Long-term variability and machine-learning-based forecasting of aerosol optical regimes in Northeast Asia using AERONET observations

This study characterizes the long-term variability of five major aerosol types: desert dust, urban industrial, clean marine, continental, and mixed, at five Aerosol Robotic Network (AERONET) sites in Northeast Asia (Beijing, Seoul, Jeju, Hong Kong, and Fukuoka) from 2012 to 2024. Daily Level 2.0 aerosol optical depth (AOD), Ångström exponent (AE), and single-scattering albedo (SSA) were analyzed using K-means clustering, with cluster interpretation guided by published optical property criteria. A support vector machine (SVM) was used to quantify the separability and reproducibility of the K-means-derived labels rather than independent ground-truth validation. The classified aerosol mixtures differed markedly among sites, with larger mixed and urban-industrial contributions in Beijing and Hong Kong and stronger continental and clean-marine influences in Jeju and Fukuoka. Using linear regression, historical changes in annual aerosol-type occurrence fractions during 2012–2024 were summarized. In addition, a regression neural network trained on the 2012–2024 time series was used to forecast the annual aerosol-type fractions for 2025–2030, from which the forecast period rates of change were derived independently. The historical analysis indicates declining urban-industrial and mixed fractions and increasing clean-marine fractions across the study region, whereas the neural network forecasts suggest site-dependent departures from historical rates. These results reveal a regional transition toward less pollution-dominated aerosol optical regimes and demonstrate that intermediate continental and mixed aerosols remain difficult to discriminate using AOD–AE–SSA alone.