the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 15 Jan 2026
Deciphering the impacts of meteorology on surface ozone variability in eastern China using explainable machine learning models
Abstract. Understanding how meteorology influences surface ozone variability is critical for interpreting trends and designing effective air quality policies. This study employs explainable machine learning (XML) with SHapley Additive exPlanations (SHAP) to interpret daily ozone variations from 2013 to 2023 across three major regions in eastern China: North China Plain (NCP), Yangtze River Delta (YRD), and Pearl River Delta (PRD). An ensemble of five machine learning models (LightGBM, XGBoost, CatBoost, Random Forest, and Extra Trees) is trained using 14 meteorological variables and two temporal indicators. XML reveals nonlinear, region-specific ozone-meteorology relationships that are broadly consistent with physical understanding, while differences in SHAP attributions across algorithms highlight structural uncertainty arising from multicollinearity among input variables. We use SHAP-derived contributions to attribute warm-season ozone trends to meteorological versus non-meteorological drivers. Before 2019, ozone increases are mainly associated with the temporal proxy for non-meteorological influences (e.g., emission changes), whereas after 2019 meteorological variability dominates regional ozone trends. Exploiting the additive nature of SHAP, we develop a de-weathering framework that partitions daily ozone into a SHAP-based climatological baseline and a meteorology-induced ozone anomaly (MOA). Across all three regions, the magnitude of positive MOA events increases over 2013–2023, while their frequency and duration show no significant trends, indicating a strengthening meteorological amplification of pollution episodes rather than more frequent events. Our results demonstrate both the utility and limitations of XML for disentangling meteorological drivers of ozone pollution and provide new constraints on how meteorology shapes surface ozone under China’s clean air actions.
- Preprint
(3131 KB) - Metadata XML
-
Supplement
(4493 KB) - BibTeX
- EndNote
Status: open (until 26 Feb 2026)
- RC1: 'Comment on egusphere-2026-74', Anonymous Referee #1, 25 Jan 2026 reply
-
RC2: 'Comment on egusphere-2026-74', Anonymous Referee #2, 31 Jan 2026
reply
The authors employed five machine learning algorithms, LightGBM, XGBoost, CatBoost, Random Forest, and Extra Trees to quantify the influences of meteorological factors on ozone variabilities across the three major city clusters in eastern China. The predictors include 14 meteorological variables from MERRA-2, together with two temporal features, the day of year and the UNIX time variable defined as a continuous, non-repeating count of days since 1-Jan 2013. Shapley Additive Explanations (SHAP) were used to quantify the overall feature importance as well as their influence on meteorology-induced ozone anomaly (MOA) during specific time period (e.g., Figure 4b).
This manuscript is generally well written and suited to the scope of ACP. It is a good practice to test and employ multiple machine learning algorithms. Such an approach is less prone to the uncertainties derived from the sole reliance of single algorithm. The authors demonstrate a solid understanding of machine learning (ML) application and potential challenges such as multicollinearity. The justification for retaining correlated predictors is reasonable. (Line 158). I therefore recommend minor revision prior to the acceptance of this manuscript.
Specific comments:
- The abstract is well written and informative. But for lines 31 to 32: I’m not sure the meaning of “indicating a strengthening meteorological amplification of pollution episodes rather than more frequent events”? Does it mean meteorology plays a more dominant role during ozone episodes compared to non-episode days?
- Line 46: Ozone has been increasing since 2013. But policy published in 2013 is cited. It seems this 2013 policy is not relevant to the post-2013 observed ozone increases?
- Line 82: The description of “global explainability” may be slightly imprecise. global explainability is more about how ML model learns and interprets the data. Thus, it is more about the data as a whole but not ML model itself. Just as the authors describe the local explainability as “the data-point level” (Line 86).
- Line 117: One might raise the skepticism that 14 variables are too many. It is best to include a short justification of why this is not an issue here. e.g., the amount of data for training is sufficient. Tree-based algos such as random forest can address collinearity through procedure such as bagging (trained on subset of features/samples and aggregation).
- Line 127: There might be a discrepancy in the variable of solar radiation between MERRA2 and ERA5. (See the TOAR paper by Lu et al.; https://doi.org/10.5194/acp-25-7991-2025, e.g., Figure S12). Modeling with MERRA2, solar radiation might be estimated less important overall. This might partially explain why SR is quantified to be less important than T over YRD here (Figure 2b). There is nothing wrong in using MERRA2. But this potential discrepancy between these two reanalysis products is worth noting.
- Section 2.2: Could the authors provide more details regarding the setups for the five machine learning algorithms (e.g., tree number, learning rate, etc.)? Perhaps put them in supplementary?
- Caption of figure 1: Could the authors explain what numbers in brackets here? Also in the right panel, what do the shaded areas along the line represent?
- L 161: There is confusion about how the ML actually being trained. For “Each region”, does it mean all data within the region of city clusters are all learned as a whole? If so, in L 171, shouldn’t the predicted MDA8 ozone be on the d-th day in year y from a particular site? That is, the dimension of site should be included here?
- Regarding the attributable de-weathering approach in section 2.3, this is a relatively novel and interesting idea. It may be helpful, if the authors could consider providing a more detailed explanation of this part and clarify my following comment:
Shapley is primarily about assigning the contribution of each “player” (i.e., each meteorological variable), to each instance of ozone value. Therefore, the shapely values from all these variables are meaningful for that particular instance (i.e., ozone in a particular day). The authors “compute the climatological mean SHAP values for each meteorological variable by averaging their daily SHAP values across an 11-year period within a 15-day moving window (Eq. (3))”. I am slightly unclear about the interpretation of directly averaging SHAP values across different instances.
For example, if we focus one variable, its SHAP value might be -5 for one instance and +5 for the next. Can these two values be averaged? While I understand that this is somewhat analogous to deriving mean feature importance from SHAP values. But in the case of quantifying feature importance, their absolute values in each instance should be computed first before averaging? I’m not implying SHAP values should not be averaged at all. But more explanations should be given.
- Lines 251 to Lines 253: this sentence might cause confusion. the feature importance appears broadly consistent across the different algorithms, with similar rankings of the key meteorological predictors. The mean |SHAP value| are indeed slightly different, but they all point that T2M is the leading predictor in NCP. Therefore, I suggest authors to rephrase this sentence.
- Line 334: A very minor point. I suggest consistent indexing for figures/tables. i.e., should it be “Figure 4b” but not “Fig. 4b”? Could the authors check through their manuscript?
Citation: https://doi.org/10.5194/egusphere-2026-74-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 146 | 66 | 20 | 232 | 44 | 13 | 14 |
- HTML: 146
- PDF: 66
- XML: 20
- Total: 232
- Supplement: 44
- BibTeX: 13
- EndNote: 14
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This paper uses explainable tree-based machine learning models (XML) to explore the role of meteorology in driving ozone variability in China. The approach shows strong performance in reproducing observed ozone variability, separating meteorological from non-meteorological influences, and identifying key meteorological drivers of ozone anomalies. It effectively addresses limitations of traditional statistical methods and CTM simulations, and the framework could serve as a useful reference for future studies in other regions or time periods.
In terms of scientific findings, the study highlights a shift in the dominant drivers of ozone trends around 2019 in China, along with an increase in meteorology-driven positive ozone anomalies between 2013 and 2023.
The paper is well written—clear, concise, and with uncertainties appropriately discussed. I recommend publication in ACP after the following comments are addressed.