Estimating 1-km PM2.5 concentrations based on a novel spatiotemporal parallel network STMSPNet in the Beijing-Tianjin-Hebei region

With the development of industry, the issue of air pollution is of great concern. Due to the sparsity of monitoring stations, acquiring full-coverage PM2.5 concentration remains challenging. The satellite remote sensing datasets provide a potential solution for estimation tasks by its spatiotemporal virtue. In this study, the ellipsoidal coordinate system was introduced for the first time to improve the spatial coding method. A two-stage algorithm using satellite datasets and ground-site values was proposed to impute the missing AOD and estimate PM2.5 concentrations. Firstly, the Multilayer Perceptron (MLP) model was utilized for imputing the missing AOD, achieving superior accuracy (R2 = 0.929). Secondly, the training efficiency and the accuracy of traditional algorithm were enhanced by innovatively utilizing an efficient attention mechanism and a novel embedding layer. The concept of combining convolutional layers with different output channels and novel spatial preprocessing methods were also innovatively proposed. Consequently, the Spatiotemporal Multi-Sample Parallel Network (STMSPNet) was constructed to estimate daily PM2.5 concentrations. Finally, the best performance of this model was obtained by 10-fold cross-validation with R2 of 0.913 and RMSE of 10.637 μg/m3. In addition, this study analyses the changing patterns of PM2.5 concentrations in the Beijing-Tianjin-Hebei region from 2019 to 2023, taking into account the COVID-19 outbreak and extreme weather. The reasons for the changes are also discussed in depth through the air pollution control policies enacted by the Chinese government and regional factors. The results show that STMSPNet has a strong estimation advantage. [Display omitted] •We developed a two-stage model for AOD imputing and PM2.5 concentration estimation.•Using MLP model to obtain full-coverage AOD datasets.•Constructing spatiotemporal model with temporal and multi-sample module.•Application of Efficient Attention and new Embedding layer.