Aerosol classification by application of machine learning spectral clustering algorithm

Abstract

Precise understanding of aerosol classification is crucial for accurately quantifying the effects of aerosols on the Earth’s energy budget, improving remote sensing retrieval algorithms, formulating climate change-related policies, and more. In this study, we used aerosol measurements from the quality assured AERosol Robotic NETwork (AERONET) and utilized a multivariate spectral clustering algorithm, a machine learning tool, to classify global aerosols. The spectral clustering algorithm is a variant of the clustering algorithm that employs eigenvalues and eigenvectors of the data matrix to project the data into a lower-dimensional space of a similar cluster. To accomplish this, we considered five aerosol optical parameters: fine-mode Aerosol Optical Depth, Extinction Angstrom Exponent, Absorption Angstrom Exponent, Single Scattering Albedo, and Refractive Index from 150 AERONET sites distributed in six continents (Africa, Asia, Australia, Europe, North and South America) during 1993 to 2022. Using the clustering analysis, we identified four primary aerosol types: dust, urban, biomass burning, and mixed aerosols. Among the continents, the African and Asian sites exhibited the highest contribution of dust aerosols, as the region has significant global dust sources. Conversely, Australia, Europe, North, and South America are predominantly influenced by fine-mode aerosols, given their considerable distance from major dust source regions.