Estimating ground-level ozone concentration in China using ensemble learning methods

Following the successful implementation of the Air Pollution Prevention and Control Action Plan (2013—2017) and the Three-Year Action Plan to Win the Blue Sky Defense War (2018—2020), the concentrations of five major pollutants (i.e., PM2.5, PM10, SO2, NO2, and CO), except for ozone, significantly dropped for most cities in China. The increasing ground-level ozone concentrations have been a key factor restricting the improvement in ambient air quality, especially during summer. Compared with the measurements from ground-based monitoring sites, satellite remote sensing technology can obtain spatially continuous total column ozone. However, given that ozone is abundantly distributed in the stratosphere, ground-level ozone has a very low contribution to the total column ozone observed from space. Therefore, the satellite total column ozone product provides limited information for estimating ground-level ozone concentrations. In this study, by combining TROPOMI ozone precursor (NO2 and HCHO) products, ERA5 meteorological parameters, and ground-based monitoring data, a machine learning model was developed to estimate the daily maximum 8-hour average ground-level ozone concentration over China from years 2019 to 2020. By comparing the performance of three ensemble learning methods, namely, extreme gradient boosters (XGBoost), extreme random trees (ERT), and gradient boost regression tree (GBRT), the averaged overall 10-fold cross-validation R2 of 2019 and 2020 are all larger than 0.89. Although the results estimated by XGBoost showed the best agreement between the model predictions and observations with an average RMSE and MAE of 15.77 μg/m3 and 10.53 μg/m3, respectively, the ERT method was eventually selected to model the daily maximum 8-hour average ground-level ozone concentration by considering the rationalization of spatial distribution. Due to the proactive emission reduction measures implemented by the Chinese government and the impact of the COVID-19 pandemic, the rising trend of ozone concentration over the years has been reversed. The annual average value of ground-level ozone concentration in 2020 reached 107.41±18.6 μg/m3 over China, which is 1.85 μg/m3 less than that recorded in 2019 (109.26±19.71 μg/m3). Severe surface ozone pollution events frequently occur from May to September of every year because the high temperatures during these months can promote photochemical reactions. The estimated ground-level ozone concentrations in the Beijing-Tianjin-Hebei, Yangtze River Delta, Pearl River Delta, and Chengdu-Chongqing regions are significantly higher than those in their surrounding areas, making these regions the key areas for ozone pollution prevention and control. © 2023 National Remote Sensing Bulletin