The study focused on temporal variability and short-term prediction at the selected site rather than long-term changes across the city. For the predictive analysis, the researchers compared deep neural network, random forest and extreme gradient boosting, or XGBoost, models. The enhanced XGBoost model performed best on the test data, with an R² of 0.9756 and a root mean square error of 6.35 micrograms per cubic metre.

These figures quantify the model’s ability to reproduce the held-out measurements at the study location; they do not establish the physical causes of the observed ozone concentrations or demonstrate that the model will perform equally well elsewhere. The team used SHapley Additive exPlanations, or SHAP, to estimate how individual variables contributed to model predictions. The ozone concentration recorded one hour earlier was the dominant predictor, showing that recent ozone conditions carried most of the model’s predictive information.

This result indicates strong short-term persistence rather than an independent estimate of a chemical forcing on each occasion. It also means that the high test performance should be interpreted in light of a model that was given the previous hour’s ozone measurement, not as a complete account of the processes that generate ozone. SHAP analysis also found negative predictive contributions from nitrogen dioxide and one-hour-lagged nitrogen dioxide.

The authors report that these associations are consistent with the expected role of nitrogen dioxide in titrating ozone under high-nitrogen-oxides conditions. Temperature and radiation-related variables had positive contributions to predicted ozone. These are statistical relationships detected in the model, not direct estimates of causal effects.

The daily and seasonal patterns provide context for the model’s results, but the attribution analysis alone cannot determine how changes in individual pollutants would alter future ozone concentrations. The study’s main contribution is a quantitative description of the conditions statistically associated with short-term ozone variability at one Wuhan monitoring site. Its interpretation is limited by the absence of direct volatile organic compound observations, which the authors say leaves ozone sensitivity to those compounds and related control implications unverified.

They recommend further investigation using volatile organic compound measurements or chemical transport and observation-based modeling.

Subject of Research: Interpretable machine-learning analysis of hourly surface ozone variability at one Wuhan monitoring site, 2019–2025.

Article Title: Factors influencing surface ozone concentrations in Wuhan: an interpretable machine learning analysis

Article References: Yang, D.-Y., Chai, S.-Y., Wang, J.-X., Yu, J.-Y., Yu, Z.-Z., Cao, Y., Wen, Y.-Y., Bai, F., Zhao, W.-H., & Guo, W.-J. (2026). Factors influencing surface ozone concentrations in Wuhan: an interpretable machine learning analysis. Environmental Monitoring and Assessment, 198(11), Article 1229. https://doi.org/10.1007/s10661-026-16066-7

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16066-7

Keywords: surface ozone, Wuhan, interpretable machine learning, XGBoost, SHAP, air quality, ozone persistence, meteorology, nitrogen dioxide, hourly monitoring, Factors, influencing