Improving the Estimation of PM2.5 Concentration in the North China Area by Introducing an Attention Mechanism into Random Forest

Abstract

Fine particulate matter with an aerodynamic diameter less than 2.5 µm (PM2.5) profoundly affects environmental systems, human health and economic structures. Multi-source data and advanced machine or deep-learning methods have provided a new chance for estimating the PM2.5 concentrations at a high spatiotemporal resolution. In this paper, the Random Forest (RF) algorithm was applied to estimate hourly PM2.5 of the North China area (Beijing–Tianjin–Hebei, BTH) based on the next-generation geostationary meteorological satellite Himawari-8/AHI (Advanced Himawari Imager) aerosol optical depth (AOD) products. To improve the estimation of PM2.5 concentration across large areas, we construct a method for co-weighting the environmental similarity and the geographical distances by using an attention mechanism so that it can efficiently characterize the influence of spatial–temporal information hidden in adjacent ground monitoring sites. In experiment results, the hourly PM2.5 estimates are well correlated with ground measurements in BTH, with a coefficient of determination (R2) of 0.887, a root-mean-square error (RMSE) of 18.31 μg/m3, and a mean absolute error (MAE) of 11.17 µg/m3, indicating good model performance. In addition, this paper makes a comprehensive analysis of the effectiveness of multi-source data in the estimation process, in this way, to simplify the model structure and improve the estimation efficiency of the model while ensuring its accuracy.