Estimation of PM2.5 concentrations in North China with high spatiotemporal resolution using the ERA5 dataset and machine learning models

Abstract

Previous studies have predominantly focused on developing high spatiotemporal resolution PM2.5 models utilizing moderate-resolution imaging spectroradiometer (MODIS) aerosol optical depth (AOD) products alongside certain meteorological factors. However, MODIS AOD is not continuous, and there is a problem of missing data, which is not conducive to the study of PM2.5 in areas without AOD. This study proposes a new method for estimating PM2.5 concentration in North China, which not only simplifies the complex process associated with the use of AOD, but also partially solves the issue of missing AOD data. This study proposes a machine learning (random forest, RF; back propagation neural network, BPNN) based method for estimating PM2.5 using meteorological parameters obtained from the fifth-generation reanalysis (ERA5) dataset released by the European Centre for Medium-Range Weather Forecasts and considering the effects of land cover type, digital elevation model, vegetation index and population data. The model performed well, with 10-fold cross-validation (coefficient of determination) R2 and (root-mean-square error) RMSE of 0.79/0.95 and 26.10/13.22 µg/m3, respectively, for BPNN and RF. The estimated hourly performance of the RF model in winter (00:00 to 23:00 BST) with an R2 ranging from 0.92 to 0.96, an RMSE of 11.45 to 16.70 µg/m3. RF model with the best performance and ERA5 were selected to build a high-resolution (0.25°×0.25°) hourly PM2.5 map (PMM), and the PMM was compared with CHAP, with R2 and RMSE of 0.75 and 20.62 µg/m3, respectively. This study further investigates the impacts of land cover types, digital elevation model, and land surface characteristics on the spatial distribution of PM2.5 in North China.