Predicting Ozone Concentrations in Ecologically Sensitive Coastal Zones Through Structure Mining and Machine Learning: A Case Study of Chongming Island, China

Abstract

Elevated O₃ concentrations pose a significant threat to human health and ecosystems, but little research has been performed on coastal wetlands near large cities. This study focuses on investigating the key factors affecting O₃ formation in the ecologically sensitive Dongtan Wetland (Chongming District, Shanghai, China) area. By comparing the performance of O₃ concentration prediction of multiple machine learning models, this study found that the random forest model achieved the highest accuracy (R² = 0.9, RMSE = 11.5). Feature importance and structure mining showed that peroxyacetyl nitrate (PAN), nitrogen oxides (NOx), temperature, wind direction, and relative humidity were the main drivers of O₃ formation. Specifically, PAN concentrations exceeding 0.1 ppb and temperatures above 3 °C were found to have a significant impact on O₃ levels, especially in spring, summer, and autumn. Trajectory analysis showed that westward urban pollution and emissions transported from the ocean were the main factors in O₃ formation in the area. This study highlights the need for targeted emission control strategies, especially for PAN precursors generated by ships and NOx generated by urban industries, providing important insights for improving air quality in ecologically sensitive coastal areas.

1. Introduction

Ozone (O₃) pollution has become an environmental issues of global concern [1]. O₃ is not only a greenhouse gas but also has a significant impact on surface air quality and human health [2]. Since 2013, China’s fine particulate matter (PM_2.5) concentration has dropped by 30–40%, but since the implementation of the “Air Pollution Prevention and Control Action Plan” in 2013, the ambient O₃ concentration in many Chinese cities has remained high [3]. The 90th percentile of the maximum 8 h average daily O₃ in Shanghai in 2023 is 158 μg/m³, which has not changed much from 163 μg/m³ in 2013 [4]. Therefore, O₃ pollution has become a growing concern in Shanghai and across China. Although the fundamental mechanisms of O₃ formation are well understood, the region-specific drivers of O₃ generation need further study to develop targeted control strategies.

The formation of O₃ involves many drivers, including precursor substances, chemical processes, and meteorological conditions [5]. Nitrogen oxides (NOx) and volatile organic compounds (VOCs) are the main precursors of O₃. They are generated through complex chemical reactions under the action of sunlight [6,7]. Meteorological conditions such as temperature, humidity, wind speed, and solar radiation significantly affect the generation and distribution of O₃. High temperatures and strong ultraviolet radiation usually promote the formation of O₃. Drought-induced decreases in relative humidity can trigger stomatal closure in forest ecosystems, reducing stomatal O₃ flux and weakening the ecosystem’s O₃ removal capacity, which may contribute to elevated ambient O₃ concentrations [8]. The generation mechanism of O₃ may vary in different regions, depending on local pollutant emissions, meteorological conditions, and geographical characteristics [9]. Secondary organic carbon (SOC) and primary organic carbon (POC) in the atmosphere also influence O₃ concentrations. In the machine learning-based driver analysis, SOC contributed approximately 8% to O₃ prediction, while POC accounted for about 4%, indicating their non-negligible roles in O₃ formation [10]. In addition, PAN, as an important secondary pollutant, plays a key role in the formation of O₃ [11]. PAN is primarily generated by the reaction of peroxyacetyl radicals with nitrogen dioxide, which are formed by the oxidation of various VOCs and oxygenated VOCs (OVOCs) [12]. PAN can act as a reservoir for NOx, affecting the generation of long-range O₃ [13]. PAN is known to be 1–2 orders of magnitude more biologically toxic than O₃ [14]. Numerous studies have found that, compared to urban areas, high concentrations of PAN are primarily observed in suburban regions [11,15,16]. Given its known biological toxicity to humans, it is crucial to incorporate PAN into the analysis of O₃ formation mechanisms in these areas. Additionally, attention should be paid to the interaction between PAN and O₃, as well as their combined impact on air quality and human health.

Conventional approaches to studying O₃ pollution rely primarily on deterministic physical and chemical models, which require extensive computing resources and precise input data and often have difficulty capturing complex nonlinear relationships [17,18,19]. Machine learning technology has significant advantages in handling large-scale data and complex nonlinear relationships [20]. With the development of machine learning technology, interpretable machine learning models have gradually gained attention. These models can not only provide high-precision predictions but also reveal the logic behind model decisions and feature importance, which helps to understand the O₃ generation mechanism. In order to comprehensively analyze the impact of different factors on O₃ concentration, in recent years, various ML methods, such as random forest (RF) [21], recurrent neural networks (RNNs) [22], and artificial neural networks (ANNs) [23], have been applied to predict the concentrations of atmospheric pollutants (O₃, PM_2.5, NOx, etc.) and analyze the causes of atmospheric pollution [24,25,26]. Although machine learning methods have obvious advantages in processing large-scale data and complex nonlinear relationships, they also have certain limitations [27]. Many machine learning models (such as RF and neural networks) are black box models, and it is difficult to directly explain the mechanism behind the prediction results [28]. This “black box” feature makes it difficult for the model to provide an interpretable basis for scientific research and policy making, and understanding the decision logic of the model is crucial in the field of environmental science. In addition, machine learning models are highly dependent on the quality and quantity of input data, and the model may be sensitive to data noise and outliers, which affects the stability and reliability of predictions [29]. Furthermore, these models usually require a lot of hyperparameter tuning, which not only increases the computational complexity but also easily leads to overfitting problems [30]. Therefore, although these methods perform well in the short term, they still have certain shortcomings in terms of model interpretability and robustness. To address this problem, interpretable machine learning models have gradually attracted attention in recent years, especially the application of interpretation methods such as SHapley Additive exPlanations (SHAP) and partial dependency plots (PDP). These tools can not only provide high-precision predictions but also effectively reveal the contribution of features within the model and their relationship with the target variable. SHAP assigns a contribution value to each feature to the prediction result, so that the relative importance and specific impact of the feature can be quantified and intuitively displayed [31]. PDP helps researchers understand the decision-making patterns of the model by showing the univariate or bivariate relationship between specific features and predicted results [32]. The combination of these two interpretation methods enables us to analyze the mechanism of O₃ generation more transparently and systematically and helps to identify key driving factors. This not only improves the credibility of the model but also provides a more reliable scientific basis for the formulation of air pollution control policies.

