An aerosol classification scheme for global simulations using the K-means machine learning method

Abstract

Abstract. The K-means machine learning algorithm is applied to climatological data of seven aerosol properties from a global aerosol simulation using EMAC-MADE3. The aim is to partition the aerosol properties across the global atmosphere in specific aerosol regimes; this is done mainly for evaluation purposes. K-means is an unsupervised machine learning method with the advantage that an a priori definition of the aerosol classes is not required. Using K-means, we are able to quantitatively define global aerosol regimes, so-called aerosol clusters, and explain their internal properties and their location and extension. This analysis shows that aerosol regimes in the lower troposphere are strongly influenced by emissions. Key drivers of the clusters' internal properties and spatial distribution are, for instance, pollutants from biomass burning and biogenic sources, mineral dust, anthropogenic pollution, and corresponding mixtures. Several continental clusters propagate into oceanic regions as a result of long-range transport of air masses. The identified oceanic regimes show a higher degree of pollution in the Northern Hemisphere than over the southern oceans. With increasing altitude, the aerosol regimes propagate from emission-induced clusters in the lower troposphere to roughly zonally distributed regimes in the middle troposphere and in the tropopause region. Notably, three polluted clusters identified over Africa, India, and eastern China cover the whole atmospheric column from the lower troposphere to the tropopause region. The results of this analysis need to be interpreted taking the limitations and strengths of global aerosol models into consideration. On the one hand, global aerosol simulations cannot estimate small-scale and localized processes due to the coarse resolution. On the other hand, they capture the spatial pattern of aerosol properties on the global scale, implying that the clustering results could provide useful insights for aerosol research. To estimate the uncertainties inherent in the applied clustering method, two sensitivity tests have been conducted (i) to investigate how various data scaling procedures could affect the K-means classification and (ii) to compare K-means with another unsupervised classification algorithm (HAC, i.e. hierarchical agglomerative clustering). The results show that the standardization based on sample mean and standard deviation is the most appropriate standardization method for this study, as it keeps the underlying distribution of the raw data set and retains the information of outliers. The two clustering algorithms provide similar classification results, supporting the robustness of our conclusions. The classification procedures presented in this study have a markedly wide application potential for future model-based aerosol studies.