This study focuses on mainland China and five major urban clusters with severe O₃ pollution, namely the Beijing–Tianjin–Hebei (BTH) region, the Yangtze River Delta (YRD), the Fenwei Plain (FWP), the Pearl River Delta (PRD), and the Sichuan Basin (SCB) (Fig. 1). These regions were selected because they have persistently high O₃ levels and are widely recognized as major O₃ pollution hotspots in China (Ozone Pollution Control Committee of Chinese Society of Environmental Sciences, 2024). They also represent distinct combinations of emission characteristics, meteorological conditions, and topographic settings, making them suitable for examining the regional heterogeneity of OFS (Wang et al., 2017; Zhang et al., 2025b).

NO_x has a short lifetime in the boundary layer, and NO is rapidly converted to NO₂. Therefore, the NO₂ column is commonly used as a proxy for ambient NO_x levels (Duncan et al., 2010). HCHO is a major intermediate product of VOCs oxidation, with a relatively short lifetime and a dominant distribution in the troposphere, making its total column abundance a useful indicator of tropospheric VOCs (Duncan et al., 2010).

In this study, daily NO₂ and HCHO column data from 2005 to 2023 were obtained from the Ozone Monitoring Instrument (OMI) onboard Aura satellite, available from the National Aeronautics and Space Administration (NASA) Goddard Earth Sciences Data and Information Services Center (https://disc.gsfc.nasa.gov/, last access: 6 September 2025). Launched in 2004, OMI provides daily near-global observations with a minimum spatial resolution of approximately 13 × 24 km and an overpass time of about 13:45 local time (LT) (Levelt et al., 2006).

Tropospheric NO₂ data were taken from the OMNO2d Level-3 product, which provides daily tropospheric NO₂ columns at a spatial resolution of 0.25° × 0.25° in units of molec. cm⁻², and retains only pixels with cloud fraction below 30 % (Krotkov et al., 2019). HCHO data were obtained from the Level-3 OMHCHOd product with a spatial resolution of 0.1° × 0.1° (Chance, 2019). This product is generated based on the Level-2 OMHCHO retrievals after applying quality control filters to exclude pixels with cloud fractions > 0.3, solar zenith angles > 70°, or those affected by the OMI row anomaly. The HCHO columns were retrieved using the Smithsonian Astrophysical Observatory (SAO) algorithm, with typical values ranging from 4 × 10¹⁵ to 4 × 10¹⁶ molec. cm⁻² and with a detection limit of approximately 1.0 × 10¹⁶ molec. cm⁻² (González Abad et al., 2015). Following Zhu et al. (2017), only values between -0.5 × 10¹⁶ and 1.0 × 10¹⁷ molec. cm⁻² were retained, while values outside this range were treated as outliers and removed. To reduce random noise, a 3 × 3 moving average based on the eight-neighborhood was applied to the HCHO fields. The smoothed HCHO data were then resampled to a 0.25° × 0.25° grid using bilinear interpolation to match the NO₂ product, following an approach similar to that of Koukouli et al. (2016) for sulfur dioxide (SO₂) data.

Meteorological data for 2005–2023 were obtained from the ERA5 reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF) (Hersbach et al., 2020; https://cds.climate.copernicus.eu/, last access: 14 September 2025). ERA5 provides monthly data at a spatial resolution of 0.25° × 0.25°, which were further regridded to match the OMI NO₂ grid. The meteorological variables used in this study include 2 m temperature (T), relative humidity at 1000 hPa (RH), surface solar radiation downwards (SSRD), 10 m wind speed (WS), surface pressure (SP), and boundary layer height (BLH) (Weng et al., 2022).

Anthropogenic emission data for 2005–2023 were obtained from the Multi-resolution Emission Inventory for China (MEIC, http://www.meicmodel.org/, last access: 14 September 2025) developed by Tsinghua University, covering five sectors: power, industry, residential, transportation, and agriculture (Li et al., 2017). MEIC provides ten major species: SO₂, NO_x, carbon monoxide (CO), ammonia (NH₃), carbon dioxide (CO₂), non-methane volatile organic compounds (NMVOCs), fine particulate matter (PM_2.5), inhalable particles (PM₁₀), black carbon (BC), and organic carbon (OC) (Zheng et al., 2018; Geng et al., 2024). Following Zhang et al. (2024a) and related work linking emissions to O₃ variability via machine learning (ML), we selected monthly emissions of NO_x, NMVOCs, CO, and PM_2.5 as key predictors for diagnosing OFS. All emission data were resampled to 0.25° × 0.25° to match the grid of satellite NO₂ data, ensuring spatial consistency across datasets.

