Estimation and impact factor analysis of 24-h near-surface ozone concentration in China using FY-4A/B collaboration and machine learning

Abstract

Ozone pollution in China's urban agglomerations poses a significant environmental challenge. Nine machine learning models were constructed based on the Extra Tree (ET) algorithm, utilizing top-of-atmosphere radiation (TOAR) data from Fengyun (FY)-4A and FY-4B Advanced Geosynchronous Radiation Imager (AGRI), to estimate 24-h near-surface ozone concentrations across China from June 2022 to May 2023. Analysis identified five TOAR channels strongly correlated with ozone concentrations: channels 7, 8, and 11–13 for FY-4A, and channels 7, 8, and 12–14 for FY-4B. The all-sky data model demonstrated superior performance in ozone estimation, achieving an R² of 0.91, outperforming models using only cloudy or clear-sky data. Through partial dependency plots and feature importance assessments, key meteorological drivers were identified: relative humidity below 60 % and temperatures between 20 and 35 °C. These findings provide valuable insights for ozone forecasting and pollution control strategies.