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

[摘要] The K-means machine learning algorithm is applied toclimatological data of seven aerosol properties from a global aerosolsimulation using EMAC-MADE3. The aim is to partition the aerosol propertiesacross the global atmosphere in specific aerosol regimes; this is done mainly forevaluation purposes. K-means is an unsupervised machine learning method withthe advantage that an a priori definition of the aerosol classes is notrequired. Using K-means, we are able to quantitatively define global aerosolregimes, so-called aerosol clusters, and explain their internal propertiesand their location and extension. This analysis shows that aerosolregimes in the lower troposphere are strongly influenced by emissions. Keydrivers of the clusters' internal properties and spatial distribution are,for instance, pollutants from biomass burning and biogenic sources, mineraldust, anthropogenic pollution, and corresponding mixtures. Several continentalclusters propagate into oceanic regions as a result of long-range transportof air masses. The identified oceanic regimes show a higher degree ofpollution in the Northern Hemisphere than over the southern oceans. Withincreasing altitude, the aerosol regimes propagate from emission-inducedclusters in the lower troposphere to roughly zonally distributed regimes inthe middle troposphere and in the tropopause region. Notably, three pollutedclusters identified over Africa, India, and eastern China cover the wholeatmospheric column from the lower troposphere to the tropopause region. Theresults of this analysis need to be interpreted taking the limitations andstrengths of global aerosol models into consideration. On the one hand,global aerosol simulations cannot estimate small-scale and localizedprocesses due to the coarse resolution. On the other hand, they capture thespatial pattern of aerosol properties on the global scale, implying that theclustering results could provide useful insights for aerosol research. Toestimate the uncertainties inherent in the applied clustering method, twosensitivity tests have been conducted (i) to investigate how various datascaling procedures could affect the K-means classification and (ii) tocompare K-means with another unsupervised classification algorithm (HAC,i.e. hierarchical agglomerative clustering). The results show that thestandardization based on sample mean and standard deviation is the mostappropriate standardization method for this study, as it keeps the underlyingdistribution of the raw data set and retains the information of outliers. Thetwo clustering algorithms provide similar classification results, supportingthe robustness of our conclusions. The classification procedures presentedin this study have a markedly wide application potential for futuremodel-based aerosol studies.

[学科分类] 土木及结构工程学