the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 23 Jun 2026
Meteorological normalization of surface ozone variability across warm and cold seasons in China during 2015–2024
Abstract. Surface ozone (O3) concentrations in China have continued to increase despite substantial reductions in primary air pollutants, but the relative roles of meteorological variability and longer-term non-meteorological changes remain uncertain across seasons and regions. Here, we applied a LightGBM-based meteorological normalization framework combined with SHapley Additive exPlanations (SHAP) to investigate interannual variations in maximum daily 8 h average O3 (MDA8 O3) across China during warm and cold seasons from 2015 to 2024. The model reproduced daily MDA8 O3 variability reasonably well, with mean testing R2 values of 0.71 and 0.77 in the warm and cold seasons, respectively. Observed national mean MDA8 O3 increased by 2.1 μg m-3 yr-1 in the warm season and 1.7 μg m-3 yr-1 in the cold season. After meteorological normalization, MDA8 O3 still increased at 1.2 μg m-3 yr-1 in both seasons, indicating that the decadal O3 increase was mainly associated with non-meteorological components. SHAP analysis revealed distinct seasonal and regional meteorological associations. During the warm season, temperature and solar radiation were more important in northern inland and basin regions, whereas relative humidity and wind-field variables were more important in southern coastal regions. During the cold season, solar radiation dominated across most regions, while relative humidity was more important in the Pearl River Delta. The net meteorological contribution shifted from generally negative during 2015–2019 to positive after 2019, indicating that the recent unfavorable meteorological conditions have amplified O3 pollution in several regions.
- Preprint
(6772 KB) - Metadata XML
-
Supplement
(264 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-2846', Anonymous Referee #1, 16 Jul 2026
-
RC2: 'Comment on egusphere-2026-2846', Anonymous Referee #2, 25 Aug 2026
This manuscript uses site-specific LightGBM models to analyse daily MDA8 O3 observations from 1173 monitoring sites across China during 2015–2024. Meteorological normalisation is applied by retaining the temporal variables while resampling meteorological conditions, and SHAP is then used to explore regional and seasonal differences in the meteorological factors associated with O3 variability. The national-scale comparison of warm- and cold-season O3 is useful, especially the emphasis on the less-studied cold-season increase and its regional differences. These results are relevant to the interpretation of recent O3 trends and may have implications for season-specific and region-specific control strategies. I think the study has the potential to make a useful contribution, but several aspects of the analysis need to be clarified or tested more carefully. In particular, I am concerned about how the “non-meteorological component” is interpreted, the use of random train–test splitting for a long-term trend analysis, the uncertainty in the quantitative attribution, and the strength of some of the SHAP-based interpretations.
1) The normalisation keeps Unix time, DOY and DOW unchanged while meteorological variables are resampled. The resulting normalised series therefore represents the long-term signal retained by the model after averaging over the selected meteorological distribution. I do not think this can be treated as an independently identified “non-meteorological contribution” in a strong sense. This is particularly important for O3 because the model contains no NOx, VOC, aerosol or other precursor-related variables. Changes in precursor emissions, NO titration, chemical sensitivity, aerosol effects on photolysis, background O3 and regional transport may all be partly represented through the time variable or remain mixed within the residual. The authors acknowledge some of this in the uncertainty section, but elsewhere the manuscript gives rather precise values for the “non-meteorological contribution” (e.g. 16.1 and 10.2 µg m-3), which reads as a stronger attribution than the model can support. Recent work on weather normalisation has also highlighted the potential “proxy trap” between temporal and meteorological predictors, particularly for secondary pollutants such as O3 (Dai et al., 2025). I suggest that the authors distinguish more clearly between a meteorologically normalised long-term component and a physically identified non-meteorological contribution.
2) The current sensitivity analysis considers the sampling window and the number of resampling iterations, but the main attribution results still depend on one machine-learning algorithm, one meteorological dataset and one normalisation framework. Recent studies of Chinese O3 have shown that estimated meteorological contributions can vary considerably across datasets and methods (Wang et al., 2025). Earlier machine-learning studies have also produced different estimates of the meteorological contribution over partly overlapping periods. These studies are not directly comparable because the periods and seasonal definitions differ, but they do show that the quantitative attribution is method-sensitive. Before values such as 16.1 µg m-3 versus 3.1 µg m-3 are treated as robust, some comparison with an alternative model or attribution approach would be helpful. The manuscript should also state clearly whether the meteorological variables are resampled together as a daily meteorological vector or independently. Independent resampling would break the covariance among temperature, humidity, radiation, cloud, precipitation and boundary-layer conditions and could generate physically unrealistic combinations.
