Abstract: The assumption underlying all clustering analyses of vegetation community data is that the plants and the related environmental quantities are distributed either discontinuously or with very steep gradients. In particular, the direct clustering under the theory of fuzzy sets starts with the definition of a fuzzy relation matrix according to attributes of the units to be clustered, and then make a number of cuts into the fuzzy relation to construct a hierarchical clustering system. Therefore the selection of appropriate cutting levels is decisive to the final clustering result for given fuzzy relation. The determination of cutting levels up to date is more or less a subjective or trial-error-trial process. Hence certain subjectivity or arbitrariness is inevitably involved in the final clustering result. The author think that the cutting levels should be determined according to the data being analyzed to reflect the structural properties of the data set. Specifically the cutting level should be set in those places where the greatest variations of fuzzy relation are found. This principle has been implemented in a general-purpose software with sample analysis showing the expected results.