Projection of Prim's minimal spanning tree into a Kohonen neural network for identification of airborne particle sources by their multielement trace patterns

Abstract A hybrid pattern recognition method has been developed as an alternative receptor modelling technique for the identification of sources of coarse airborne particles. The Kohonen self-organizing neural network is first applied to yield a topological map of an m-dimensional variables space. Unfortunately, during the projection into a low-dimensional subspace, most of the information about the correct distance between the sample vectors is lost. However, the Kohonen network is a useful a priori step of data compression before application of the minimal spanning tree. Prim's minimal spanning tree partly compensates for this loss yielding the distance interrelationships between groups of the samples. This combination of both projection techniques can overcome some of their individual deficiencies. Several illustrative examples are presented to demonstrate the nature of the analysis results. Then a set of airborne particle compositions for samples obtained at a single sampling site were analysed. After transferring the combined map to a geographical unit circle (GUC), a correct pattern of the main industrial emission sources around a sampling site in Granite City (Illinois, USA) has been obtained by decoding a 35-dimensional space of chemical-analytical variables into a visually and geographically interpretable 2-D space.