Surveying Southern Mongolia: application of multivariate classification methods in drylands with low diversity and long floristic gradients

Abstract

AbstractQuestion: How do standard multivariate classification algorithms perform for Mongolian dryland vegetation characterized by low diversity and very long floristic gradients?Location: Southern Mongolian Gobi.Methods: We compared several widely used classification methods based on 1231 relevés obtained with a similar sampling method. We applied agglomerative cluster algorithms (flexible beta/FLEX, average linkage/UPGMA, weighted linkage/WPGMA, complete linkage, Ward's clustering) and a divisive classification technique (TWINSPAN); data were reduced to presence/absence. We compared results against a published phytosociological classification (PHYTO), against environmental background data, and with respect to the presence of significant indicator species.Results: Complete linkage was inferior to other methods. TWINSPAN, UPGMA, flexible beta and WPGMA gave partly similar clusters, with FLEX and WARD showing the highest pair‐wise similarity. Classifications of all methods except CL partly agreed with PHYTO classification. Clusters of all methods had significant indicator species, but Ward's method had the highest number of indicator species, followed by the PHYTO classification and FLEX, TWINSPAN, UPGMA and WPGMA. The latter four methods all yielded clusters that differed in terms of precipitation, but TWINSPAN, FLEX and Ward's method performed best under this criterion. PHYTO and CL ranked last in partitioning the precipitation gradient. Comparisons with ordinations indicated that classification algorithms capture the main floristic gradient but were less successful than the phytosociological approach to elucidate the finer structures.Conclusion: Performance of classification methods differed depending on the applied validation approach and we thus caution against uncritically adopting a single evaluation/validation criterion. Most numerical approaches can aid sorting of large data sets, while details of manual syntaxonomic classifications are not easily reproduced. Choice of the most appropriate classification and validation method thus clearly depends on the overall aim of a given study.