Geographical and temporal encoding for improving the estimation of PM2.5 concentrations in China using end-to-end gradient boosting

Abstract

Fine particulate matter with aerodynamic diameters less than 2.5 μm (PM2.5) profoundly affects environmental systems and human health. To dynamically monitor fine particulate matter over large geographic areas, some machine learning methods have been utilized to estimate its concentration using satellite-based aerosol optical depth (AOD). To improve the estimation of PM2.5 concentration across large areas, a geospatial-temporal joint code is proposed in this paper to characterize the influence of spatial-temporal information hidden in satellite-based aerosol products. This encoding method can reveal the relationship between the PM2.5 concentration and its geospatial location and observation time. Instead of aggregating observation data over neighbors, the method directly encodes the spatial-temporal information as features of the end-to-end gradient boosting model for the estimation of PM2.5. Experimental results of PM2.5 concentration in 2019 across China show that the state-of-the-art method is outperformed by the proposed method by a large margin, with R2 from 0.89 to 0.92, RMSE from 10.35 to 7.89 μg/m3, and MAE from 6.71 to 5.17 μg/m3. In addition, overall partial dependence plots (PDPs) are used for the first time to visualize the complicated relationship between satellite-based aerosol products and PM2.5 concentrations.