Source apportionment by Receptor modeling
Multivariate statistical techniques are extensively used to apportion atmospheric aerosol sources on the basis of the internal correlations of observational data collected at a measurement point, called “receptor site”. The so-called Receptor modeling by this techniques, such as Factor analysis and Clustering, does not need any detailed a prior knowledge of source profiles and it is therefore very useful also for the determination of secondary origin aerosol fractions.
Recent developments pointed to be especially applicable for working with environmental data forcing all the values in the solutions to be non-negative, which is more realistic and meaningful from a physical point of view.
Different statistical techniques are in this way commonly applied to very different chemical-physical ambient datasets coming from both on-line and off-line measurements:
- Cluster analysis, by canonical Hierarchical Clustering and other techniques (such as K-means algorithms);
- Principal Component Analysis (PCA);
- Positive Matrix Factorization (PMF) or the multi-linear engine (ME-2) algorithm implemented both in the US-EPA open-source software (EPA-PMF v5.0 and previous ones) and in the toolkit SoFi, developed by Canonaco et al. (2013) at the Paul Scherrer Institute;
- Non-Negative Factor Analysis (N-NMF), employing a projected gradient bound-constrained optimization (Lin, 2007), or a multiplicative update approach (Lee and Seung, 2001);
- Multivariate Curve Resolution (MCR), according to two different algorithms: the classical alternating least square approach (MCR-ALS, Jaumot et al., 2005; Tauler 1995) and a weighted alternating least square method (MCR-WALS, Wentzell et al., 2006);