Global Aerosol Models Considering Their Spatial Heterogeneities Based on Aeronet Measurements

The imprecision inherent in aerosol models poses a significant challenge in satellite- and surface- based aerosol retrieval. The correct classification of global aerosol types and understanding of their spatial heterogeneity are prerequisites for accurately deriving aerosol optical and microphysical properties and assessing their impacts on climate, air quality, and human health. In this study, a deep clustering algorithm is developed and employed to classify global aerosol types by using >160,000 records of AERONET measurements at 411 sites during 2000–2020. The proposed algorithm utilizes deep learning techniques to optimize the feature space of input variables, thereby improving the distinguishability among various aerosol types. This algorithm outperforms the commonly used K-means method with 77.4% of AERONET sites experiencing increased explained variance. Based on their unique properties (e.g., single scattering albedo, complex refractive index and dust ratio, etc.) and geographic distribution, global aerosols are classified into six types: mineral dust, smoke, two urban aerosols with different particle sizes, and two mixtures with different dominant components. The spatial heterogeneity of aerosol optical/microphysical properties is distinctly presented in our clustering results. Through the subdivision of the global domain into 11 sub-regions, a comprehensive elucidation has been undertaken to delineate the variances in volume size distribution, absorption, and scattering characteristics exhibited among disparate aerosol types, as well as within a same aerosol category. These new understandings of aerosol models and their spatial heterogeneities in this study will be highly helpful to update or optimize aerosol retrieval algorithms in the future.