Relationship between NO₂ and meteorological parameters
The relationship between NO₂ concentrations and meteorological variables revealed distinct nonlinear dependencies, as captured by all the four algorithms used. Figure 4 presents the predicted versus observed NO₂ concentrations using all seven meteorological parameters. Among the models, the RF and GB algorithms demonstrated superior predictive capability, while the DT and NN exhibited moderate performance. Quantitatively, the GB model achieved the lowest mean squared error (MSE = 2.07) and the highest coefficient of determination (R2 = 0.826) when three input features, day-of-year, dew point, and humidity were used (Table 2). The RF model followed closely (MSE = 2.39; R2 = 0.799). Conversely, the DT and NN models recorded relatively higher errors, reflecting their limited ability to generalize across the full range of meteorological variability. The superior performance of ensemble algorithms (RF and GB) can be attributed to their robustness in handling nonlinear, multivariate interactions typical of atmospheric datasets. Both models leverage multiple weak learners to minimize bias and variance, thus capturing subtle dependencies between meteorological predictors and pollutant concentration.
Performance comparison of four machine learning algorithms NN, DT, RF, and GB in predicting nitrogen dioxide concentrations from meteorological parameters.
From a physical standpoint, the strong dependence of NO2 on humidity and dew point highlights the influence of atmospheric moisture on pollutant chemistry. High humidity promotes heterogeneous reactions that enhance the conversion of NOx into secondary nitrogen compounds. Furthermore, seasonal variations captured by the day-of-year variable indicate temperature-driven photochemical processes, where higher solar radiation during summer enhances photolysis of NO2 to form ozone. Similar findings were reported by8] and [18, who observed comparable seasonal and humidity-dependent trends. Interestingly, including additional meteorological features beyond the top three did not significantly improve model accuracy, in fact MSE increased marginally. This suggests model saturation, where redundant or weakly correlated features introduce noise rather than additional explanatory power24. Therefore, optimal model complexity in environmental AI applications depends not only on algorithmic depth but also on judicious feature selection.
Relationship between CO and meteorological parameters
The relationship between CO concentration and meteorological variables exhibited a more diffuse predictive pattern than that of NO2 (Fig. 5). Table 3 summarizes the statistical indicators derived from each model. The Gradient Boosting (GB) algorithm consistently yielded the most reliable results (MSE = 136.47; R2 = 0.456), outperforming NN, DT, and RF models. While overall explanatory strength was moderate (R2 < 0.5 for most models), distinct meteorological influences were evident. The day-of-year, dew point, and weather condition variables emerged as the most relevant predictors. The temporal dependence indicates that CO levels vary seasonally, reflecting differences in combustion activity, temperature-driven atmospheric mixing, and photochemical oxidation rates. During cooler months, reduced atmospheric turbulence and lower boundary layer height enhance CO accumulation near the surface. Conversely, higher temperatures in summer accelerate oxidation to CO₂, reducing ambient CO concentrations.
Comparative performance of four machine learning algorithms NN, DT, RF, and GB in predicting carbon monoxide (CO) concentrations from meteorological inputs.
The inclusion of dew point and weather condition variables generally enhanced model performance, particularly for decision tree and ensemble-based models, implying that atmospheric moisture and dispersion state significantly affect CO distribution. Higher dew points often correspond to stagnant air and limited vertical mixing, both of which promote CO persistence. Similar trends have been reported in previous studies conducted17, where low wind speeds and temperature inversions contributed to increased CO accumulation. The moderate predictive strength compared with NO2 suggests that CO levels are less sensitive to short-term meteorological fluctuations and more influenced by emission dynamics, particularly vehicular and industrial combustion sources. AI-based models can thus complement emission inventories by distinguishing meteorological effects from anthropogenic contributions. However, the relatively lower R2 values indicate that improving CO prediction accuracy would require incorporating traffic density, emission rates, and atmospheric boundary layer height parameters that were not included in the present dataset.
Relationship between PM10 and meteorological parameters
The relationship between PM10 and meteorological factors was generally weak across all algorithms (Fig. 6). As summarized in Table 4, none of the models achieved satisfactory performance, with R2 values ranging between − 0.3 and 0.29 and high MSE values (> 250). This poor model performance indicates that PM10 concentrations in the study area are governed by highly variable and non-stationary factors, many of which are not fully captured by basic meteorological parameters.