This study utilized a variety of machine learning models to predict O₃ concentrations, including recurrent neural networks (RNNs), support vector machines (SVMs), extreme gradient boosting (XGBoost), random forests (RFs), and artificial neural networks (ANNs). These models not only exhibited high predictive performance but also provided in-depth insights into the relationship between input features and O₃ concentrations. Compared with existing studies on air pollution control, while machine learning has been widely applied, the use of structure mining as a method remains relatively rare. This study compared the performance of different machine learning models (see Supplementary Materials for details) and ultimately selected the random forest model for further feature analysis due to its superior performance. The optimal parameters for the random forest model were determined using grid search and ten-fold cross-validation (see Figure S1), ensuring robust and accurate predictions. The selected model revealed the effects of multiple driving factors, such as meteorological conditions, PAN, NOx, and VOCs concentrations, on O₃ formation. In addition, traditional methods were combined to calculate the sources of airflow, providing a more comprehensive scientific basis for the prevention and control of O₃ and its precursors. The machine learning results identified the most critical driving factors, while structure mining analysis explored the interactions between paired driving factors, offering valuable references for the prevention and control of O₃ in the Dongtan area of Chongming, Shanghai.

2. Methods and Materials

2.1. Sampling and Monitoring

The data were collected from 1 January 2021 to 31 December 2021 at the Dongtan Supersite in Chongming District, Shanghai (31.51° N, 121.96° E). This is a typical island area, close to a megacity, near the ocean, with a relatively small population density. It belongs to the offshore environmental ecological zone close to the city, as shown in Figure 1. Multiple instruments were used to measure the hourly concentrations of PAN, NOx, O₃, PM_2.5, and 56 types of VOCs. Meteorological parameters include temperature, relative humidity [33], wind speed [34], and wind direction. The time resolution was 1 h.

2.2. Random Forest Model

Random forest (RF) is an integrated supervised learning method that can be seen as an extension of the decision tree [35]. In this study, the RF prediction model was built using the “sklearn” package in Python (version 3.12). Prior to model training, we used Python’s Pandas (in combination with sklearn’s preprocessing tools) to identify and remove blank values and outliers, ensuring high-quality data for modeling. Specifically, blank values were removed using the ‘dropna()’ function, and outliers were detected and removed based on the Z-score method—data points with an absolute Z-score greater than 3 were considered outliers and excluded from the analysis. To obtain the optimal parameters for the model, GridSearch was employed to identify the best combination of parameters for the RF model. Ultimately, the optimal configuration consisted of 200 decision trees, with a maximum tree depth of 10. The dataset was split into 80% for training and 20% for testing. The model’s performance was evaluated by calculating the R² (coefficient of determination) and RMSE (root mean square error) between the predicted and observed values. The model inputs included the concentrations of ‘Isoprene’, ‘Ethane’, ‘PAN’, ‘m/p-Xylene’, ‘Toluene’, ‘Propylene’, ‘o-Xylene’, ‘1-Butene’, ‘Ethylbenzene’, ‘n-Pentane’, ‘n-Butane’, ‘Propane’, ‘POC’, ‘SOC’, ‘PM_2.5’, ‘temperature’, ‘relative humidity’, ‘wind speed, ‘wind direction’, ‘NOx’, and ‘PSA’. The output was the concentration of O₃. The top ten VOCs were selected based on their O₃ formation potential (OFP) ranking, which was calculated using the maximum incremental reactivity (MIR) method from a total of 56 VOCs. For a detailed explanation of the calculation method, please refer to the Supplementary Materials, Note S2, Table S2. The RF model provides two tools to estimate the importance of individual features: the mean squared error (MSE) increase and node purity increase. To enhance the robustness of feature importance, both metrics were converted into percentages to determine the final variable importance. The training dataset for the RF model included data on O₃, PM_2.5, NOx, SOC, POC, PSA, PAN, and VOCs from 1 January 2021 to 31 December 2021. Importantly, feature calculations were not performed in isolation but were fully integrated across all records in the observed dataset, ensuring the model reflects real-world meteorological diversity. A total of 6924 valid samples were obtained during this period, of which 5539 were used for training and 1385 for testing.

2.3. Feature Importance

The importance of features is determined by evaluating their impact on the model’s predictive performance using mean squared error (MSE) increment. The MSE increment method evaluates feature importance by calculating the increase in the model’s MSE after the features are randomly shuffled or removed. The greater the increase, the more important the feature [36].

SHAP (SHapley Additive exPlanations) is an explanation method based on game theory. It uses a theoretical method based on Shapley values to assign a specific estimated importance value to each variable and determines the individual and interactive effects of each variable on O₃ formation from the optimal model. It treats each combination of features as a combination [37]. The SHAP value of a feature is calculated by averaging its marginal contribution across all alliances. The SHAP formula is shown below [38]:

$$\begin{array}{c}{y}_i=y_{base}+f\left(x_i,1\right)+f\left(x_i,2\right)+\cdots ++f\left(x_i,k\right)\end{array}$$

(1)

$y_i$ denotes the prediction, $y_{base}$ is the base value (typically the average prediction over the dataset), and $f\left(x_i,k\right)$ is the base value (typically the average prediction over the dataset) and represents the SHAP value for feature ($k$), quantifying its contribution to the prediction.

2.4. RF-PDP Method

RF-PDP (random forest partial dependence plot) is a method that combines the RF model and a partial dependence plot (PDP) to analyze and visualize the impact of features on model predictions [39,40]. First, an RF model is trained to make predictions and evaluate the importance of features. Next, partial dependencies are calculated, that is, the value of a feature is changed while other features are fixed, and the average predicted value of the model is calculated [41]. Finally, these calculation results are visualized as partial dependency plots to show how the model’s predicted value changes when the feature value changes. The RF-PDP method provides a deep understanding of the relationship between features and predicted results by combining the nonlinear modeling capabilities of RF with the intuitive visualization of PDP, revealing the overall impact of features on predicted results.

