Every Hour, Across East Asia: 24-hour, Full-Coverage Ground-level Ozone Estimation Techniques using Machine Learning

As global warming intensifies, ground-level ozone pollution has emerged as a critical environmental issue. Ground-level ozone is formed through photochemical reactions and exhibits a characteristic diurnal pattern, with concentrations rising during the day and declining at night. Most studies have primarily focused on high ozone levels during the day or the Maximum Daily 8-hour Average, often neglecting nighttime variations. Moreover, satellite-based ozone monitoring typically excludes areas under cloud cover, leading to gaps in the spatial distribution of the output maps. This highlights the need for continuous and comprehensive ozone monitoring over both time and space.

In response, this study developed a machine learning-based hourly ground-level ozone estimation model for South Korea and China, regions severely impacted by ozone pollution. The model utilizes Himawari-8 AHI brightness temperatures (BT), ECMWF ERA5 meteorological data, tropospheric NO2 vertical column densities from the TROPOspheric Monitoring Instrument (TROPOMI), and additional auxiliary data. A two-way coupled framework was employed, integrating a clear-sky model that uses satellite data and a cloud-sky model that operates without it. This approach enabled the creation of continuous spatiotemporal ozone concentration maps.

In random cross-validation, the clear-sky model demonstrated high accuracy with an R² of 0.90 and RMSE of 8.82 ppb, while the cloud-sky model achieved similarly strong performance with an R² of 0.88 and RMSE of 9.07 ppb. Notably, incorporating BT data in the clear-sky model significantly enhanced the model's ability to estimate ozone concentrations. The resulting ozone maps effectively captured diurnal variations in ozone levels, providing more detailed spatial resolution compared to ECMWF CAMS reanalysis data. This study offers valuable insights into the dynamic changes in ozone levels and underscores the potential of the proposed model as a reliable tool for environmental monitoring and research.