Spatiotemporal variations of PM2.5 and ozone in urban agglomerations of China and meteorological drivers for ozone using explainable machine learning

Ozone pollution was widely reported along with PM2.5 reduction since 2013 in China. However, the meteorological drivers for ozone varying with different regions of China remains unknown using explainable machine learning, especially during the COVID-19 pandemic. Here we first analyzed spatiotemporal variations of PM2.5 and ozone from 2015 to 2022 in eleven urban agglomerations of China. PM2.5 decreased in all regions, with the largest drop in Beijing-Tianjin-Hebei (BTH). In contrast, ozone declined initially but rose during the pandemic in most regions, especially in Cheng-Yu. Probability density curves showed pronounced increase (24.7%) and slight change in the proportion of PM2.5 and ozone meeting the pollution criterions during the pandemic, respectively. Leveraging Random Forest with SHAP analysis, we further established ozone models in typical urban agglomerations with good performance (CV-R2 = 0.80-0.90; CV-RMSE = 8.52-19.20 mu g/m3) during the pandemic, and compared their relative importance of meteorological variables. Particularly, temperature and incoming shortwave flux at top of atmosphere were identified with high importance in high-ozone regions such as Middle Plain and BTH. Increasing importance of PM (e.g., PM10) was found in southern China, e.g., Yangtze River Delta and Pearl River Delta regions. The western China was characterized with more importance of meteorology, especially in Tibet. Surface albedo and sensible heat flux from turbulence were noted distinctively with high importance in Tibet, partly due to their impacts on ozone formation by generating heat source and sink. In addition, sea level pressure (SLP) was revealed with the highest importance (25.2%) in Cheng-Yu, consistent with the fact that synoptic patterns characterized by SLP field could affect ozone pollution in Sichuan Basin. Our results not only provide an understanding of meteorological factors in regional ozone formation in China, but also highlight the feasibility of explainable machine learning in ozone studies.