Estimating hourly PM2.5 concentrations using MODIS 3 km AOD and an improved spatiotemporal model over Beijing-Tianjin-Hebei, China

Abstract

The spatiotemporal distribution of PM2.5 during heavy pollution is a short-term dynamic change process, and quantifying the dynamic change process of PM2.5 is the premise and guarantee for short-term PM2.5 exposure research. However, given the low temporal resolution of polar-orbiting satellites and late launch time of geostationary satellites, the application of remote sensing aerosol optical depth (AOD) data in hourly PM2.5 spatial distribution prediction is greatly limited, which brings uncertainty to short-term PM2.5 exposure research. This study introduces the PM2.5 concentration predicted by Moderate Resolution Imaging Spectroradiometer (MODIS) 3 km AOD data and the PM2.5 concentration of monitoring stations into a spatiotemporal autoregressive (STAR) model to generate hourly PM2.5 spatial distribution and quantify the short-term dynamic change process of PM2.5. The monitoring data in the Beijing–Tianjin–Hebei (JingJinJi) region of 2014 were used to test the model performance. Time-based 10-fold cross-validation (CV) R2 was 0.82, and the root-mean-square prediction error (RMSE) was 37.37 μg/m3. The CV R2 and RMSE were higher by 0.04 and lower by 3.4 μg/m3 than the STAR model without monitoring station PM2.5 concentration as predictors, which indicated that the monitoring station PM2.5 concentration could improve the performance of the model. Hourly performance statistics results showed that the model's accuracy increased when the time was closer to the MODIS transit time compared with that at other hours. The farther away from the MODIS transit time, the greater the monitoring stations' PM2.5 concentration improved the performance of the model. The predicted results of the spatial distribution of PM2.5 showed that the spatial distribution of the average PM2.5 concentration in each hour varied greatly in JingJinJi, and the maximum difference reached 30 μg/m3. The model in this paper not only demonstrates high prediction accuracy but also provides high spatiotemporal resolution of PM2.5 for short-term PM2.5 exposure studies.