Surface O₃ concentration data for 2015–2023 (µgm-3) used in this study were obtained from China's national air quality monitoring network. Since 2015, this network has expanded to include more than 330 cities and over 1400 monitoring stations nationwide. The data were collected from the National Real-time Urban Air Quality Publishing Platform provided by the China National Environmental Monitoring Centre (https://air.cnemc.cn:18007/, last access: 20 January 2026).

OFS was diagnosed using the satellite-based HCHO / NO₂ ratio (FNR), defined as: 1FNR=[HCHO][NO2], where [HCHO] and [NO2] represent the tropospheric column densities of HCHO and NO₂, respectively. Because the two quantities have the same units, FNR is dimensionless.

OFS regimes were then identified according to the corresponding FNR threshold ranges. The threshold ranges were determined mainly following the method of Jin et al. (2020). In this study, region-specific FNR thresholds across China were derived by combining OMI satellite observations with surface O₃ measurements. This approach is also consistent with previous studies that derived threshold ranges from fitted O₃–FNR relationships (Wang et al., 2021; Ren et al., 2022; Chen et al., 2024). In practice, the transition of OFS from VOC-limited to NO_x-limited regimes is gradual rather than abrupt. This transition can be influenced by multiple factors, including topography, meteorological conditions, and the spatial resolution of satellite observations (Jin et al., 2020). Because these influencing factors vary across China, the corresponding FNR threshold ranges are also expected to differ among regions. For this reason, OFS regimes were identified using region-specific FNR threshold ranges. China was divided into six regions, including five major urban clusters with severe O₃ pollution, namely BTH, FWP, YRD, SCB, and PRD, and one additional category comprising all remaining areas, referred to as other regions of China (ORC) (Lu et al., 2018). These six regions were used as the basic units for deriving region-specific FNR threshold ranges, which were then applied to identify OFS regimes in each region.

Because O₃ pollution in China occurs mainly during the warm season (April–September) (Lu et al., 2020b), the analysis was restricted to April–September in 2015–2023. To match the OMI overpass time, surface O₃ concentrations at 13:00 and 14:00 LT were averaged for each site and day. The resulting daily O₃ concentrations were then collocated with OMI FNR values at the corresponding grid cells. For each site, the high-O₃ probability was then defined as the fraction of days during the study period on which the 13:00–14:00 mean O₃ concentration exceeded 160 µgm-3. The resulting paired samples were subsequently pooled within each region for threshold derivation.

To focus on polluted conditions, samples with NO₂ vertical column densities below 1.5 × 10¹⁵ molec. cm⁻² were excluded, following previous studies (Jin et al., 2020; Wang et al., 2021). In addition, high-end outliers (5 % or 10 %, depending on the region) were removed prior to fitting to reduce the influence of extreme values (Chen et al., 2024). After quality control, paired high-O₃ probability and FNR samples in each region were grouped into 100 bins according to FNR and a cubic polynomial was fitted to the binned relationship between high-O₃ probability and FNR. The FNR interval corresponding to fitted high-O₃ probability values above the 90th percentile was then defined as the threshold range for that region (Jin et al., 2020).

To quantify the spatiotemporal trends of NO₂ and HCHO columns during the warm season (April–September) from 2005 to 2023, we applied the nonparametric Mann–Kendall (MK) test for trend significance (Hirsch et al., 1982) and the Sen's slope for trend magnitude (Sen, 1968). The MK statistics are computed from all pairwise differences in a time series: 2S=∑i=1n-1∑j=i+1nsgnxj-xi, where n is the number of observations, xi and xj are values at times i and j, respectively (j > i), and sgn(⋅) is the sign function that quantifies the direction of change between two observations: it returns 1 when a later value (xj) is greater than an earlier one (xi), -1 when it is less, and 0 when they are equal. A positive (negative) S indicates an increasing (decreasing) trend. Statistical significance is assessed using the standardized test statistic Z: 3Z=S-1VarSS>00S=0S+1VarSS<0, where Var(S) represents the variance of the statistic S. When |Z| ≥ 1.96, the trend is considered statistically significant at the 95 % confidence level (p < 0.05).

