High-resolution prediction of the spatial distribution of PM2.5 concentrations in China using a long short-term memory model

The concentration of fine particulate matter (PM2.5) has a significant impact on the environment and human health. However, strong spatial heterogeneity and spatiotemporal dependence increases the difficulty of prediction. Moreover, due to the lag of the update of auxiliary variables at national scale in the prediction application, it is still difficult to achieve the timely nationwide PM2.5 prediction at present. To better model and predict real time concentrations and spatial distributions of PM2.5, this study developed a workflow of future PM2.5 concentrations prediction based on long short-term memory (LSTM) model. Using ground-based station PM2.5 data in 2014–2018, the 1 km Multi-Angle Implementation of Atmospheric Correction (MAIAC) aerosol optical depth (AOD) product and other auxiliary data to predict PM2.5 concentrations in the next year and generate a high-resolution national PM2.5 concentration spatial distribution map. The LSTM model outperformed random forest (RF) and Cubist approaches for prediction PM2.5 because of its recurrent neural network structure that can capture time dependence and nonlinear relationships among PM2.5 concentrations and other independent variables, and exhibited a stable accuracy with an R2 of 0.83, by applying the annual time series, with an improvement of 0.04–0.09, compared to daily and monthly data. The results indicated that PM2.5 pollution had gradually decreased in 2019 after application of pollution controls, with annual mean PM2.5 concentrations of 27.33 ± 15.56 μg m−3, although there were still some areas with severe pollution, including the North China Plain, parts of the Loess Plateau, and the Taklimakan Desert. The LSTM model makes it possible to predict fine-scale PM2.5 spatial distributions nationwide in the future and may thus be useful for sustainable management and control of air pollution at a national scale.