating more persistent long-range effects. In Cluster C, Yushu forms a continuous mid-level belt stretching eastward, which is absent in Guoluo. These contrasts highlight the different source regions and transport pathways shaped by local geography and weather.

At 500 m in Guoluo, PM_2.5 transport was mainly influenced by local and nearby regional sources. A pronounced high-contribution belt extended southeastward across Guoluo, with core values above 75 μg/m³ and peak areas exceeding 115 μg/m³. The maximum and average WCWT concentrations were 147.08 μg/m³ and 65.60 μg/m³, indicating strong local and medium-range effects. In contrast, Cluster B showed lower overall contributions, mostly below 75 μg/m³, with only a few discontinuous moderate patches. Its maximum and average values were 147.08 μg/m³ and 50.60 μg/m³, suggesting weak long-range influence.

At 1000 m, the influence range broadened and long-distance effects became more evident. Cluster A formed a continuous medium-concentration belt from Guoluo to eastern Qinghai, with localized peaks over 95 μg/m³ (max 111.83 μg/m³, avg 48.74 μg/m³). Cluster B followed a west–east corridor dominated by values below 75 μg/m³, while Cluster C extended southwestward with concentrations surpassing 135 μg/m³ (max 147.08 μg/m³, avg 61.06 μg/m³), reflecting increasing regional and cross-border transport contributions.

At 1500 m, the transport range expanded further, but contribution levels declined. Cluster A showed scattered low-value cells (mostly <55 μg/m³; max 61.08 μg/m³, avg 46.69 μg/m³), indicating weak regional transport. Cluster B formed a westerly transport belt with occasional peaks up to 135 μg/m³ (max 147.08 μg/m³, avg 62.90 μg/m³), suggesting diluted long-range inputs. Cluster C, mainly over Qinghai and adjacent regions, displayed generally low concentrations (<65 μg/m³; max 71.68 μg/m³, avg 44.78 μg/m³).

Overall, at 500 m the transport was dominated by concentrated local and medium-range sources, while at 1000 m the signals from distant sources strengthened along continuous corridors. At 1500 m, the spatial coverage widened but high-value areas became sparse. Despite differences in intensity, the main transport pathways remained consistent across heights, confirming the robustness of PM_2.5 transport patterns and the progressive dilution of pollutants with altitude.

3.3.3. Case Study of Extreme Pollution Episodes and External Transport Characteristics

To elucidate further the role of external transport in the development of extreme pollution events, this section examines a PM_2.5 extreme event in Guoluo on 28 February 2021, and an O₃ extreme event in Yushu on 1 June 2024. Backward trajectory simulations are employed to reveal the external transport characteristics associated with these pollution events.

• Guoluo PM_2.5 Pollution Episode (28 February 2021)

On 28 February 2021, Guoluo experienced a severe particulate pollution event. The PM_2.5 concentration reached 147.1 μg/m³, while PM₁₀ peaked at 468.1 μg/m³, giving a PM₁₀/PM_2.5 ratio of about 3.2. This high ratio (coarse fraction ≈ 68.6%) suggests that dust made a dominant contribution. The MLR model predicted a PM_2.5 concentration of 73.31 μg/m³, which is roughly half of the observed value. This indicates that local sources alone could not explain the pollution. The remaining ≈50% (≈73.8 μg/m³) likely came from external transport, representing about a twofold increase over the local baseline. The 72 h backward trajectories (Figure 17) showed that air masses at 500–1000 m mainly came from the Qaidam Basin and the Hexi Corridor, while those at 1500 m extended southwestward. This pattern reveals a multi-level inflow of dust. Together with the WCWT results, these findings confirm that about half of the PM_2.5 originated from long-range dust transport, while poor dispersion conditions acted as a secondary factor.

2.

Yushu O₃ Pollution Episode (1 June 2024)

On 1 June 2024, Yushu experienced a notable O₃ pollution episode, with the MDA8 O₃ concentration reaching 189.9 μg/m³. The day was marked by poor atmospheric ventilation, high relative humidity, and precipitation, while precursor levels remained low. The MLR model predicted only 90.77 μg/m³, less than half of the observed value, indicating that local photochemical processes alone could not explain the peak. The remaining ≈52% (≈99 μg/m³) can be attributed to external transport. The 72 h backward trajectories (Figure 18, at 500 m) showed that air masses over Yushu mainly originated from the southwest and west, extending toward South Asia. This pattern reveals strong external influence on local air quality. Together with the WCWT results, these findings confirm that long-range transport was the dominant driver of this O₃ episode, while unfavorable local dispersion acted as a secondary amplifying factor.