Comparison of modeled and observed PM₁₀ concentrations obtained from NN, DT, RF, and GB algorithms.
The three best-performing predictors wind speed, day-of-year, and humidity provided marginal improvements in predictive capability when combined. The NN and RF algorithms performed slightly better than DT and GB, though still with limited accuracy. The negative R2 values observed for several models suggest that linear and ensemble-based approaches struggled to generalize PM10 patterns in the dataset25. Several factors may explain these outcomes. PM10 levels in arid regions such as Dammam are strongly influenced by non-linear dust resuspension events, which depend on threshold wind speeds, soil moisture, and surface roughness. While moderate winds facilitate dispersion, strong gusts (> 6 m/s) can mobilize large quantities of mineral dust, temporarily elevating PM10 concentrations independent of humidity or temperature. Also, sporadic local activities such as construction, road traffic, and industrial processes can introduce abrupt emission spikes that disrupt model continuity. These stochastic influences are difficult to capture with purely meteorology-based models. The weak model performance therefore highlights a limitation of AI correlation analyses when applied to heterogeneous particulate datasets. Similar findings were reported by18, who noted that PM levels in coastal and desert regions exhibit poor predictability based solely on meteorological inputs. Incorporating aerosol optical depth (AOD), soil particle index, and surface dust emission data could substantially enhance future models.
The observed partial correlations suggest certain meteorological controls: higher relative humidity enhances particle growth through hygroscopic processes, increasing measured PM10 mass. Moderate winds promote dispersion, whereas strong winds increase dust entrainment. The day-of-year variable reflects recurring patterns of dust storm frequency and intensity, which typically peak during late spring and summer. These results emphasize the multifactorial nature of PM10 dynamics and the need for hybrid models that combine meteorological predictors with emission and land-surface data for improved reliability. The low and, in some cases, negative R2 values observed for PM10 prediction indicate that models driven solely by standard meteorological variables are unable to adequately represent the dominant controls on particulate matter variability in arid environments. Unlike gaseous pollutants, PM10 concentrations in desert regions are strongly influenced by episodic and non-linear dust resuspension events, which depend on surface conditions, soil moisture, land cover, and threshold wind speeds rather than on gradual meteorological variations alone. Negative R2 values therefore reflect the inability of purely meteorology-based models to generalize PM10 behavior, rather than numerical instability or algorithmic inadequacy.
Comparative assessment of algorithmic performance
The comparative evaluation of the four algorithms revealed clear differences in their predictive accuracy, robustness, and capacity to capture nonlinear dependencies between meteorological parameters and air pollutant concentrations. These differences reflect both the intrinsic properties of the algorithms and the nature of the atmospheric data analyzed. Overall, ensemble-based algorithms (RF and GB) consistently outperformed individual learners (DT and NN) across all pollutant categories. The GB model achieved the highest accuracy, particularly for NO2 and CO prediction, demonstrating its superior capability to model subtle nonlinear interactions and compensate for data heterogeneity. The model’s iterative learning process where successive trees correct the residual errors of previous ones enables it to systematically minimize bias and variance, thereby achieving better generalization in small to medium-sized datasets. This is particularly advantageous for environmental data, where nonlinear meteorological-pollutant interactions are common and measurement noise is often unavoidable.
The RF model also performed reliably, ranking second overall. Its bagging approach, which constructs numerous decision trees trained on random subsets of the data and predictor features, provides high stability against overfitting and robust handling of multicollinearity among meteorological variables. RF’s capability to compute feature importance further enhances interpretability, allowing identification of the dominant meteorological drivers for each pollutant. For instance, humidity and dew point consistently ranked among the top predictors for NO₂, while temperature and wind speed were most influential for CO and PM10. By contrast, the DT algorithm showed comparatively lower accuracy. Although its hierarchical structure facilitates intuitive interpretation, its performance deteriorated in the presence of continuous, noisy, and interdependent environmental variables. The single-tree framework is prone to high variance and overfitting, leading to inconsistent predictions when applied to heterogeneous meteorological datasets. The NN model also underperformed relative to ensemble methods. This can be attributed to the limited sample size available which constrains the NN ability to learn complex mappings without overfitting. Neural networks generally require larger and more diverse datasets to effectively approximate nonlinear functions, particularly when multiple hidden layers and activation functions are involved. Nevertheless, the NN model showed potential when key features such as humidity, dew point, and wind speed were emphasized suggesting that further optimization with larger datasets and additional regularization techniques could enhance its predictive strength.
