PM2.5 estimated directly from satellite data and from fused data produced by an interpretable multi-model stacking ensemble method

Abstract

Rapid urbanization and industrialization in China have resulted in an increase of PM_2.5 concentrations. In this study, an interpretable multi-model stacking ensemble method (IMSEM) with top-of-the-atmosphere reflectance (TOAR) from the Himawari-8 satellite were used to acquire high-resolution PM_2.5 data in China. In contrast to the traditional approach whereby PM_2.5 is estimated with single models, using TOAR data, IMSEM outperformed single models in terms of several skill scores. The hourly average R² (RMSE) of 10-fold-cross validation reached 0.84 (9.52 μg/m³) in 2021 by IMSEM. The feature importance results of IMSEM showed the significant contributions of TOAR and meteorological variables. The PM_2.5 estimates of IMSEM were also fused with surface observations using interpolation for correction and optimization. When this was done for PM_2.5 concentrations in 2022, it was found that, among the four seasons, the fusion-based estimate of PM_2.5 concentration was highest in winter (49.94 μg/m³), followed by autumn (31.59 μg/m³) and spring (29.07 μg/m³), and lowest in summer (19.25 μg/m³).