Refined spatiotemporal estimation model of PM2.5 based on deep learning method

The concentration distribution of fine particulate matter (PM2.5) on the surface of the atmosphere has a strong temporal and spatial heterogeneity. Due to the limited spatial coverage of traditional PM2.5 monitoring sites, it is difficult to reflect the complexity of PM2.5 concentration in time and space. This paper proposed a temporal and spatial prediction model of ground PM2.5 concentration based on deep learning methods(PM2.5-DNN). Based on the AOD data from Kuihua-8satellite and the observation data from PM2.5 monitoring and meteorological station, hourly high-precision simulations of the surface PM2.5 concentration in Beijing had been carried out. The results show that the 1km resolution hourly ground PM2.5 concentration in Beijing area estimated by the PM2.5-DNN model had good consistency with the observation data from the surface monitoring station. The model estimation accuracy could reach R2=0.88, which was better than the performance of current mainstream method. The method proposed in this paper was suitable for fine-grained modelling and estimation of the temporal and spatial distribution of PM2.5 concentration at a regional scale. The end-to-end training method is used to construct the model, which provides a simple and effective method model for fine PM2.5 concentration estimation. © 2021, Editorial Board of China Environmental Science. All right reserved.