Trend magnitude is estimated by the Sen's slope β, defined as the median of all pairwise slopes: 4β=Medianxj-xitj-ti, where ti and tj represent the time points of observations i and j, respectively. This MK–Sen framework is distribution-free and robust to outliers, making it well suited to satellite-derived time series.

To further identify structural shifts in the long-term evolution of O₃ precursors and OFS, and to provide an objective basis for subsequent phase-based analyses, we applied the Pettitt test to the annual time series of NO₂, HCHO, and FNR in each region. The Pettitt test is a non-parametric method for detecting a single abrupt change point in a time series and does not require the data to follow a normal distribution (Pettitt, 1979). It has been widely used in environmental studies to detect statistically significant shifts in long-term observations (Baruah et al., 2022). The test statistic Ut,n is defined as: 5Ut,n=∑i=1t∑i=t+1nsgnxj-xit=1,2,…,n, where xi and xj are the values of the time series, and sgn(⋅) denotes the sign function. The change point is identified as the time corresponding to the maximum absolute value of Ut,n, and its significance is evaluated as: 6p=2exp⁡-6kt2n3+n2, where kt = max⁡1≤t<n|Ut,n|. A p value below 0.05 was considered statistically significant.

In this study, the Pettitt test was applied to the annual warm-season (April–September) series of NO₂, HCHO, and FNR for each region during 2005–2023. For each variable, the year corresponding to the maximum absolute test statistic was taken as the detected change point. Because the Pettitt test identifies only one abrupt change point, it may not fully capture series with multiple shifts or several equally strong candidates (Zhang and Song, 2015). Therefore, the detected change points were interpreted together with the temporal evolution of each variable and the broader policy context.

Random Forest (RF) is an ensemble machine learning (ML) algorithm based on multiple decision trees and is widely used for classification and regression (Breiman, 2001). It implements Bootstrap aggregation (bagging): each tree is trained on a bootstrap replica of the training data with feature randomness at each split, and predictions are combined by majority vote for classification or arithmetic averaging for regression to improve generalization and robustness (Breiman, 1996).

Departing from most studies that regress O₃ concentrations on candidate factors (Luo et al., 2024; Yao et al., 2024), we frame the problem as classification to identify the key factors distinguishing OFS regimes. In this setting, the RF seeks to minimize overall misclassification. Given an input and an ensemble of trees, the predicted class is the plurality vote: 7y^=argmaxk∑t=1TIht(x)=k, where I(⋅) is the indicator function, equal to 1 if the tth tree assigns the sample to class k, and 0 otherwise.

This study constructs RF classification models for five city clusters, including BTH, FWP, YRD, SCB, and PRD, to characterize the relationships between OFS, defined based on the FNR classification, and multiple environmental factors. Predictor variables included anthropogenic emissions, namely NO_x, NMVOCs, CO, and PM_2.5, as well as meteorological parameters, including T, RH, SSRD, WS, SP, and BLH. It should be noted that, although OFS is diagnosed here using the FNR indicator, the spatiotemporal distribution of FNR-derived OFS states is jointly shaped by precursor emissions, chemical transformation, transport, dilution, and boundary-layer mixing (Bai et al., 2018; Chen et al., 2020; Liu and Shi, 2021). Meteorological conditions therefore do not enter the OFS definition directly but can influence the probability of different OFS states by modulating the underlying precursor environment. Based on this rationale, we treat the diagnosed OFS regimes as response classes and use RF to statistically map their relationships with meteorological and emission-related predictors.

