Combining Kohonen neural networks and variable selection by classification trees to cluster road soil samples

Kohonen neural networks, or Self-Organizing Maps (SOMs), were used to study the data sets generated in a survey of soil pollution along a four-season study. Each sampling season comprised 89 road soil samples and 12 analytical variables; namely, nine heavy metals (Cd, Co, Cu, Cr, Fe, Mn, Ni, Pb, and Zn) and three physicochemical parameters (loss on ignition, pH and humidity). The SOMs provided a rapid and intuitive means to recogniz3e four different groups of samples: roadside of a highway, highway transects, roadside of a main avenue and urban gardens. They became defined essentially by the physical characteristics of the sampling sites and by the intensity of the road traffic. In order to simplify the chemical understanding of the patterns defining the different groups of samples, to avoid noisy and/or redundant variables and to reduce the time required to develop a suitable SOM, the usefulness of a previous variable selection step using CART, Classification and Regression Trees, was investigated.