3.4. Integrated Discussion

The present study revealed a clear altitude-dependent pattern in the pollution mechanisms over the Tibetan Plateau, driven by both local and transboundary processes. This finding is consistent with previous observations of elevation-related gradients in northwestern China and the central Plateau [74,75,76]. At lower altitudes, both PM_2.5 and O₃ are mainly controlled by local factors. PM_2.5 is influenced by anthropogenic emissions and frequent winter inversions, while O₃ is affected by local meteorological conditions such as temperature, solar radiation, and boundary-layer height. These results are in line with previous studies in Xining and Lanzhou, where seasonal variations of PM_2.5 and O₃ were largely determined by local emissions and weather conditions [77,78,79]. At higher elevations, the influence of local factors becomes weaker, and stronger westerly winds promote long-range transport of dust and precursors. This agrees with trajectory and modeling studies that identified dominant westerly and South Asian inflows over the Plateau [80,81]. However, this study further shows a clearer altitude threshold above which local effects decline and external transport becomes dominant. Overall, these features distinguish the Plateau from lowland basins, where human emissions remain the main driver of pollution throughout the year, and highlight the combined effects of meteorological, topographic, and chemical processes on air pollution formation.

Building upon these findings, several limitations and future directions should be noted. While the observational period covers only five years (2020–2024), it ensures consistent data across all sites and captures multiple seasonal cycles. Similar time spans have been used in previous studies [82,83,84,85], but we acknowledge that longer datasets would better reflect interannual variability and will be considered in future work. First, the relatively short observational period and the absence of certain precursor species may have constrained the comprehensiveness of the results. Second, although MLR and GAM provided valuable insights into the relationships between local factors and pollutant concentrations, these statistical models remain inadequate for fully capturing the complex non-linear chemical processes underlying atmospheric pollution. Third, the reliability of the HYSPLIT and WCWT analyses is highly dependent on the accuracy of meteorological inputs and the sample size, which introduces uncertainties, particularly under extreme conditions [86,87,88]. In addition, the use of ERA5 reanalysis data introduces uncertainties over the complex terrain of the Tibetan Plateau. ERA5 tends to overestimate rainfall frequency and duration over the eastern margin [89] and shows a wet bias of about 20–30% in western China [90]. It also underestimates near-surface temperature by 1–2 °C in the Qilian Mountains [91], overestimates shortwave radiation by around 15 W m⁻² [92], and underestimates wind speed in high-altitude areas [93]. These biases suggest that the overall uncertainty of ERA5 meteorological forcing over the Plateau is roughly within ±20–40%, which may slightly affect this study’s results. Future work will use higher-resolution or bias-corrected datasets to improve model reliability. Finally, the current analysis qualitatively assessed the balance between local influences and external transport without explicitly quantifying their respective contributions. Future research should aim to refine this framework by incorporating regression coefficient or residual decomposition methods to achieve more robust quantification, and by coupling with three-dimensional chemical transport models (e.g., WRF-Chem, GEOS-Chem) to simulate chemical mechanisms with greater fidelity [94,95,96]. In addition, integrating high-resolution remote sensing products and multi-source datasets may further enhance the attribution of pollution dynamics on the Plateau and extend the applicability of the framework to other complex regions.

4. Conclusions

This study combined multivariate statistical modeling with trajectory analysis to build an integrated MRG–HSW framework for identifying the drivers of PM_2.5 and O₃ pollution across eight prefectures in Qinghai Province. In low-altitude areas, PM_2.5 was mainly affected by local human activities (R² = 0.63–0.80), while O₃ responded more strongly to meteorological conditions (R² = 0.53–0.78). O₃ increased with temperature, solar radiation, and boundary layer height but decreased with precipitation and wind speed. PM_2.5 was mainly influenced by PM₁₀, CO, and NO₂, with SO₂ having a minor effect. At high altitudes, the explanatory power of local factors dropped sharply (R² = 0.34–0.36), indicating that local processes alone could not explain pollution peaks. External transport thus became the main driver. Quantitative results show that the external-transport intensity ($I_{ext}$) ranged from 0.16–0.31 in low-altitude regions and 0.27–0.44 in high-altitude regions, meaning that external inflow contributed about 20–40% of the total pollution and had a stronger impact in higher areas. In Guoluo, long-range dust from the Qaidam Basin and Hexi Corridor added to PM_2.5, while O₃ was affected by both short-range inflow from the southeast and long-range transport from the west. In Yushu, O₃ was influenced mainly by air masses from the northwest and South Asia. These results confirm that the spatial differences of PM_2.5 and O₃ on the Tibetan Plateau arise from both local emission–meteorology interactions and large-scale external transport.

Collectively, this study deepens the understanding of pollution formation mechanisms on the Tibetan Plateau and provides evidence for differentiated regional management and cross-regional cooperation. In lower-altitude areas, PM_2.5 mitigation should prioritize reducing local emissions, particularly from industrial and vehicular sources, while O₃ control should focus on meteorological influences and coordinated regional actions. In higher-altitude prefectures, where external transport plays a dominant role, collaboration with neighboring provinces and transboundary partners such as South Asia is essential. These differentiated strategies contribute directly to sustainable air quality governance by aligning local mitigation efforts with regional atmospheric circulation patterns. Integrating the results into long-term planning-such as the optimization of energy and transportation structures, early warning networks, and regional emission-trading mechanisms-can help achieve both pollution control and carbon reduction co-benefits. The MRG-HSW framework provides a transferable and science-based decision-support tool for evidence-driven policymaking, facilitating adaptive and cooperative air quality management in the ecologically sensitive Plateau region.