The 3D partial dependence plot (3D-PDP) is a visualization tool used to explain the relationship between features and prediction targets in machine learning models. In research, 3D-PDP can help this study understand the impact of two features changing at the same time on the target variable [42]. In this study, this study used the 3D partial dependence plot (3D-PDP) to explore how changes in two features jointly affect O₃ concentration. 3D-PDP enables this study to intuitively observe the response of the model to different feature combinations by displaying the relationship between features and predicted values in three-dimensional space. The 3D-PDP can reveal the interaction effect between two features and can also reveal the nonlinear relationship between features and target variables. The specific formula is as follows [10]:

$$\begin{array}{c}f\left(X_S\right)={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}RF\left(X_s,X_C^{\left(i\right)}\right)\end{array}$$

(2)

Among them, $f\left(X_S\right)$ means when the input is $X_S$, The predicted value of the RF model; RF represents the trained RF model; $X_S$ is the feature selected when calculating the partial dependence function in the RF model; $X_C^{\left(i\right)}$ represents the features not selected in the RF model and is input into the model with a fixed value; and n represents the number of samples.

2.5. Structure Mining Method

Structural mining methods, specifically the minimum depth (min-depth) method, are a technique used to analyze tree-based models such as RF [43]. The method aims to quantify the importance of features by examining their structural position in the ensemble of decision trees that make up the model. It provides a measure of how early a particular feature is used to split the data within each tree. In the context of decision tree algorithms such as those used in RF, the minimum depth of a feature refers to the layer or depth at which the feature is first used as a decision criterion in the tree structure. More formally, the minimum depth of a feature ff in a decision tree is defined as the distance from the root node (depth = 0) to the first occurrence of ff in any node in the tree. This metric provides insight into the relative importance of features: features that are closer to the root (i.e., have a smaller minimum depth) are generally more influential in predicting the target variable. The main rationale behind minimum depth is that important features tend to be selected earlier in the tree building process because they provide the most information gain or impurity reduction at higher levels of the tree. Therefore, features with lower minimum depth values are considered more important because they contribute more to the predictive performance of the model.

The minimum depth method provides an interpretable and rigorous way to understand feature importance in RF models. By focusing on the depth of the first selected features, it provides a direct and intuitive measure of feature importance that complements traditional importance metrics.

2.6. Trajectory Concentration Clustering Calculation

The backward trajectory clustering method is used to study the potential transport pathways of pollutants. Backward trajectory analysis uses the global reanalysis database developed by NOAA Air Resources Laboratory [8]. This method can provide a comprehensive understanding of the transport and distribution of pollutants in relation to meteorological dynamics and atmospheric circulation patterns. Using the GBL format file from the official NOAA website, the Hybrid Single Particle Lagrangian Integrated Trajectory (HYSPLIT) module is used to generate the backward trajectory of the air mass [44]. Next, the cluster model was used to calculate pollutant concentrations using the geographic information system [45]-based TrajStat software (version 1.5.5) from Meteoinfo [46].

3. Results and Discussion

3.1. Time Series Analysis of O₃ and Its Related Factors

By monitoring the hourly variations in O₃ and its influencing factors at the Chongming Dongtan supersite in Shanghai from 1 January to 31 December 2021 (as shown in Figure 2), this study identified patterns in O₃ concentrations based on long-term observations of atmospheric pollutants and meteorological parameters in the Chongming region. It was found that O₃ concentrations were higher in both spring and autumn. Specifically, Figure 2a illustrates the seasonal variation of O₃ alongside NOx, PM_2.5, and PAN. O₃ concentrations gradually increased during spring (March–May), slightly declined in summer, rose again in autumn, and dropped to their lowest levels in winter. This pattern is similar to findings from numerous coastal regions, where O₃ concentrations in spring and autumn are higher than in summer [47,48]. In contrast to northern regions, where O₃ peaks typically occur in summer, the highest concentrations in coastal areas are observed in spring and autumn [49]. This trend indicates that O₃ formation is primarily driven by photochemical reactions, with significant increases in O₃ production. The decline in O₃ concentrations during summer can be attributed to the reduction in pollutant levels. NOx concentrations were higher during the winter months (December and January) and lowest in summer, likely due to increased fossil fuel combustion in winter and the rapid photochemical consumption of NOx in summer. The seasonal variation in PAN is closely correlated with that of PM_2.5, suggesting a possible shared source. NOx in the atmosphere is involved in converting into nitrates to promote PM_2.5 formation and also in PAN formation. The similar temporal characteristics of PAN and PM_2.5 imply that PAN may share a similar origin with PM_2.5. We will explore this further in the backward trajectory analysis later.

Figure 2b illustrates the seasonal variation in different VOCs components, including alkanes, alkenes, aromatics, alkynes, and total VOC concentrations. The total concentration of VOCs peaks in winter and gradually declines through spring and summer. This seasonal variation reflects increased combustion activities in winter, coupled with poor atmospheric dispersion conditions. Alkanes constitute the largest proportion of total VOCs, particularly during winter, indicating that their primary source may be combustion emissions. Although alkenes, aromatics, and alkynes represent a smaller proportion, they make significant contributions to the formation of O₃ and PAN. In summer, O₃ levels decreased primarily due to relatively lower precursor concentrations of VOCs and NOx, as shown in Figure 2, while higher wind speeds, as indicated in Figure S2, facilitated the diffusion of these precursors, further contributing to the reduced O₃ concentrations. The seasonal variation in VOCs is closely related to atmospheric photochemical reactions, particularly as they serve as precursors for O₃ and PAN. The seasonal fluctuations of VOCs directly influence the formation processes of atmospheric pollutants.

Figure 2c presents the seasonal variations in temperature, relative humidity [33], and wind speed [34] throughout the year. The temperature peaks in summer and reaches its lowest levels in winter, displaying typical seasonal fluctuations. High temperatures promote photochemical reactions, thereby facilitating the formation of O₃ and PAN. Relative humidity is higher in spring (around 80%) and lower in winter (approximately 60%), with humidity variations influencing the formation of particulate matter and the dispersion of pollutants. Wind speeds are lower in winter and slightly higher in summer. Elevated wind speeds assist in the dispersion and dilution of pollutants, whereas lower wind speeds can lead to pollutant accumulation, particularly in winter. In combination with higher humidity, lower wind speeds in winter may contribute to the accumulation of particulates such as PM_2.5. The seasonal variations in temperature, relative humidity, and wind speed have a significant impact on pollutant concentrations, particularly in summer, when higher temperatures and wind speeds contribute to O₃ formation and the dispersion of particulate matter.

