MULTIPLE-SITE RECEPTOR MODELING WITH A MINIMAL SPANNING TREE COMBINED WITH A NEURAL-NETWORK

A combination of two pattern recognition methods has been developed that allows the generation of geographical emission maps from multivariate environmental data. In such a projection into a visually interpretable subspace by a Kohonen self-organizing feature map, the topology of the higher dimensional variables space can be preserved, but parts of the information about the correct neighborhood among the sample vectors will be lost. This loss can partly be compensated for by an additional projection of Prim's minimal spanning tree into the trained neural network. This new environmental receptor modeling technique has been adapted for multiple sampling sites. The behavior of the method has been studied using simulated data. Subsequently, the method has been applied to mapping data sets from the Southern California Air Quality Study (SCAQS). The projection of 17 chemical variables measured at up to eight sampling sites provided a two-dimensional, visually interpretable, geographically reasonable arrangement of air pollution sources in the South Coast Air Basin.