Statistical modelling of roadside and urban background ultrafine and accumulation mode particle number concentrations...

Quantification of the exposure of urban residents to ultrafine particle number concentrations (UFP) is challenging due to its high spatial and temporal variability. Hence, statistical models, e.g. generalized additive models (GAM), may be used to estimate time series or spatial characteristics of UFP. The GAM approach allows the representation of non-linear relations of a response variable with explanatory variables without the need to pre-define model functions. Up to now, GAMs were usually fitted to UFP data from a single site or from mobile measurement campaigns with limited temporal coverage. In this study, GAMs were used to determine UFP, accumulation mode particle (ACC) and total number concentration (TNC) at five urban sites in the cities of Leipzig and Dresden, Germany for the period 2011-2013. As explanatory variables, reanalysis data sets of meteorological quantities, urban geometry and traffic volume data were evaluated. Variables causing concurvity, which is the equivalent to collinearity in non-linear model approaches, were neglected to guarantee the interpretability of the final models. The models were then validated in a ten-fold cross-validation approach. The final models contained smooth functions for the building surface fraction, planetary boundary layer height, traffic volume, air temperature, wind direction, atmospheric pressure, relative humidity, global radiation and precipitation. Adjusted coefficients of determination (R²_adj.) for the final models were R²_adj. = 0.44 for UFP, R²_adj. = 0.51 for ACC and R²_adj. = 0.48 for TNC. Coefficients of determination of the cross-validation were in a similar range (0.44 for UFP, 0.51 for ACC, 0.49 for TNC). Finally, our study shows that GAMs are able to represent important processes that contribute to the particle number concentration from the smooth functions, i.e. emission, dilution, nucleation, deposition and long-range transport.