Full-coverage spatiotemporal estimation of surface ozone over China based on a high-efficiency deep learning model

Abstract

Ozone concentration Monitoring is essential to atmospheric pollution prevention and control. Against the background of severe ozone pollution over China in recent years, a spatiotemporal contiguous ozone concentration mapping method was developed. We imputed the significant data gaps of the Ozone Monitoring Instrument's tropospheric NO2 content by using an efficient machine learning model named LightGBM. Then, we developed a deep learning model based on three-dimensional Convolutional Neural Network architecture for daily maximum 8 h average ozone concentration estimation over China. With the support of the satellite-retrieved precursor, meteorological data and other ancillary data, our model achieved excellent performance with sample-based 10-fold cross-validation R2 = 0.88. Furthermore, we generated daily maximum 8 h average ozone datasets covering the whole of China from 2016 to 2020. This study presents a novel method for surface ozone modeling, which can provide fundamental data for the study of the ecological changes caused by ozone pollution, such as crop loss, or the harmful effects on humans, such as the increased incidence of respiratory diseases.

Keywords

• Ozone pollution
• China
• Deep learning
• CNN