3) Many of the predictors used here are strongly correlated, particularly temperature, solar radiation, relative humidity, cloud cover, precipitation and boundary-layer height. This may not greatly affect predictive performance, but it can affect how SHAP importance is distributed among predictors. For this reason, statements such as SSRD being the dominant factor at 80.48 % of cold-season sites should be interpreted as model attribution rather than evidence that solar radiation is physically the uniquely dominant control at those sites. This issue becomes more important because the manuscript builds several regional process interpretations around the top-ranked SHAP variable. Some assessment of the stability of these rankings, for example across alternative models, variable groupings or importance metrics, would strengthen this part of the paper. It would also be useful to clarify whether the annual sum of signed meteorological SHAP values is expected to agree quantitatively with the independently calculated meteorological contribution (O-N). If not, the black “net meteorological contribution” lines in Figs. 6 and 7 represent a different quantity from that used in Sect. 3.3 and this distinction should be made clear.
4) The observations are described as covering 2015–2024, while the cold season for a given year is defined as November–December of the previous year plus January–April of the named year. Under this definition, the 2015 cold season requires data from November–December 2014. However, the manuscript reports a 2015 cold-season value and uses 2015–2024 to calculate the cold-season trend. It is not clear whether 2014 data were actually used but not described, or whether the 2015 cold season contains only January–April. If the latter is the case, the first cold season is not directly comparable with the later years and this may affect the estimated trend.
Here are some minor comments:
- The definition of a “short gap” in the KNN imputation procedure is not given.
- It is unclear which variables were used to calculate KNN distance and whether they were standardised before applying Euclidean distance.
- Please clarify whether all 1173 sites have sufficiently continuous coverage over the full 2015–2024 period.
- The statement that approximately 89 % of sites showed increasing trends “in both seasons” is ambiguous.
- The relationship among “observation”, positive contribution, negative contribution and normalised concentration in Fig. 5 is difficult to follow from the current caption and legend.
- There are several minor typographical and citation-formatting issues, including “variablesthat”, “severalsources”, “asstatistical”, inconsistent spacing around year ranges, and the incomplete citation “Wang, L. et al.” in the Introduction.
- “China National Environmental Monitoring Centre” and “China National Environmental Centre” are used inconsistently.
- Some of the proposed physical explanations, particularly those related to RH and wind, are phrased more strongly than would be supported by a statistical association alone.
Citation: https://doi.org/10.5194/egusphere-2026-2846-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 58 | 36 | 11 | 105 | 10 | 9 | 6 |
- HTML: 58
- PDF: 36
- XML: 11
- Total: 105
- Supplement: 10
- BibTeX: 9
- EndNote: 6
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General comments
This manuscript applies LightGBM-based meteorological normalization and SHAP analysis to investigate warm- and cold-season ozone variability across China during 2015–2024. The national-scale dataset and direct comparison between the two seasons are valuable, and the results may provide useful information for ozone management. However, the main quantitative interpretation is still not sufficiently supported by the analysis results. I think major revision is required before the manuscript can be considered for publication.
Specific comments
Chapter 2: Details about the methodology are required. Please clarify the variables used to calculate the KNN interpolation (maximum gap length, proportion of imputed observations by site and year,…), MDA8O3 (the minimum number of valid hourly values, …), etc…. Also, the detailed setup of hyperparameters for LightGBM and for SHAP should be described.
Line 128-: Unix time was included as a predictor to represent long-term changes and was kept fixed during meteorological resampling. However, long-term changes in temperature, radiation, humidity, or other meteorological variables can also be correlated with Unix time. Therefore, the model may assign part of the meteorological trend to the temporal predictor, and the normalized component is not necessarily a uniquely identified non-meteorological signal. I think sensitivity tests excluding Unix time, or using alternative representations of the long-term trend, are necessary before concluding that non-meteorological changes were dominant.
Line 133-: The daily dataset was randomly divided into training and testing subsets at an 80:20 ratio. However, ozone and meteorological variables have strong temporal autocorrelation, so adjacent days or the same pollution episode can be included in both datasets. This may overestimate the model’s performance. I recommend using cross-validation based on selected time series blocks (not randomly).
Lines 270-: The cold season is defined as November–April, which combines winter with the spring. Because ozone can increase rapidly during March and April, especially in southern China, the reported cold-season trend may be mainly driven by springtime changes. Please provide monthly normalized trends or at least separate November–February and March–April analyses.
Lines 307-: The meteorological contribution defined by the normalization method, O−N, and the annual sum of meteorological SHAP values are not necessarily the same quantity because they are calculated relative to different baselines. I think these two estimates should be directly compared for each year and region. Without this comparison, the net meteorological contributions in Figs. 6b and 7b may be interpreted as equivalent to the decomposition in Fig. 5, although this equivalence is not guaranteed.
Lines 362-: Mechanistic interpretations are stronger than the presented analysis can support. Direct NOx/VOC observations and transport or chemistry diagnostics were not included. I think these should be clearly presented as hypotheses unless they are supported with additional chemical observations or process-based modeling.