The seasonal variations of O₃, NOx, PM_2.5, PAN, and VOCs are influenced by a combination of atmospheric photochemical reactions and meteorological conditions. The summer peaks in O₃ and PAN are primarily driven by active photochemical reactions, with VOCs and NOx serving as precursors interacting under conditions of sunlight and high temperatures. The concentrations of NOx and PM_2.5 are higher in winter, largely due to anthropogenic emissions and meteorological factors such as combustion emissions and poor atmospheric dispersion. VOC compositions are dominated by alkanes in winter, whereas in summer, they participate in atmospheric chemical reactions and are rapidly consumed, exacerbating O₃ formation. The seasonal fluctuations in meteorological conditions (temperature, relative humidity, and wind speed) play a critical role in pollutant formation, dispersion, and deposition. In particular, the higher temperatures and wind speeds in summer facilitate the generation or dispersion of pollutants.

The seasonal analysis of PAN, O₃, and NO₂ concentrations reveals distinct temporal patterns across spring, summer, autumn, and winter. In spring, summer, and autumn, PAN concentrations peak between 10:00 and 12:00, reaching approximately 0.81 μg/m³, 0.44 μg/m³, and 0.53 μg/m³, respectively. In contrast, O₃ concentrations exhibit a delayed peak between 14:00 and 16:00, with values of approximately 120 μg/m³, 130 μg/m³, and 120 μg/m³ for the same seasons. A similar pattern emerges in winter, where PAN peaks at around 0.75 μg/m³ between 10:00 and 12:00, followed by an O₃ peak of approximately 90 μg/m³ between 14:00 and 16:00. This consistent 2–4 h lag between the PAN and O₃ peaks suggests a potential mechanistic link between PAN decomposition and subsequent O₃ formation.

As shown at Figure 3, the observed temporal relationship is consistent with known photochemical processes. PAN serves as a temporary reservoir for nitrogen oxides (NOx) and organic radicals, which are released upon its decomposition under conditions of elevated temperature and sunlight. This decomposition produces NO₂ and RO₂, both of which are critical precursors in O₃ formation. The peak of PAN concentrations in the late morning, followed by elevated O₃ levels in the early to mid-afternoon, supports the hypothesis that PAN decomposition contributes significantly to O₃ production. Notably, this pattern persists across all seasons, indicating that the underlying photochemical mechanism is robust and not confined to specific seasonal conditions.

While the data provide compelling evidence for the role of PAN in driving O₃ formation, other factors warrant consideration. For example, the availability of VOCs plays a key role in radical cycling within photochemical systems and may enhance O₃ production. Additionally, meteorological variables—such as temperature, relative humidity, and solar radiation—likely influence the rates of PAN decomposition and O₃ formation. Nevertheless, the consistent lag between PAN and O₃ peaks across multiple seasons suggests that PAN decomposition is a primary contributor to the observed O₃ trends, even if modulated by these external factors.

These findings highlight the necessity of incorporating PAN into models of O₃ formation, particularly in suburban environments where PAN concentrations are often elevated. Given its toxicity to human health and its role as a precursor to O₃, a deeper understanding of PAN’s behavior and its interactions with O₃ is essential for evaluating air quality and associated public health risks.

3.2. Model Performance and Feature Important

The dataset used in this study was obtained through the use of multiple instruments that measured the hourly concentrations of various atmospheric parameters. These parameters include PAN, NOx, O₃, PM_2.5, and VOCs. In addition, meteorological parameters such as temperature, relative humidity [33], wind speed [34], and wind direction [47] were also recorded.

Too many features may lead to overfitting of the model, so this study calculated the O₃ formation potential (OFP) of all VOCs using the maximum incremental reactivity [27] method. Based on the OFP, 10 volatile organic compounds were selected as VOC features in descending order according to their OFP. These selected VOCs were then used as input features in the RF model.

In total, 6924 valid samples were obtained during this period. These samples were divided into two subsets: 5539 samples were used for training the model, while the remaining 1385 samples were reserved for testing. The training dataset was used to develop and fine-tune the RF model, enabling it to capture the relationships between the input variables (PAN, NOx, O₃, PM_2.5, selected VOCs, meteorological factors) and the target variable (O₃ concentration). The test dataset was then employed to evaluate the model’s performance and generalizability. For more details on the observation instruments, please refer to the Supplementary Materials.

Based on the results shown in Figure S1, both random forest (RF) and XGBoost demonstrated high predictive accuracy for O₃ concentration prediction. However, a closer examination revealed that RF exhibited a lower R² variance (0.03335 for RF versus 0.04082 for XGBoost) and a slightly higher average R² score, indicating more consistent performance across different cross-validation folds. In contrast, other models such as ANN, RNN, SVR, and KNN produced noticeably lower R² values. Given the ecological sensitivity of the nearshore region under investigation, where robust and reliable predictions are critical, we ultimately selected random forest as the primary model for further analysis and interpretation.

Figure 4a shows the comparison between the O₃ concentration predicted by the RF model and the actual observations, and the results show that the model has good performance, with an R² value of 0.90 and a root mean square error (RMSE) of 11.498. This indicates that the model has high accuracy in predicting O₃ concentration. For a detailed model performance comparison, see the Supplementary Materials. The scatter plot of actual O₃ concentration and model prediction values shows that most data points are closely distributed near the 1:1 line, especially in the medium and low concentration range (about 50–150 μg/m³), showing a strong linear correlation. This indicates that the model has high prediction accuracy. The research by Watson et al. (2019) [33] yielded analogous results when forecasting wildfire-related O₃ exposure in California, with random forest demonstrating superior predictive capability among ten evaluated machine learning models. Zhan et al. (2018) [1] successfully applied random forest (RF) to predict O₃ concentrations in China with high accuracy. RF is widely recognized for its strong predictive performance, particularly in capturing nonlinear relationships between input variables and the target output—a capability that distinguishes it from alternative methods such as support vector machines (SVMs) and neural networks [50,51]. However, there is a slight underestimation trend in the high concentration area (>150 μg/m³), which may be caused by the complexity and scarcity of extreme pollution events.