Model training and evaluation were based on annual mean gridded data for the warm seasons (April–September) from 2005 to 2023. Because the primary aim of this study was not to predict OFS in future years, but to characterize the relationships between OFS regimes and their environmental factors over the full study period, a year-wise stratified random sampling strategy was adopted. In studies aimed at predicting future years, a common practice is to train the model on earlier-period data and validate it on later-period data in order to assess temporal extrapolation skill (Ni et al., 2024; Wang et al., 2023). Such a strategy was not adopted here because it would bias model evaluation and subsequent SHAP interpretation toward the late-period data distribution. Instead, samples from each year were randomly divided into 80 % for training and 20 % for validation (Özüpak et al., 2025), so that both subsets contained data from all years. This design preserves interannual variability associated with key periods, including major policy interventions and the COVID-19 pandemic, while maintaining comparable sample distributions across the training and validation sets. In addition, given the class imbalance in the dataset, this stratified sampling strategy helps maintain a more balanced class distribution in both subsets, thereby improving the reliability of model evaluation (Sadaiyandi et al., 2023). It also provides a more suitable basis for subsequent SHAP-based interpretation across different temporal subsets, thereby facilitating comparisons of OFS regimes before and after policy implementation.

To account for class imbalance and optimize model performance, the Synthetic Minority Over-sampling Technique (SMOTE; Fernandez et al., 2018) was applied to the training data. To reduce the structural differences of the model and facilitate consistent comparison across the five regions, a unified set of model parameters was adopted for all regional models. Although this does not eliminate differences in model baseline and data distribution across regions, it helps ensure that the regional models are constructed under a consistent modeling framework. Hyperparameter optimization was performed using GridSearchCV (Hastie et al., 2009; Ahmed et al., 2025) over combinations of candidate parameter values. Based on validation performance across all regions, a single parameter set was selected as the optimal configuration. The final model configuration consisted of 200 trees (n_estimators = 200), a maximum tree depth of 25 (max_depth = 25), and a minimum of 1 sample per leaf (min_samples_leaf = 1).

Model performance on the validation set was evaluated using multiple metrics, including precision score (PS), recall score (RS), F1-score (F1), accuracy score (AS), geometric mean score (GMS), and the Area Under the Curve of the Receiver Operating Characteristic (ROC-AUC) score (Thieu, 2024; Ahmed et al., 2025). All of these metrics range from 0 to 1, with higher values indicating better performance.

SHapley Additive exPlanations (SHAP) is an interpretable ML framework based on cooperative game theory that explains model prediction by assigning each feature a Shapley value representing its contribution to the prediction (Lundberg and Lee, 2017). It is defined as follows: 8φi=∑S⊆1,…,n\iS!n-S-1!n!fS⋃i-fS, where φi is the SHAP value of feature i, S represents the feature subset excluding feature i, fS⋃i and fS are the model predictions for the corresponding subsets, and S!n-S-1!n! ensures that the contributions of all feature subsets are fairly distributed.

A positive SHAP value indicates that a given feature increases the model prediction, whereas a negative SHAP value indicates that it decreases the prediction. In this study, SHAP values were calculated for the RF models in different regions to identify the key factors of OFS regimes, including NO_x-limited, transitional, and VOC-limited regimes, and to further interpret the direction and relative magnitude of their effects.

For each regional model, the mean absolute SHAP value [mean(|SHAP|)] was calculated for each feature. Because SHAP values are defined relative to the model-specific expected output and are influenced by the underlying data distribution, the absolute magnitudes of SHAP values are not directly comparable across different regional models. Therefore, to facilitate cross-region comparison, these overall mean(|SHAP|) values were further normalized to obtain the relative contribution of each variable within each regional model. It should be noted that this metric reflects the statistical contribution of individual features to model predictions rather than their direct causal effects in atmospheric physical or chemical processes.

The derived region-specific FNR thresholds were used to diagnose OFS during the warm season (April–September) from 2005 to 2023. Overall, BTH, FWP, and YRD show a clear two-stage evolution in FNR (Fig. 4). From 2005 to 2012, FNR remained relatively stable, with only small interannual variations, because both NO₂ and HCHO increased during this period and their changes largely offset each other in the ratio. In contrast, FNR increased markedly during 2013–2014. This change was mainly driven by a sharp decline in NO₂ together with a continued rise in HCHO, which led to a pronounced increase in FNR and a gradual shift of OFS toward NO_x-limited regimes. Based on the derived thresholds, BTH and FWP were dominated by VOC-limited and transitional regimes during the warm season, whereas YRD was characterized primarily by transitional and NO_x-limited regimes. This evolution likely reflects the combined influence of emission-control policies and climate-related factors. On the one hand, major clean-air policies implemented after 2013 substantially reduced NO_x emissions. On the other hand, warmer conditions and stronger solar radiation may have enhanced O₃ production efficiency and prolonged the persistence of HCHO (Wu et al., 2024), thereby contributing to higher FNR values and stronger NO_x-limited characteristics.