A clear trend across all models was the diminishing return of adding more meteorological parameters. As the number of predictors increased beyond three to four, the models exhibited higher MSE and lower R2 values. This effect reflects the classical curse of dimensionality in machine learning, where redundant or weakly correlated features introduce noise and reduce overall generalization. Therefore, feature selection and dimensionality reduction remain critical preprocessing steps in environmental modeling. Retaining only meteorological parameters with strong physical and statistical relationships to pollutant dynamics such as humidity, wind speed, and temperature enhances both accuracy and computational efficiency. From a theoretical perspective, the superior performance of GB and RF supports growing evidence that ensemble learning methods are better suited for environmental systems characterized by nonlinearity, stochasticity, and feedback loops26. These models provide a balance between flexibility and interpretability, making them ideal for translating complex atmospheric processes into quantitative predictive frameworks.
Environmental and modeling implications
The results of this study have significant implications for both environmental management and the advancement of data-driven modeling in atmospheric sciences. The differential response of pollutants to meteorological factors shows the need for pollutant-specific predictive frameworks, rather than a single universal air quality model. From an environmental standpoint, the findings reveal that NO2 concentrations are primarily governed by atmospheric moisture content and seasonal variation. The strong predictive relationship with humidity and dew point suggests that local meteorological conditions substantially influence nitrogen oxide chemistry, particularly through reactions involving aqueous aerosols and heterogeneous catalysis. This implies that even in arid regions such as the Eastern Province of Saudi Arabia, transient increases in atmospheric humidity associated with coastal winds or early morning dew formation can intensify secondary pollutant formation. Consequently, integrating high-resolution meteorological monitoring with real-time air quality forecasting could enable early warnings of ozone and NO2 pollution episodes, which are of major health concern.
The moderate relationship of CO with temperature and atmospheric mixing conditions indicates that pollutant persistence is strongly tied to boundary layer dynamics. The relatively low R2 values suggest that meteorological parameters alone cannot fully explain CO variability, emphasizing the dominant role of anthropogenic emissions. Incorporating real-time traffic data and emission inventories would substantially enhance model performance. Nevertheless, the current results demonstrate that ensemble-based AI models can successfully disentangle the meteorological contribution from background emission signals, enabling more accurate source attribution analyses. The weak correlations observed for PM10 emphasize the complexity of dust-dominated environments. In arid regions, PM10 variability arises from both meteorological forcing (e.g., wind gusts, humidity, and surface dryness) and episodic dust entrainment from natural and anthropogenic activities. The stochastic nature of these processes explains the poor predictive performance of models trained exclusively on standard meteorological inputs. To improve PM₁₀ predictability, future studies should adopt more comprehensive modeling frameworks that integrate meteorological predictors with additional explanatory variables. These may include land-surface parameters (e.g., soil moisture, vegetation cover, and surface roughness), emission-related factors (e.g., construction activity, traffic density, and industrial sources), and satellite-derived products such as aerosol optical depth (AOD) and dust indices. Furthermore, hybrid modeling approaches that combine physical dust emission and transport models with machine learning techniques may offer a more robust representation of PM₁₀ dynamics. Such hybrid frameworks can leverage physical process understanding while retaining the flexibility of data-driven learning, thereby improving predictive reliability in dust-dominated arid regions.
Beyond pollutant-specific observations, the study provides broader methodological insights into ML applications for atmospheric research. The superior performance of the GB algorithm indicates that hybrid models combining ensemble learning with feature selection techniques can effectively balance predictive accuracy and interpretability15. The implications of this study extend to policy and operational forecasting. The demonstrated ability of ML to predict air quality parameters based on routine meteorological measurements provides a cost-effective alternative for regions with limited sensor networks. Implementing such models in regional air quality management systems could support Short-term pollution forecasting to inform public health advisories, the emission control strategies, especially for transport and industrial sources as well as climate adaptation planning, by linking pollutant dynamics with extreme meteorological events (e.g., dust storms, temperature inversions).
