A uniform classification of aerosol signature size distributions based on regression-guided and observational cluster analysis

Fine and ultrafine particles affect human health, especially in urban areas with a large number of emission sources and significant particle number concentrations. The mixture of different aerosol sources, their temporal evolution over the diurnal course and the impact of naturally occurring processes, e.g. particle nucleation, all contribute to the spatio-temporal exposure variability across a city and its suburban areas. A differentiation between different aerosol source types is of importance for the assessment of urban aerosol pollution and exposure in a city at a certain point in time. Therefore three years of particle number size distribution (NSD) measurements from three European cities were analysed by using cluster analysis in order to identify characteristic size distributions (signatures) which are associated to specific meteorological and spatio-temporal properties. A ‘triple-site’ approach was established with data from roadside, urban background and rural sites measured over a diameter size range of 8 < Dp < 700 nm. After applying two different clustering approaches – a ‘regression guided’ and an ‘observational’ K-Means clustering – a number of 7 clusters were best representing the occurring individual aerosol “signature types”: two ‘low pollution’ clusters that showed the highest relative occurrence of around 40% on average, an ‘aged combustion’ type (∼20%), a ‘traffic’ (∼13%), a ‘remote transport’ cluster (∼10%), a cluster consisting of various unspecified local emission sources (∼10%) and a ‘particle formation’ cluster (∼7%). The cluster size distributions were distinguishable by their total number concentration, peak mode diameters and shape of the size distribution. A specific signature was assigned to each of these cluster types based on the temporal and seasonal occurrence and the prevailing meteorological conditions. The ‘regression guided’ approach was able to slightly better differentiate between homogeneous data whereas the weakness mainly lies in the time-intensive data preparation.