Capturing synoptic-scale variations in surface aerosol pollution using deep learning with meteorological data

Abstract

The estimation of daily variations in aerosol concentrations using meteorological data is meaningful and challenging, given the need for accurate air quality forecasts and assessments. In this study, a 3×50-layer spatiotemporal deep learning (DL) model is proposed to link synoptic variations in aerosol concentrations and meteorology, thereby building a “deep” Weather Index for Aerosols (deepWIA). The model was trained and validated using 7 years of data and tested in January–April 2022. The index successfully reproduced the variation in daily PM2.5 observations in China. The coefficient of determination between PM2.5 concentrations calculated from the index and observation was 0.72, with a root mean square error (RMSE) of 16.5 µg m−3. The DeepWIA performed better than Weather Forecast and Research (WRF)-Chem simulations for eight aerosol-polluted cities in China. The simulating power of the model also outperformed commonly used PM2.5 concentration retrieval models based on random forest (RF), extreme gradient boost (XGB), and multilayer perceptron (MLP). The index and the DL model can be used as robust tools for estimating daily variations in aerosol concentrations.