In contrast, SCB shows a relatively high annual mean FNR of 3.34 but no significant change point or clear long-term trend. In addition, SCB did not show a statistically significant change point in the Pettitt test, which further supports the decision not to interpret its FNR evolution using the same two-stage framework. Together with the relatively small changes in both NO₂ and HCHO, this suggests that the precursor structure in SCB remained comparatively stable over the study period, and OFS stayed largely within the transitional regime.

PRD differs from the other two patterns by exhibiting a significant long-term increase in FNR, at a rate of 0.06 yr⁻¹ (p < 0.05). This trend indicates a shift in OFS from VOC-limited or transitional regimes toward more NO_x-limited regimes, likely associated with reductions in NO_x-rich sources in the region, particularly from road traffic and industrial emissions, as suggested by previous studies (Bian et al., 2019; Li et al., 2025).

The timing of the detected change points broadly coincides with the implementation of major clean-air policies in China, suggesting that precursor control patterns in the major urban clusters changed substantially after the early 2010s. At the same time, under conditions of climate warming and enhanced radiation, and against the background of anthropogenic precursor reductions, continued increases in HCHO may have further promoted the shift toward more NO_x-limited chemistry (Itahashi et al., 2022; Wu et al., 2020).

Based on the OMI-derived FNR, OFS in China during the warm season from 2005 to 2023 shows a clear spatial pattern (Fig. 5). VOC-limited regimes are concentrated mainly in eastern urban clusters, whereas NO_x-limited regimes are more common in surrounding suburban and less developed areas. In regions such as BTH and FWP, strong anthropogenic NO_x emissions lead to relatively low FNR values, resulting in predominantly VOC-limited regimes in most years. Away from urban cores, NO_x emissions gradually decrease, while the relative contribution of BVOCs becomes more important, leading to higher FNR values and a shift toward NO_x-limited regimes (Zhang et al., 2017; Wu et al., 2020). This spatial pattern is generally consistent with previous studies and reflects regional differences in emission structure, land cover, and urbanization level (Chen et al., 2020; Ren et al., 2022).

The areal fractions of different OFS regimes also show a clear temporal shift before and after the implementation of the Air Pollution Prevention and Control Action Plan (Fig. 5d). During 2005–2012, the OFS distribution in eastern China shifted toward more VOC-limited regimes, mainly through transitions from NO_x-limited to transitional regimes and from transitional to VOC-limited regimes. Over this period, the area fraction of NO_x-limited regimes decreased by 7.9 %, while the fractions of transitional and VOC-limited regimes increased by 5.6 % and 2.3 %, respectively. This change is consistent with the faster increase in NO₂ than in HCHO during this period, which lowered FNR and favored VOC limitation. PRD showed a different evolution, with a decline in VOC-limited regimes and a predominance of transitional regimes by 2012.

During 2013–2023, the national OFS pattern shifted back toward NO_x-limited and transitional regimes. This change was associated with substantial reductions in NO_x emissions, comparatively slower decreases in anthropogenic VOC emissions (Fig. S5), and continued enhancement of BVOCs emissions under warmer and more strongly irradiated conditions (Zeng et al., 2023; Zhang et al., 2025a). By 2023, VOC-limited regimes accounted for only 2.8 % of the total area, whereas NO_x-limited regimes expanded to 76.5 %. This evolution indicates that OFS in China did not change monotonically over time but instead underwent a pronounced reversal under the combined influence of emission-control policies and climate-related factors (Fang et al., 2026). It also highlights the asymmetric reduction of O₃ precursors during China's clean-air actions and the resulting shifts in regional O₃ chemistry.

The nationwide COVID-19 lockdown in early 2020 caused a sharp but short-lived reduction in anthropogenic emissions and provided a useful case for examining the response of OFS to abrupt emission changes. To capture this process, the period was divided into three phases: pre-lockdown (1–22 January 2020; 22 d), during lockdown (23 January–13 February 2020; 22 d), and post-lockdown (14 February–7 March 2020; 23 d).