Feature importance analysis (Figure 4b) reveals the key factors affecting O₃ concentration. PAN and NOx (nitrogen oxides) are identified as the most important predictors, followed by temperature, wind direction, and relative humidity [33]. This ranking emphasizes the combined role of precursors, meteorological conditions, and particulate matter in O₃ generation. Yao et al.’s [52] RF-based O₃ prediction study identified temperature and NOx as the most influential features through importance analysis, aligning broadly with our results. The dominance of NOx highlights the importance of controlling traffic and industrial emissions, while the higher importance of temperature illustrates the significant impact of climate factors on O₃ pollution levels. Figure S3, which presents SHAP values, further supports this analysis. The SHAP plot is similar to the feature importance chart, but it reveals that NOx shows a negative effect on O₃ when its concentration is high, and a positive effect when its concentration is low. This dynamic behavior of NOx emphasizes the complex relationship between precursor levels and O₃ formation.

3.3. Partial Dependence Plot

Based on the machine learning analysis using partial dependence plots (PDPs), the figure reflects the role of four key factors—PAN, temperature, NOx, and relative humidity [33]—in O₃ generation. These figures, respectively, show the impact of each factor on O₃ concentration changes within a specific range, from which we can identify how these factors affect O₃ concentration and thus deduce their positive and negative effects in predicting O₃ concentration.

Figure 5a shows that PAN has a strong positive effect on the production of O₃. As PAN increases from 0 to about 3.5 ppb, O₃ concentration rises from approximately 60 µg/m³ to around 110 µg/m³, suggesting that PAN contributes to O₃ formation. While PAN and O₃ are both products of VOC oxidation in the presence of NOx and may share similar variation characteristics, the observed trend from the PDP highlights that PAN, as a precursor, can indeed elevate O₃ levels under certain conditions. Consistent with Liu’s findings, our analysis reveals that PAN contributes to O₃ formation under NOx-rich conditions. As demonstrated in Figure 3 and Figure 5, the nocturnal increase in NOx concentrations leads to subsequent PAN-driven O₃ enhancement, mirroring the mechanisms reported in Liu’s study [53].

Figure 5b shows the effect of temperature on O₃ concentration. It can be seen from the figure that as the temperature increases, the O₃ concentration gradually increases, showing an obvious positive effect. This is consistent with previous studies [54,55]. When the temperature increases from 10 °C to 30 °C, the O₃ concentration increases from approximately 80 µg/m³ to over 100 µg/m³. This indicates that higher temperatures help promote the generation of O₃, mainly because higher temperatures are usually accompanied by stronger solar radiation and more active photochemical reactions. This phenomenon is especially obvious in summer, because high temperature provides favorable conditions and accelerates the photochemical reaction rate of VOCs and NOx, resulting in a significant increase in O₃ concentration. Therefore, temperature shows a positive contribution in O₃ concentration prediction, especially in high-temperature seasons, when the impact is particularly significant.

Figure 5c shows the impact of NOx on O₃ concentration, showing a typical negative correlation. As the NOx concentration increases from approximately 10 µg/m³ to 50 µg/m³, the O₃ concentration gradually decreases from 100 µg/m³ to approximately 60 µg/m³. This negative correlation can be explained by the “NOx suppression effect”. When the NOx concentration is high, NO will react with O₃ to generate NO₂ and consume O₃, thus inhibiting the generation of O₃. When the NOx concentration is low, O₃ generation is more active; however, when the NOx concentration is too high, O₃ is consumed instead. This negative effect is especially obvious in urban areas with high NOx emissions, indicating that NOx is one of the keys limiting factors for O₃ generation [45]. In the prediction model of O₃, NOx usually shows a negative contribution, and its inhibitory effect on O₃ increases as the concentration increases.

Figure 5d shows the effect of relative humidity [33] on O₃ concentration, which also shows a negative effect. This is similar to previous studies [56]. As the relative humidity increases from 20% to 100%, the O₃ concentration decreases from 110 µg/m³ to 80 µg/m³. High humidity conditions usually mean more cloud cover and weaker solar radiation, inhibiting photochemical reactions from occurring. In addition, high humidity helps the sedimentation of pollutants, further reducing O₃ production. Therefore, increased humidity has a suppressive effect on O₃ concentrations, which is particularly significant in humid climate conditions, especially during cloudy or rainy seasons. This shows that the influence of relative humidity contributes negatively to the prediction of O₃, and higher humidity usually means a lower O₃ production rate.

Similarly to Wang et al.’s [57] study using random forest for O₃ prediction, our partial dependence plot (PDP) analysis of feature impacts on O₃ concentrations revealed comparable patterns: temperature showed limited influence below 0 °C, and NOx exhibited increasing effects at higher concentrations. However, our study demonstrated a stronger impact of relative humidity [33] on O₃ levels compared to Wang’s findings, which is potentially attributable to our study’s unique ecoregion characteristics.

Overall, these four variables show different directions of influence in the prediction of O₃ concentration. PAN and temperature have significant positive contributions to O₃ generation, while NOx and relative humidity have significant negative effects on O₃ concentration. In model predictions, the combined effect of these factors reveals the complex change mechanism of O₃ concentration, emphasizing the different impact paths of different meteorological conditions and precursors in the O₃ generation process.

3.4. Feature Importance in Four Seasons

Since the response of O₃ to different factors may vary by season, this study divides the months into different seasons, trains RF models for each season, and estimates the feature importance for each factor influencing O₃ concentrations. Feature importance plots are then drawn to examine the response of O₃ concentrations to different factors across the seasons.

In Figure 6, we observe the contributions of various chemical components and meteorological factors to the prediction of O₃ concentrations in spring. Notably, temperature, relative humidity, NOx, and PAN all play significant roles in O₃ formation. Among these, NOx is the largest contributing chemical factor in spring. Temperature and relative humidity also have key roles, which is closely related to the accelerated photochemical reaction rates and the influence of humidity on O₃ generation. Changes in relative humidity notably impact O₃ concentration prediction, while higher temperatures promote O₃ formation [54,55]. It is worth noting that VOCs, such as ethylbenzene, n-pentane, and isoprene, contribute less in spring, likely due to the lower reactivity of VOCs under cooler conditions.

