干旱区生态环境知识资源中心

This paper presents a new clustering program named RCLUS that was developed for species (R-mode) analysis of plant community data. RCLUS identifies clusters of co-occurring species that meet a user-specified cutoff level of positive association with each other. The "strict affinity" clustering algorithm in RCLUS builds clusters of species whose pairwise associations all exceed the cutoff level, whereas the "coalition" clustering algorithm only requires that the mean pairwise association of the cluster exceeds the cutoff level. Both algorithms allow species to belong to multiple clusters, thus accommodating both generalist and specialist species. Using a 60-plot dataset of perennial plants occurring on the Beaver Dam Slope in southwestern Utah, we carried out RCLUS analyses and compared the results with 2 widely used clustering techniques: UPGMA and PAM. We found that many of the RCLUS clusters were subsets of the UPGMA and PAM clusters, although novel species combinations were also generated by RCLUS. An advantage of RCLUS over these methods is its ability to exclude species that are poorly represented in a dataset as well as species lacking strong association patterns. The RCLUS program also includes modules that assess the affinity of a given species, plot, or environmental variable to a given cluster. We found statistically significant correlations between some of the RCLUS species clusters and certain environmental variables of the study area (elevation and topographical position). We also noted differences in clustering behavior when different association coefficients were used in RCLUS and found that those incorporating joint absences (e.g., the phi coefficient) produced more clusters and more even numbers of species per cluster than those not incorporating joint absences (e.g., the Jaccard index). In addition to the species association application described in this paper, the RCLUS algorithms could be used for preliminary data stratification in sample (Q-mode) analysis. The indirect link between sample plots and RCLUS species clusters could also be exploited to yield a form of "fuzzy" classification of plots or to characterize species pools of plots.