Full-coverage estimation of PM2.5 in the Beijing-Tianjin-Hebei region by using a two-stage model

Abstract

The accurate estimation of fine particulate matter (PM2.5) is significant for both environmental protection and health assessment. However, the sparsity of monitoring stations poses a challenge to provide continuous and large-scale PM2.5 monitoring data. The extensive coverage and continuous spatial distribution of satellite remote sensing products make them popular for estimating ground-level PM2.5 concentrations through inversion modeling. Although aerosol optical thickness (AOD) is commonly utilized for modeling PM2.5 concentrations, achieving accurate and comprehensive PM2.5 estimation remains a significant challenge. One of the main limitations is the random missing data in AOD. In this study, we propose a two-stage model that combines satellite remote sensing and ground monitoring data to achieve AOD filling and PM2.5 estimations in Beijing-Tianjin-Hebei region. In the first stage, the daily full-coverage AOD was filled by the Light Gradient Boosting Machine (Light-GBM), and achieved a superior performance (R2: 0.93, RMSE: 0.043). In the second stage, a spatio-temporal feature extraction model for PM2.5 concentration estimation was designed based on graph neural network (GNN), namely the spatio-temporal estimation model (ST-GAT). Finally, the ST-GAT obtained the performance of R2 = 0.88, RMSE = 12.66 μg/m3, and MAE = 8.66 μg/m3 on five-fold cross-validation, and the results showed that the method could provide reliable full-coverage PM2.5 estimation.