Assimilating a blended dataset of satellite-based estimations and in situ observations to improve WRF-Chem PM2.5 prediction

Abstract

Continuous improvement in remote sensing observation-based techniques has enabled data inversion assimilation methods to play an increasingly important role in the field of numerical forecasting. Using an artificial intelligence fusion inversion method can overcome the multisource heterogeneity of data, thereby providing a better blended source for data assimilation (DA) and improving the forecasting capability. This study proposed a multimodel stacking machine learning algorithm to estimate surface concentrations of PM2.5 (particulate matter with an aerodynamic equivalent diameter of ≤2.5 μm) using Himawari-8 satellite data. The derived satellite-based PM2.5 estimation was then blended with the in situ observations as physical constraints to obtain a high-resolution multisource blended dataset. To determine the level of improvement provided by the blended dataset, four parallel experiments were conducted during February 2022: a control experiment without DA (noDA), an experiment that assimilated satellite-based estimation dataset (sate-basedDA), an experiment that assimilated multisource blended dataset (blendDA) and an experiment that assimilated in situ observations (siteDA) based on the Gridpoint Statical Interpolation system. Statistically, in comparison with the direct satellite-based PM2.5 estimations, the blended dataset better matched the in situ observations. The accuracy of PM2.5 predictions can be optimized by DA and for the first 24-h period the blended dataset performed better than the traditional ground observations as the assimilation source. Throughout the initial 24-h period, the results of blendDA (siteDA, sate-basedDA) showed an improvement in terms of root-mean-square error, with reductions varying from 6.65 to 20.20 μg/m3 (4.50–19.79 μg/m3, 2.29–10.86 μg/m3).