In summer, the most significant contributor to O₃ formation is PAN, followed by NOx and relative humidity. This result indicates that O₃ concentrations in summer are primarily driven by high temperatures and photochemical reactions. The rise in temperature accelerates the decomposition of PAN, further intensifying O₃ formation. Additionally, the longer and more intense sunlight during the summer provides abundant energy for O₃ generation. Although relative humidity continues to contribute, its influence is weaker compared to spring.

As we transition to autumn, the influence of PAN on O₃ concentrations becomes more pronounced. At the same time, m/p-Xylene also contributes to O₃ formation. According to research by Xiao et al. [58], the main contribution of m/p-Xylene in Shanghai comes from fuel evaporation. Moreover, PAN still accounts for a significant portion of the contribution, indicating that precursor substances remain crucial for O₃ formation. The photochemical reaction rate in autumn is slower than in spring and summer, so the dominant drivers of O₃ formation shift toward chemical components rather than meteorological conditions.

In winter, the impact of drivers on O₃ formation differs significantly from the previous seasons. The contribution of NOx sharply increases and becomes the most important factor, likely due to the lower temperature and reduced VOC reactivity in winter. Furthermore, the contribution of PM_2.5 increases in winter, potentially due to the higher particle emissions from heating activities, which indirectly affect O₃ formation. It is important to note that meteorological factors have a minimal effect on O₃ in winter, suggesting that under low-temperature and weak-light conditions, O₃ formation is primarily controlled by chemical precursors (especially NOx) and particulate matter. The contribution of VOCs, such as ethylbenzene and isoprene, is negligible in winter, further emphasizing their limited chemical reactivity under low-temperature conditions.

Overall, seasonal variations have a clear impact on the main drivers of O₃ concentrations. In spring, summer, and autumn, PAN is consistently the dominant driver of O₃ formation, followed by meteorological factors and NOx [54,55]. This result is in line with the typical mechanism of O₃ generation, which is driven by photochemical reactions. In the hot and bright summer months, meteorological conditions dominate O₃ formation, while in the colder seasons, O₃ generation is more dependent on the supply of precursor substances. For pollution control, O₃ emission reduction measures should focus on different aspects in different seasons. For example, in summer, emphasis should be placed on controlling O₃ formation driven by meteorological factors, while in autumn and winter, reducing emissions of NOx, VOCs, and other precursor substances should be prioritized.

3.5. Two-DimensionalPartial Dependence Plot of Four Seasons

Figure 7 shows the influence of different meteorological factors and pollutants on O₃ concentration by 2D-PDP analysis and compares them in spring, summer, autumn, and winter. Overall, the generation of O₃ is closely related to season, precursor substances (such as PAN and NOx) and meteorological factors (such as temperature and relative humidity), and their influences show significant differences in different seasons.

Firstly, Figure 7a shows the influence of PAN concentration on O₃ generation. In all seasons, the increase in PAN concentration leads to an increase in O₃ concentration, but this upward trend is most significant in summer. This can be attributed to the fact that under the conditions of abundant sunshine and high temperature in summer, PAN, as a product of photochemical reaction, greatly promotes the generation of O₃. Especially in summer, the O₃ concentration rises rapidly with the increase in PAN, reflecting the strong driving effect of summer photochemical reaction. In contrast, the O₃ concentration in winter is lower and less affected by the change in PAN, which reflects the characteristics of insufficient light and inactive photochemical reaction in winter.

Secondly, temperature, as another important meteorological factor, its influence on O₃ generation is reflected in Figure 7b. The O₃ concentration increased significantly with the increase in temperature in spring and autumn, especially in spring, where the increase in temperature was most closely related to the increase of O₃ concentration. This indicates that the photochemical reaction in spring is more active, and the increase in temperature promotes the transformation of precursor substances and the generation of O₃. However, the O₃ concentration changes more slowly in summer and winter, especially in summer. Although high temperature is a favorable condition for the generation of O₃, at a certain temperature, the O₃ concentration no longer increases significantly due to other inhibitory factors (such as high humidity or the saturation effect of NOx). Similarly, the O₃ concentration in winter does not increase significantly with the change in temperature, indicating that the photochemical reaction is limited in winter. Figure 7c shows the effect of NOx concentration on O₃. In all seasons, the increase in NOx leads to a decrease in O₃ concentration, especially in winter. This phenomenon can be attributed to the so-called “NOx saturation” effect, that is, when the NOx concentration is too high, it will not only not promote the generation of O₃ but will reduce its concentration by reacting with O₃. In winter, due to the weak photochemical reaction, the increase in NOx will significantly inhibit the generation of O₃. In contrast, although O₃ concentrations also decreased in spring and autumn, the negative effects of NOx were not as significant as in winter due to moderate sunlight and temperature. Figure 7d shows the effect of relative humidity on O₃ generation. In all seasons, an increase in relative humidity will lead to a decrease in O₃ concentration, especially in spring and autumn, when the humidity exceeds 60%, the O₃ concentration decreases significantly. This may be because O₃ in the atmosphere is more easily removed under high humidity conditions, thereby reducing its concentration. In addition, increased humidity may also inhibit the occurrence of photochemical reactions, further reducing the generation of O₃. The O₃ concentration in winter was originally low and did not change much with humidity, indicating that humidity had a limited effect on O₃ in winter.

In summary, the figure reveals the sensitivity of O₃ concentration to multiple driving factors in different seasons. In summer, the increase in PAN and temperature significantly promoted the generation of O₃, while NOx and relative humidity had an inhibitory effect on it. In winter, the effect of NOx was the most significant, and high concentrations of NOx significantly inhibited the generation of O₃, while humidity had little effect on O₃. Spring and autumn are relatively balanced seasons, with O₃ concentrations rising when temperature and PAN increase, but high humidity or NOx concentrations also inhibit their production. This seasonal difference reveals the complexity of the O₃ generation mechanism and emphasizes the need to develop O₃ pollution control strategies for different seasons.

