Optimizing the classification of species composition data by combining multiple objective evaluators toward selecting the best method and optimum number of clusters

摘要

Aims: Classification is an appropriate tool for the summarizing of species data in community ecology. Researchers need to select the effective classification method(s) and the optimum number of clusters to perform a reasonable classification. The aims of the present research are to assess the efficacy of various classification algorithms and to select the optimum number of clusters. Study area: We used a dataset of 197 400 m(2) releves recorded from Tarbiat Modares University research forest located in the north of Iran. Methods: For each releve, a species list and the canopy cover were recorded by using Braun-Blanquet cover-abundance scale modified by van der Maarel. We considered seven classification methods: flexible-beta linkage (beta = -0.25), Ward's linkage, complete linkage, average linkage, Modified TWINSPAN, k-means, and PAM. Using each of these algorithms, data were classified into 2-21 cluster levels. Then, values of eight internal evaluators viz. ASW, 1-C.index, PARTANA, PBC, 1-ISA.pval, ISA.sig.inds, ISAMIC, 1-Morisita as well as mean lambda index were calculated for each classification level resulted from algorithms. These values were applied in three methods to select the appropriate classification algorithm(s). Also, we used those values to choose the optimum number of clusters in the selected algorithm(s). A discriminate analysis opted for the verification of the selected optimums. Results: Our results revealed that, for our data, flexible-beta linkage was the proper classification algorithm with 12 the optimum number of clusters. Despite the vast number of available classification algorithms, there is no ultimate best one for all vegetation datasets. Therefore scientists need using multiple criteria to choose their specific appropriate method. With respect to this, our methods and findings could provide a generalized framework for choosing the effective method(s) for the subsequent classification analyses.