Figure 6 shows the spatial distributions of O₃ precursors and FNR during these three phases. Before the lockdown, high tropospheric NO₂ columns formed a continuous belt across the urban-industrial regions of North, East, and South China. During the lockdown, both the intensity and spatial extent of high NO₂ decreased markedly, consistent with strong reductions in industrial activity and traffic emissions. In the post-lockdown phase, NO₂ columns recovered partially in many regions, although they remained below pre-lockdown levels (Qiao et al., 2022; Zhao et al., 2022). HCHO showed a broadly similar pattern, but its spatial changes were less uniform. During the post-lockdown phase, HCHO increased over parts of South and Southwest China. This increase was likely related to the combined influence of wildfire emissions from Myanmar and northern Vietnam and enhanced BVOCs emissions (Stavrakou et al., 2021). These results suggest that seasonal and BVOCs sources played an important role in sustaining HCHO levels and increasing FNR during this period.

The corresponding OFS distributions also changed substantially across the three phases. Before the lockdown, VOC-limited and transitional regimes dominated most major urban regions in eastern China. During the lockdown, the sharp decline in NO₂ led to a rapid increase in FNR and a broad expansion of NO_x-limited regimes. This shift was especially evident in PRD and YRD, where OFS changed rapidly from VOC-limited toward transitional or NO_x-limited regimes. Such changes are consistent with strong NO_x reductions that weakened urban O₃ titration and altered precursor limitation. In the post-lockdown phase, as NO_x emissions partly recovered, parts of the North China Plain and northern YRD shifted back toward VOC-limited and transitional regimes. However, the spatial extent of these regimes did not return to pre-lockdown levels.

The short-term OFS response during the COVID-19 lockdown is broadly consistent with the longer-term evolution observed from 2005 to 2023. In both cases, rapid reductions in NO_x promoted a shift toward more NO_x-limited regimes, whereas partial NO_x recovery or enhanced VOC emissions tended to move OFS back toward transitional or VOC-limited regimes. These results indicate that OFS in China is highly sensitive to short-term emission perturbations, especially under meteorological conditions favorable for photochemical production. They also suggest that the effectiveness of O₃ control depends strongly on precursor reduction pathways and background climatic conditions. In warm-season urban regions, sustained NO_x reductions are more likely to support O₃ mitigation, whereas coordinated VOC control remains important in regions or periods where OFS may shift back toward transitional or VOC-limited regimes.

Table 1 presents a comparison of the five model configurations. Overall, all models demonstrate relatively high classification performance, with AS, GMS, and ROC-AUC all exceeding 0.70. Among the regions, the models for the YRD and PRD perform the best, with all category-specific metrics generally above 0.70.

In contrast, the BTH and FWP models show relatively low performance for minority classes, particularly the NO_x-limited category, with F1 scores of 0.60 and 0.57, respectively, while other categories generally achieve metrics above 0.60, indicating moderate performance in identifying this category. In the SCB model, the NO_x-limited category performs the worst, with PS, RS, and F1 all below 0.50, reflecting the difficulty for the model in effectively capturing its characteristics.

Overall, after sampling and hyperparameter optimization, the models achieve robust performance on the majority classes, but classification performance for minority classes remains limited. Ahmed et al. (2025) noted that SMOTE can effectively improve classification performance by addressing data imbalance, but it may also lead to decreased accuracy for majority classes, highlighting the trade-offs involved when handling imbalanced datasets.

For each regional model, SHAP values were first calculated for each sample and for each OFS class. For each class, mean(|SHAP|) was then computed for each feature across all samples, and these class-specific mean(|SHAP|) values were subsequently averaged across the three OFS classes to represent the overall importance of each feature in that regional model. To facilitate comparison among regions, the resulting overall mean(|SHAP|) values were further normalized and expressed as relative contribution percentages. Figure 7 shows the relative contribution of each feature to OFS classification in the five city clusters. Overall, emission-related variables contribute more than meteorological variables, with their total contribution ranging from 50.1 % in YRD to 72.2 % in PRD. This result indicates that OFS classification is primarily constrained by emission-related factors in most city clusters. In YRD, meteorological and emission-related variables contribute almost equally to OFS classification, accounting for 49.9 % and 50.1 %, respectively.