3.6. Three-DimensionalPartial Dependence Plot

Figure 8 shows the predicted impact of different combinations of factors on O₃ concentration through 3D-PDP analysis. The four sub-graphs in the figure, respectively, combine different meteorological and pollutant factors in binary form to explore their synergistic effects on O₃ generation. The color represents the O₃ concentration, and the color scale changes from light blue (low concentration) to orange (high concentration), which illustrates the trend of O₃ concentration under specific conditions. From these charts, it can be seen that O₃ concentration is affected by multiple factors, and the change pattern of O₃ concentration under different combinations of factors shows significant differences.

First, the interaction between PAN and NOx on O₃ concentration. Overall, as the concentrations of PAN and NOx increase, the concentration of O₃ also increases. This is similar to previous studies [53]. In particular, when the concentration of PAN gradually increases from 0.01 ppb, the concentration of O₃ increases rapidly, indicating that PAN significantly promotes the generation of O₃ under high NOx concentration conditions. This is consistent with the characteristics of PAN as a photochemical reaction product, and PAN can accelerate the generation of O₃ in a high NOx environment. This phenomenon shows that PAN and NOx have a synergistic effect in the generation of photochemical smog, especially in areas with more serious pollution. When NOx and PAN increase at the same time, the O₃ concentration may rise sharply, leading to more serious photochemical pollution events.

From the synergistic effect of temperature and PAN concentration on O₃. It can be seen that under higher PAN concentrations (1–2 ppb) and higher temperatures (20–30 °C), the O₃ concentration increases significantly. This shows that high temperature and high PAN concentration jointly promote the generation of O₃, especially when the temperature exceeds 20 °C, the O₃ concentration rises rapidly, indicating that the photochemical reaction under high temperature conditions is very active. Elevated temperatures accelerate the thermal degradation of PAN, thereby enhancing photochemical reaction rates and subsequently increasing O₃ concentrations [53]. However, at low PAN concentrations (less than 1 ppb), the O₃ concentration does not change much with increasing temperature. This may be due to insufficient reactants at low PAN concentrations, which cannot fully promote the generation of O₃, so the change in temperature has limited effect on the O₃ concentration. This shows that PAN concentration is a prerequisite for temperature-driven O₃ generation.

The interaction between NOx and temperature found that at lower NOx concentrations, increased temperature significantly promoted O₃ generation [59], especially when NOx concentration was less than 10 µg/m³, and increased temperature significantly increased O₃ concentration. However, as NOx concentration increased to nearly 30 µg/m³, O₃ concentration showed a downward trend regardless of temperature changes and could not be reversed even under high temperature conditions. This suggests that under high NOx concentrations, a “NOx suppression effect” may occur, that is, excessive NOx will react with O₃, resulting in a decrease in O₃ concentration. This result highlights the complexity of the effect of temperature on O₃ generation under high NOx conditions, indicating that a single control of temperature is not sufficient to effectively reduce O₃ pollution in a highly polluted environment.

Finally, the lower right figure shows the effect of relative humidity and temperature on O₃ concentration. It can be seen that under higher temperature conditions (20–30 °C), the increase in humidity has a significant inhibitory effect on O₃ concentration, and when humidity exceeds 60%, the O₃ concentration decreases significantly. This may be because the presence of water vapor under high humidity conditions accelerates the scavenging reaction of O₃ or inhibits the photochemical reaction of precursor substances, thereby reducing the generation of O₃. On the other hand, under low temperature conditions (less than 10 °C), the effect of humidity on O₃ concentration is relatively gentle, indicating that temperature is the key factor determining the generation of O₃, while the effect of humidity is mainly reflected in the inhibitory effect on O₃ concentration under high temperature conditions.

In summary, this figure reveals the effect of the interaction of different meteorological factors and pollutants on O₃ concentration through the 3D-PDP method. High temperature, low humidity, higher PAN, and moderate NOx concentration help promote the generation of O₃, while high humidity and high NOx concentration have the effect of inhibiting the generation of O₃. These findings provide important insights into the formation mechanism of O₃ pollution and emphasize the need to consider the interaction of multiple factors when controlling O₃ pollution, especially to formulate targeted response strategies under different meteorological and pollution conditions.

3.7. Min Depth

The above RF-based analysis of O₃ drivers briefly illustrates their respective impacts. Since O₃ is affected by multiple complex drivers, the interaction between multiple drivers also plays a vital role in the formation of O₃. Therefore, it is necessary to study the different factors. Although RF cannot be directly used to study the interaction between paired drivers, exploring the structure of RF decision trees can help extract information from different factors. The main method of this structural mining analysis is based on the concept of minimum depth, that is, the minimum distance from the depth of a factor to the root of the tree. The minimum depth is within 3, which represents a strong contribution to the prediction of O₃. For a single factor, if its MD is shallow, it is considered to be more important for the prediction of O₃. As shown in Figure 9, in the minimum depth analysis of the RF model, NOx, and PAN are the most important features for predicting O₃ concentration. The same as the previous feature importance and SHAP analysis. Although PAN itself is a secondary pollutant, it has a far-reaching impact in atmospheric chemistry, especially through its thermal decomposition to release NOx, which continues to affect O₃ production. In addition, as one of the VOCs, propane also has a relatively low minimum depth, showing its greater contribution in the model. This is attributed to the reaction of propane with OH· in the atmosphere to form PAN precursors. Although PM_2.5 (fine particulate matter) has a complex relationship with O₃ generation, it can indirectly affect O₃ concentration by adsorption, scattering, or changing the chemical environment in the atmosphere. The effects of meteorological factors such as relative humidity [33] and wind direction on O₃ generation are more reflected in the transmission and diffusion of pollutants. For example, changes in wind direction may affect the accumulation and dilution of pollutants between regions, while humidity can inhibit the occurrence of certain photochemical reactions and indirectly reduce O₃ generation.

For the effect of temperature, although temperature is one of the variables with a more significant impact on O₃ concentration in the previous feature importance and SHAP value analysis, its importance is lower in this minimum depth analysis. This phenomenon may be caused by a variety of factors. First, NOx and PAN show a stronger dominant role in the model, especially in the atmospheric photochemical reaction chain, where NOx is directly related to O₃ production, while temperature indirectly affects O₃ production mainly by regulating the reaction rate. Therefore, although temperature has an impact on atmospheric reactions, its effect may be more indirect than that of NOx and PAN, resulting in a relatively large minimum depth. Secondly, the RF model is able to capture the complex nonlinear interactions between features, and temperature may have a strong interactive effect with other meteorological variables (such as humidity or wind speed), resulting in a small contribution of temperature alone, but it is still important in the overall impact. The feature importance and SHAP value focus more on the direct impact of quantitative variables on the prediction results and may ignore the complexity of this interaction to a certain extent. This shows that although temperature still has an impact that cannot be ignored in O₃ prediction.

