A physical knowledge-based machine learning method for near-real-time dust aerosol properties retrieval from the Himawari-8 satellite data. (1st July 2022)

• Record Type: Journal Article
• Authors: Li, Jing; Wong, Man Sing; Lee, Kwon Ho; Nichol, Janet Elizabeth; Abbas, Sawaid; Li, Hon; Wang, Jicheng
• Is Part Of: Atmospheric environment. Volume 280(2022)
• Publication Date: 2022-07-01
• Subjects: Natural dust aerosol -- Artificial neural network -- XGBoost -- Third-generation geostationary satellite; Air -- Pollution -- Periodicals; Air -- Pollution -- Meteorological aspects -- Periodicals; 551.51
• DOI: 10.1016/j.atmosenv.2022.119098
• Languages: English
• ISSNs: 1352-2310

Abstract

Abstract: Monitoring dust aerosol properties is critical for the studies of radiative transfer budget, climate change, and air quality. Aerosol optical thickness (AOT) and effective radius (R eff ) are two main parameters describing the optical and microphysical properties of airborne dust aerosol. Satellite remote sensing provides an opportunity for estimating the two parameters in spatial coverage and continuously. To take the merits of machine learning algorithms and also utilize the physical knowledge discovered in the conventional retrieval algorithms, a physical-based machine learning method was proposed and applied on the Himawari-8 geostationary satellite for robust retrieval of dust aerosol properties. The main concepts of this study comprise i) constructing the model input data by extracting highly informative features from the Himawari observations according to physical knowledge and ii) exploiting the utility of six state-of-the-art machine learning algorithms in dust aerosol retrieval. The algorithms include artificial neural network (ANN), extreme boost gradient tree (XGBoost), extra tree (ET), random forest (RF), support vector regression (SVR), and kernel Ridge regression (Ridge). The ground-truth AOT and R eff data from AERONET stations were supplied as output labels. The cross-validation technique was adopted for model training and the results show that the ANN model is superior to the other machine learning models for both AOT and R eff estimation, which exhibits the lowest mean absolute error (MAE = 0.0292 and 0.0981) and the highest correlation coefficient ( r = 0.98 and 0.84). When validated on an independent dataset, the ANN model achieved the lowest MAE (0.0334 and 0.1487), and the highest r (0.94 and 0.63). More importantly, when compared against representative physical-based algorithms, the developed ANN model still retains the best performance. Furthermore, the ANN model shows an overall better performance than other machine learning models and also the JAXA Himawari-8 Level-2 AOT product, with examples exhibited in three dust storm events and for continuous monitoring of one of the dust storm events. Additionally, feature importance analysis implies that the important features of dust aerosol identified by the ANN model are consistent with that in physical model-based algorithms. In summary, this study shows great potential for generating near-real-time products of dust aerosol properties from Himawari satellite data. These products can provide a scientific basis for climate and meteorological study regarding severe dust storms. Highlights: Dust optical thickness and effective radius are retrieved simultaneously from Himawari-8 data. Physical knowledge about dust aerosols is integrated with the machine learning models. The ANN outperforms the XGBoost, extra tree, random forest, SVR, and Ridge regression models. The ANN model is superior to the official aerosol retrieval algorithm of Himawari-8. Near-real-time estimation of dust aerosol properties is achievable.