A closer examination of individual emission predictors reveals substantial regional heterogeneity. In BTH, FWP, and YRD, CO and PM_2.5 make the largest contributions (17.5 % and 14.0 % in BTH; 16.3 % and 16.5 % in FWP; 14.3 % and 13.3 % in YRD). NMVOCs also show a relatively large contribution in BTH, reaching 13.0 %. By contrast, SCB and PRD are characterized by high contributions from CO and NO_x (19.0 % and 15.9 % in SCB; 24.4 % and 17.6 % in PRD). The strong model importance of CO and PM_2.5 likely reflects both the influence of their significant long-term trends in the model and their broader association with regional photochemical environments. First, both variables show a two-stage trend in most urban clusters from 2005 to 2023, with an initial period of non-significant change followed by a significant decline (Fig. S5), a pattern that is similar to the phase-like changes in OFS identified in this study (Fig. S6). Second, CO can enhance hydroperoxy radical (HO₂) production through the reaction CO + OH → HO₂, thereby increasing O₃ production efficiency under conditions of declining NO_x (Ren et al., 2013; Seinfeld and Pandis, 2016). Third, a reduction in PM_2.5 may weaken its heterogeneous uptake of hydroxyl radicals (HO_x) and NO_x and thus allow more radicals and reactive nitrogen to remain available for photochemical O₃ reactions (Li et al., 2019). Therefore, these results suggest that CO and PM_2.5 function not simply as co-emitted pollutants, but also as informative predictors of broader changes in oxidation capacity and regional pollution conditions.

Meteorological predictors also show robust and regionally distinct contributions. Surface pressure (SP), relative humidity (RH), and temperature (T) are consistently important across regions, highlighting the roles of weather system stability, moisture conditions and thermal conditions in modulating photochemistry and pollutant accumulation. In the YRD, the contribution of RH is particularly notable (25.5 %), while in the FWP, SP contributes more substantially (23.3 %), indicating that OFS in these regions is also strongly influenced by meteorological factors. More broadly, meteorological factors can influence atmospheric oxidation capacity, boundary-layer structure, and pollutant dispersion, thereby shaping both the spatial distribution and temporal evolution of OFS. Therefore, the relatively high contributions of certain meteorological factors in the YRD and FWP may reflect a greater sensitivity of OFS classification to regional humidity and atmospheric stability conditions.

To further characterize the temporal evolution of the factors influencing OFS, we compared the relative contributions of emission-related and meteorological variables between the warm-season periods of 2005–2012 and 2013–2023 (Fig. 8). Overall, emission-related variables remained the dominant predictors across most urban clusters and OFS categories in both periods, with their contributions generally exceeding 50 %. This suggests that the long-term evolution of OFS in China was primarily governed by changes in precursor emissions, although the relative role of meteorology varied by region and OFS category.

For the NO_x-limited regime, the contributions of meteorological and emission-related factors in YRD were broadly comparable, whereas emission-related contributions in BTH, YRD, and PRD increased slightly after 2013 (+4.6 %, +2.2 %, and +3.5 %, respectively). By contrast, meteorological contributions increased moderately in FWP and SCB (+4.2 % and +4.3 %, respectively). In the VOC-limited and transitional regimes, emission-related factors also remained dominant overall. However, meteorological contributions to the VOC-limited regime increased markedly after 2013 in BTH, FWP, and PRD, by +8.6 %, +6.7 %, and +13.0 %, respectively. For the transitional regime, meteorological contributions increased in all urban clusters except YRD, accompanied by corresponding declines in emission-related contributions of approximately 2.0 %–7.3 %.

From an overall perspective (Fig. 8d), the relative contribution of emission-related factors decreased slightly after 2013 in BTH, FWP, SCB, and PRD, by about 2.0 %–5.6 %. In YRD, the relative contributions of meteorological and emission-related factors remained broadly comparable, although the balance shifted slightly toward emission-related influences after 2013. Taken together, these results indicate that, while OFS classification remained predominantly emission-driven, the regulatory role of meteorology became more pronounced after 2013. This pattern suggests a modest transition from a regime controlled mainly by emissions toward one jointly modulated by emissions and meteorological conditions. Such a shift is consistent with the substantial reduction in precursor emissions after 2013 and implies that, under a cleaner emission background (Fig. S5), OFS classification may become increasingly sensitive to meteorological variability.

