Coupling the linear mixed effects model with random forest improves hourly PM2.5 estimation from Himawari-8 AOD over the Yangtze River Delta

Abstract

PM2.5 harms human health and has been the focus of many research and monitoring activities. Although high temporal and spatial resolution information of PM2.5 is critically required by many applications, under most circumstances it is in general unobtainable. However, the recently released Himawari-8 AOD product, which has a high-frequency observation at 10 min, may provide a potential solution to the dilemma. In this study, we adopted three of the commonly used parametric models as well as the Random Forest (RF), to estimate the PM2.5 concentration based on the hourly Himawari-8 AOD product over the Yangtze River Delta (YRD) in 2021. Furthermore, a two-stage strategy combining both parametric models and the RF has been investigated. Among all three parametric models, the linear Mixed Effects Model (MEM) has the best performance with a test R2 of 0.77 and an RMSE of 12.13 μg/m3. The results have revealed that the performance was improved based on the two-stage strategy. In particular, the combination of MEM and RF performed best, reaching a cross-validation R2 of 0.85, while having a test R2 of 0.82 and an RMSE of 10.67 μg/m3. Consequence analyses on the average PM2.5 concentration over the YRD region suggested that the PM2.5 concentration reached the highest in winter and the northern part of this region. Our results suggested that the Himawari-8 AOD product is a feasible bridge providing a fast way to grasp the highly dynamical PM2.5 concentration at a large scale and thus provides an invaluable data source for PM2.5 monitoring and assessment.