A Rapid Classification and Ordination Method and Its Application to Vegetation Mapping

In a previous communication (Crawford & Wishart 1967) a rapid multivariate method was described for the classification of ecological data by a monothetic divisive process. The method differed from standard numerical taxonomy techniques in that it was designed to detect sets of quadrats in terms of groups of co-incident species and not, as is more usual, in terms of quadrat homogeneity. The stopping rule applied therefore (see Macnaughton-Smith 1965) was determined by the degree of co-incidence between species and not the attaining of a set level of similarity or absence of dissimilarity, as in most agglomerative and divisive methods (see Sokal & Sneath 1963; Williams & Dale 1965). Owing to the manner in which the species group correlations are carried out the method is very rapid even with large surveys and is relatively unaffected by the number of species in the survey. The time necessary for analysis is dependent solely on the number of samples to be analysed and increases linearly with the sample number. However, in common with all other monothetic divisive methods, no indication is obtained of the relationships between the various terminal groups; and further there is always the danger of misclassification due to the chance occurrence, or erroneous diagnosis of a dividing species. This paper describes firstly, a rapid agglomerative method that can be used after the initial divisive process to check for any misclassifications, and secondly, a means of representing the variance both within and between the terminal groups by an ordination procedure. The need in ecological surveys for such a representation of the results of classification is demonstrated by the present dichotomy in the use of ordination and classification methods. A review of the problems involved in this divergence is given by Greig-Smith (1964). The recently devised polythetic agglomerative methods of Jancey (1966) and Orloci (1967) certainly give a solution to these problems, in that the ordination of the group centroids displays the relative distances between the resulting groups. However, as developed at present these methods are difficult to use with large surveys owing to the quadratic relationship between computation time and sample number. In an attempt to overcome this difficulty the methods described in this paper achieve the economy in core storage which enables them to be used with large surveys by applying the same strategy as that used in the previous communication. The basis of this approach was to look for the occurrence of groups produced by dividing the set of quadrats on the presence or absence of each species in turn. Thus each division of the data involved only N tests, where N2 would have been necessary using existing methods. Thirdly, a method is presented for mapping vegetation that has been sampled on a grid. Maps of distribution of vegetation types usually involve the drawing of lines on paper that may or may not represent a discrete vegetation boundary in the field. In this present study the conventional mapping of several discrete vegetation types is replaced by a