Machine Learning Reveals China’s Ozone Keeps Rising Even in Cold Winters

Surface ozone pollution in China has continued to climb over the past decade, and a new study shows the increase is not confined to the smog-prone summer months. Using a decade of nationwide monitoring data and machine-learning-based meteorological normalization, researchers found that ozone rose significantly in both the warm and cold seasons from 2015 to 2024, with unfavorable weather shifts after 2019 further amplifying pollution in several regions. The findings suggest that ozone control strategies built around summer photochemistry alone may miss a substantial part of the problem.

The study, published in Atmospheric Chemistry and Physics, analyzed hourly surface ozone observations from 1,142 monitoring stations operated by the China National Environmental Monitoring Centre. The team focused on maximum daily 8-hour average ozone, or MDA8 O3, the standard metric for assessing ground-level ozone exposure. They divided each year into a warm season from May to October and a cold season from November to April, with November and December assigned to the following cold-season year so that complete winters could be compared across the decade.

To separate the influence of weather from longer-term changes, the researchers built a LightGBM model, a gradient-boosting decision-tree algorithm well suited to capturing nonlinear relationships in large environmental datasets. Separate models were fitted for each monitoring site and season, using nine meteorological variables from the ERA5 reanalysis, including temperature, solar radiation, cloud cover, boundary-layer height, wind components, surface pressure, relative humidity, and precipitation, alongside temporal predictors such as day of year and day of week. The models reproduced daily ozone variability reasonably well, with mean testing R2 values of 0.62 in the warm season and 0.68 in the cold season.

Meteorological normalization then worked by repeatedly resampling weather conditions from a window around each target date across all available years, effectively asking what ozone concentrations would have looked like under average meteorological conditions. Each prediction was averaged over 1,000 resampling iterations. The difference between observed and normalized concentrations represents the meteorological contribution, while the normalized time series reflects a longer-term component that weather alone cannot explain. Importantly, the authors caution that this normalized component is a composite signal that may include changes in precursor emissions, background ozone, regional transport, and other processes the model does not explicitly represent, and should not be read as a pure measure of emission changes.

The observed numbers are striking. National mean warm-season MDA8 O3 increased at a rate of 2.0 micrograms per cubic meter per year, rising from 98.3 micrograms per cubic meter in 2015 to 117.0 in 2024, a cumulative increase of roughly 19 percent. The cold season rose even faster, at 2.2 micrograms per cubic meter per year, climbing from 66.7 to 86.7 micrograms per cubic meter. Among all monitoring sites, 88.6 percent showed increasing warm-season trends and 94.3 percent showed increasing cold-season trends, with statistically significant increases at 38.7 and 52.0 percent of sites respectively.

After meteorological normalization, the increases persisted. Normalized ozone still rose at 1.1 micrograms per cubic meter per year in the warm season and 1.2 in the cold season, both highly significant, indicating that meteorological variability alone cannot account for the observed trends. The observed increases exceeded the normalized increases by about 3.6 and 4.2 micrograms per cubic meter over the decade, meaning weather contributed to part of the change but was not the dominant driver. Positive normalized trends appeared at 86.7 percent of sites in the warm season and 92.0 percent in the cold season.

Regional patterns revealed sharp geographic contrasts. The Fenwei Plain showed the strongest warm-season increase, with a trend of 3.4 micrograms per cubic meter per year and a cumulative rise of 44.2 percent, reaching a regional mean of 145.2 micrograms per cubic meter by 2024. The North China Plain, Yangtze River Delta, and Sichuan-Chongqing Basin followed with trends of 2.7, 2.5, and 2.0, while the Pearl River Delta increased more moderately at 1.3. In the cold season, normalized trends were actually higher than warm-season trends in four of the five key regions, with the Fenwei Plain again leading at 2.0 micrograms per cubic meter per year and the Sichuan-Chongqing Basin showing the smallest increase at 0.6.

To probe which weather variables mattered most, the team applied SHAP analysis, a game-theory-based method that attributes model predictions to individual input variables. The results showed clear seasonal and regional structure. During the warm season, temperature and solar radiation dominated in northern inland and basin regions such as the North China Plain, Fenwei Plain, and Sichuan-Chongqing Basin, consistent with the role of heat and sunlight in accelerating photochemical ozone production. In the southern coastal Pearl River Delta, relative humidity and meridional wind were instead the leading variables, likely reflecting humid monsoonal air masses, cloudiness, and regional ventilation. In the cold season, solar radiation was the dominant variable at 84.5 percent of sites nationwide, while relative humidity remained prominent in the Pearl River Delta.

The SHAP analysis also tracked how meteorological contributions shifted over time. In the warm season, summed regional SHAP values were generally negative during 2015 to 2018, turned positive in most regions in 2019, dipped again around 2020 and 2021, and became broadly positive after 2021, with all five regions showing positive sums in 2024. In the cold season, the sums showed marked interannual variability and regional heterogeneity: they were negative in all five regions in 2015 and 2016, positive in all five regions in 2018 and 2022, and in 2024 positive in four regions while the Pearl River Delta was negative. The authors note that SHAP values describe how the fitted model associates predictors with predicted ozone; they do not by themselves establish physical causality or quantify direct physical effects.

The study carries important caveats. The absence of long-term precursor observations, the coarse 0.25-degree resolution of the reanalysis meteorology, and the inherent limits of statistical attribution all constrain interpretation. Even so, the central message stands: ozone has intensified across China in every season, and meteorological conditions have increasingly worked against pollution control since 2019. The authors argue that future ozone control plans should account for seasonal and regional differences, including the often-overlooked cold-season increases in southern coastal China.

Subject of Research: Meteorological normalization of surface ozone trends across warm and cold seasons in China from 2015 to 2024

Article Title: Meteorological normalization of surface ozone variability across warm and cold seasons in China during 2015–2024

Article References: Ye, Z., Liu, R., Hu, T., Chen, G., & Wu, M. (2026). Meteorological normalization of surface ozone variability across warm and cold seasons in China during 2015–2024. Atmospheric Chemistry and Physics, 26(20), 15291-15307. https://doi.org/10.5194/acp-26-15291-2026

Image Credits: AI Generated

DOI: 10.5194/acp-26-15291-2026

Keywords: ozone pollution, China, meteorological normalization, machine learning, LightGBM, SHAP analysis, air quality, cold-season ozone, precursor emissions, ERA5 reanalysis, MDA8 ozone, regional trends