Global classification of aerosols based on ground observations: A comparison between an empirical method and machine learning algorithms

Aerosols are atmospheric particles that remain suspended in both solid and liquid states, significantly influencing climatic processes through the scattering and absorption of sunlight. They possess distinct physical, chemical, optical, and microphysical properties. Moreover, identifying aerosol types is also essential, as it provides crucial insights into their characteristics and behavior. Therefore, the current study investigates the global distribution of aerosol's optical and microphysical properties across 50 AERONET (Aerosol Robotic Network) sites located on six continents. Based on these properties, aerosols were classified using an empirical method into six major types: Dust Aerosols (DA), Continental Aerosols (CA), Highly Absorbing (HA), Low Absorbing (LA), Non-Absorbing (NA), and Uncertain (UC). The dominance of aerosol types, their mean frequency, and their spatial and temporal variations were analyzed across all sites during the overall study period as well as during summer and winter seasons. The findings indicate that DA and UC aerosol types are prevalent in West Africa and the Middle East during both the overall period and summer. The CA type shows relatively uniform distribution globally during the overall and winter seasons, with the exception of the East Asian region. Conversely, HA and LA types are less frequent in West Africa during the overall period and display reduced occurrences in both West Africa and the Middle East during summer. The NA type is predominantly observed in East Asia across all periods—overall, summer, and winter. To complement the empirical classification, three machine learning (ML) algorithms—Support Vector Machines (SVM), Naive Bayes, and Logistic Regression—were applied for aerosol classification and their performance compared with the empirical method. The effectiveness of the models was evaluated using performance metrics, including accuracy, precision, recall, F1-score, and a confusion matrix. Among the ML methods, SVM and Naive Bayes achieved the highest overall accuracy of 98 % and 96 %, respectively, with perfect classification (100 % precision, recall, and F1-score) for the NA type. Logistic Regression also performed well, with an accuracy of 96 %, along with high precision (97 %), recall (100 %), and F1-score (99 %) for the NA type.