Predicting long-term ground-level ozone concentration in China using satellite remote sensing data and machine learning models

Abstract:

Overwhelming epidemiological evidence has shown that ground-level ozone is associated with a series of adverse health outcomes. However, research efforts to quantify the health impact of ozone exposure in China are hindered by the limited spatiotemporal coverage of ground monitoring data. It is, therefore, crucial to establish a full-coverage ozone prediction model at a fine-resolution. In the present study, we developed a two-stage random forest model to predict monthly ground ozone concentrations in China at a spatial resolution of 0.05°. We first obtained the Ozone Monitoring Instrument (OMI) ozone profile product (OMO3PR) from the Smithsonian Astrophysical Observatory (SAO). Its missing values were imputed with a random forest model incorporating the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2) assimilated meteorological data. With the gap-filled ozone profile, another random forest model was trained to predict the monthly mean values of the daily maximum 8-hour average ozone (MDA8) in China from 2017 to 2019. The prediction model also included 3-D MERRA-2 meteorological data, land-use terms, and population density. Our model achieved a high prediction accuracy with an overall out-of-bag (OOB) R² of 0.93, and a ten-fold cross-validation R² of 0.92. The prediction map showed that the worst ozone pollution occurred in the most industrialized regions of China, i.e., the Yangtze River Delta and the North China Plain, highlighting the effect of anthropogenic emissions on tropospheric ozone production. Stratospheric intrusion, long-range transport, and biomass burning also contribute substantially to surface ozone pollution, especially in springtime. The prediction model established in this study would facilitate further epidemiological studies to investigate the health effect of ozone exposure.