3.8. Backward Trajectory Concentration Clustering

The trajectory clustering is shown in Figure 10, and the pollutant concentration clustering table is shown in

Table 1. Pollutant concentration under airflow clustering.

Season	Cluster	PAN (ppb)	PM2.5 (μg/m3)	O3 (μg/m3)	NO2 (μg/m3)
Spring	1	0.4822	29.1791	99.1284	20.4573
	2	0.4551	25.7500	123.8935	2.1351
	3	0.4299	13.4615	122.7578	2.2000
Summer	1	0.2827	14.1898	88.0709	9.0567
	2	0.0197	7.8243	47.7703	2.4729
	3	0.0602	9.1570	96.6904	2.1979
Autumn	1	0.6678	18.8267	114.4603	10.1527
	2	0.4825	24.4539	65.8519	17.2098
	3	0.6928	41.4580	49.7769	45.4563
Winter	1	0.9121	41.8201	68.1357	34.8380
	2	0.4456	17.6980	87.4980	11.7059
	3	0.5257	22.0623	88.8301	6.9723

. Starting from spring, the air mass clustering shows that the airflow from the inland southwest (passing through the Shanghai urban area) accounts for 40.46%, the airflow from the eastern sea direction accounts for 23.14%, and the northward airflow passing through the Shandong Peninsula along the Chinese coastline to Chongming accounts for 36.4%. The clustering analysis of pollutant concentrations reveals that the main sources of PAN and O₃ are located in the inland and marine directions. This suggests that the primary pollution in Chongming Dongtan originates from Shanghai city and from the marine direction. Unlike local sources in Chongming, nitrogen oxides (NOx) emitted by ships have a positive effect on O₃ generation in the marine area farther from the land. Furthermore, time series analysis shows that the variation trends of PM_2.5 and PAN are similar, indicating that PAN and PM_2.5 may share common sources, and PAN is likely to originate from long-range transport.

In summer, the prevailing wind direction (Class I) is from the south of Chongming, accounting for 57.24%, passing through the Jinshan Chemical Park in Shanghai before reaching Chongming. Class II and III wind directions account for 10.01% and 32.75%, respectively, and both come from the sea. Notably, the main contribution to PAN is a result of Class I winds, especially those passing through the Jinshan Chemical Park in Shanghai. In addition, the O₃ concentration contribution is higher under Class II winds, further confirming that ship emissions from the sea lead to an increase in O₃ levels, which are then transported to Chongming Dongtan.

In autumn, the dominant wind direction remains the sea breeze, with Class I winds from the sea accounting for 57.12%. For this wind direction, PAN and O₃ concentrations are generally higher, which further demonstrates that PAN contributes positively to the O₃ levels in Chongming Dongtan after being transported from a distance. In winter, inland winds account for only 21.74%. Although PAN concentrations are higher under Class I winds, its effect on O₃ concentration increase is limited due to low temperatures. However, Class II and III winds account for nearly 80%, and the O₃ clustering concentrations are 87.5 μg/m³ and 88.83 μg/m³, respectively. This suggests that O₃ generation in winter primarily originates from the sea and is transported long-distance, significantly influencing O₃ levels in Chongming Dongtan. PAN’s promoting effect on O₃ generation may exist in certain seasons (such as spring and autumn), but in other seasons (such as summer and winter), its effect may be regulated by other atmospheric factors. Therefore, it can be inferred that the role of PAN in O₃ generation is seasonal and condition-dependent, rather than a simple co-transport relationship. According to the results of machine learning, PAN contributes the most to O₃ in summer, indicating that under high-temperature conditions in summer, the life of PAN is reduced and it decomposes into NOx and VOCs, resulting in a higher contribution to O₃ generation than in other seasons. Additionally, considering the backward trajectory analysis, it is believed that the transport of O₃ from the ocean to Chongming Dongtan affects the local O₃ concentration, but the high concentration of NOx in the local area has a negative impact on O₃ generation.

In summary, although ship emissions play a leading role in the generation of O₃ in spring and summer, especially in high-temperature and strong-light conditions, NOx emitted by ships generates a large amount of O₃ through photochemical reactions; urban emissions are also not to be ignored, especially in autumn and winter, when urban emissions contribute significantly to pollutants such as PAN and PM_2.5. This seasonal difference reveals the combined effect of ship and urban emission sources on the concentration of atmospheric pollutants.

4. Conclusions

In this study, we employed an interpretable random forest model to investigate the key factors influencing O₃ concentrations in the Chongming region of Shanghai, with a focus on the complex interactions among precursors and meteorological conditions. The model demonstrated high predictive accuracy (R² = 0.9), and feature importance analysis identified PAN, NOx, temperature, and relative humidity as primary drivers of O₃ formation. Our analysis revealed clear seasonal patterns, with peak O₃ levels in spring, and uncovered complex, nonlinear relationships between O₃ and its precursors. Notably, while temperature and NOx were positively correlated with O₃, NOx exhibited a dual effect—enhancing O₃ formation at lower concentrations while inhibiting it at higher concentrations, particularly in summer due to intensified photochemical reactions. Relative humidity above 60% suppressed O₃ formation, underscoring the intricate interplay between chemical and physical processes. Structural mining further confirmed that NOx and PAN significantly contribute to rising O₃ levels. Backward trajectory clustering revealed that the remote transport of PAN and O₃ from marine and urban industrial sources plays a critical role in the pollution observed on Chongming Island. These findings highlight the combined influence of precursor emissions and meteorological factors on O₃ pollution, emphasizing the need for temperature-specific, multi-dimensional, and seasonally adjusted pollution management strategies. Overall, while our results are consistent with previous studies on the dual role of NOx and the positive influence of temperature, this work advances our understanding by elucidating the complex interactions between meteorological conditions and precursor emissions, particularly in ecologically sensitive wetland regions.