Estimating daily full-coverage surface ozone concentration using satellite observations and a spatiotemporally embedded deep learning approach

Abstract

China has experienced an increasing and spreading trend of ozone (O3) pollution in recent years, which can be of significant threat to human health. High-resolution full-coverage O3 data will be highly valuable for O3 pollution prevention and control. To this end, a spatiotemporally embedded deep residual learning model (STE-ResNet) is proposed in this study to obtain daily high-resolution surface O3 concentration data, by the integration of surface station O3 measurements, satellite O3 precursors, reanalysis data, and emissions data. The proposed model uses a novel temporal and spatial embedding technique to represent the temporal (year, month, and day) and spatial (latitude and longitude) information, rather than directly inputting the temporal and spatial information into the neural network. Meanwhile, a gap-filling approach is developed to reconstruct the missing data in the satellite-retrieved O3, so that daily full-coverage O3 data can thus be generated. The proposed approach was applied to a heavily polluted region of China, namely, the Guangdong-Hong Kong-Macao Greater Bay Area (denoted as GBA), where a world-class city agglomeration is being built. The sample-based cross validation (CV), spatial-based CV, temporal-based CV, and external validation demonstrate high consistency with the station measurements, with R2 values of 0.93, 0.90, 0.89, and 0.70, respectively, at a daily level. The spatiotemporal embedding promotes the estimation accuracy compared to directly inputting the temporal and spatial information, with the sample-based CV R2 value increasing from 0.91 to 0.93. The daily full-coverage surface O3 concentrations were obtained at a spatial resolution of 0.05°, and it was found the O3 pollution hotspot is located in the center of the GBA, with mean values of mostly > 80 μg/m3. In the worst case, about 60% of the days in autumn have O3 concentrations exceeding the standard (i.e., 160 μg/m3) in the city of Zhongshan. Furthermore, for a serious pollution incident that occurred on April 9, 2020, the reconstructed O3 data show superior monitoring capabilities, due to the full coverage, when compared with the station measurements and satellite retrievals. The approach proposed in this study will be of great value for the fine-scale and continuous monitoring of O3 pollution.