Revealing the drivers of surface ozone pollution by explainable machine learning and satellite observations in Hangzhou Bay, China

Abstract

Surface ozone (O₃) pollution is an emerging concern in China. Hangzhou Bay (HZB), where the petrochemical industry is clustered, has become one of China's most O₃ polluted areas due to exposure to volatile organic compounds (VOCs) emissions and land-sea breezes. It is urgently need to investigate the multiple drivers of surface O₃ generation in HZB more specifically. The spatial distribution of O₃ trends from April to September (2015–2022) in HZB depicts a general upward trend, with an observed trend of 0.26 μg/m³ a⁻¹, where meteorological factors contribute to 54°% based on the stepwise multiple linear regression (MLR). Ensembled machine learning is more efficient and accurate, especially the Light Gradient Boosting model (LightGBM, R² = 0.84) outperforms other machine learning algorithms. The Shapley additive explanation (SHAP) technique allows for more in-depth quantification of the contribution of specific factors to O₃ trends. The results of the LightGBM-SHAP algorithm present that solar radiation plays a leading role in O₃ generation. More importantly, stronger solar radiation can still lead to high O₃ concentration accumulation even at lower temperature based on the interaction of SHAP values. For the precursor's emissions, the ratio of formaldehyde-to-NO₂ (HCHO/NO₂) obtained from the Tropospheric Monitoring Instrument (TROPOMI) satellite observations, shows the study area is located in the VOCs-limited and transitional regimes, highlighting that VOCs control is more cost-effective.