These findings are broadly consistent with previous regional studies, while also highlighting the strong spatial heterogeneity of OFS influencing factors across China. In the Guangdong–Hong Kong–Macao Greater Bay Area, FNR has been reported to be mainly influenced by temperature, shortwave radiation, total column water, and surface pressure (Chen et al., 2020), whereas in Guangdong Province it is primarily associated with surface solar radiation, relative humidity, and temperature (Liang et al., 2024b), although emission-related variables were not explicitly considered. By contrast, in the Chengdu–Chongqing region, summer OFS has been shown to be mainly driven by CO, PM_2.5, NMVOCs, and NO_x emissions, with increasing pollution shifting OFS toward VOC-limited regimes (Xu et al., 2025). In the North China Plain, the decline in PM_2.5 concentrations reduces the loss of HO₂, enhances their reaction with NO, thereby strengthening NO_x sensitivity and leading to increased O₃ levels (Li et al., 2019). Together, these studies support the view that OFS evolution reflects the combined influence of emissions and meteorology, but that their relative roles differ substantially among regions. Compared with these regional analyses, our results provide a unified national-scale framework showing that emission-related factors generally dominate OFS classification, while meteorological modulation becomes more evident under a cleaner emission background.

To further interpret how individual predictors influence OFS classification, we examined the SHAP value distributions of different variables for each urban cluster (Fig. 9). The results reveal both common features and regional differences in the directional and nonlinear effects of the predictors. In general, many variables show broadly opposite SHAP responses between VOC-limited and NO_x-limited regimes, indicating that the same predictor can exert contrasting effects under different chemical sensitivity conditions.

For emission-related variables, samples classified as NO_x-limited are generally associated with relatively low emission levels. By contrast, under VOC-limited regimes, higher emission values are more consistently associated with positive SHAP values, indicating a greater probability of being classified as VOC-limited. This pattern is broadly consistent with the spatial distribution of OFS, in which NO_x-limited regimes are more common in suburban or peripheral areas, whereas VOC-limited regimes are concentrated in urban cores with stronger anthropogenic emissions (Xue et al., 2014; Wang et al., 2021; Johnson et al., 2024). Overall, the regional differences in the SHAP responses of emission variables likely reflect differences in local emission intensity and chemical sensitivity.

Meteorological variables also show substantial but regionally distinct effects on OFS classification. SP, RH, and T emerge as influential predictors in several regions, but their directional effects vary geographically. In BTH and YRD, higher temperatures are more closely associated with NO_x-limited classification, which may reflect enhanced photochemical activity and BVOCs emissions under warm conditions (Pétron et al., 2001; Duncan et al., 2009). In several other regions, however, higher temperature is more often linked to VOC-limited classification, which may also be related to the spatial distribution of urban heat islands (Zhou et al., 2015). SP influences pollutant accumulation and vertical mixing through atmospheric stability and subsidence; in general, higher SP favors VOC-limited classification, whereas lower SP is more closely associated with NO_x-limited regimes, with particularly asymmetric SHAP responses in YRD. RH is especially important in BTH and YRD, where high humidity tends to promote VOC-limited classification. Higher relative humidity may promote aerosol growth, thereby reducing the radiation reaching the surface and limiting the photochemical production of HCHO (Ye et al., 2016).

Compared with VOC-limited and NO_x-limited regimes, the transitional regime shows less consistent directional responses across variables and regions. This suggests that transitional regimes are governed by a more complex combination of emissions and meteorological influences, with stronger nonlinearity and greater classification uncertainty. Such complexity implies that transitional regimes may be the most difficult to manage using simple precursor-reduction strategies alone and may require more detailed process-based analysis to support effective control design.

Overall, the SHAP analysis highlights that OFS classification is shaped by both shared and region-specific predictor responses. Emission-related variables remain the dominant contributors in most regions, whereas meteorological variables modulate the OFS transitions in a region-dependent manner. These results provide a more process-informed interpretation of the regional heterogeneity and temporal evolution of OFS across China during the warm season.