A comparative evaluation of techniques for ecological ordination using simulated vegetation data and an integrated ordination-classification analysis of the Alpine and Subalpine Plant communitites of the Mt. Field Plateau, Tasmania.

This thesis consists of two, inter-related sections. In Part 1, the relative utility of ordination techniques, for the purpose of indirect gradient analysis in plant ecology, is studied empirically. Using simulated data derived from flexible models of vegetation-gradient relationships. Part 2 is an integrated analysis of vegetation data obtained by intensive sampling of the subalpine and alpine vegetation of the Mt. Field Plateau, Tasmania. The results of the direct gradient analysis of the Mt. Field vegetation were taken into account when formulating the models used in Part 1. In return, the selection of the most appropriate methodo1ogy for indirect gradient analysis of the Mt. Field data drew upon the results of the comparative evaluation of techniques in Part 1. Previous studies which have evaluated the performance of ordination techniques using simulated vegetation data are reviewed. Most such studies have employed models in which the response profiles of species along the artificial gradients were Gaussian. Some preliminary experiments with plausible alternative models have suggested that even the most promising of recently introduced ordination methods may be undesirably sensitive to small variations in the model. Since the available observational evidence does not support the general applicability of the Gaussian model in nature, the relevance of the results of comparative studies based on Gaussian models is in doubt. Noting that present evidence is insufficient to permit the formulation of a general model of vegetational response to environmental gradients, the present study attempts to identify ordination methods which are robust to variations in those features of the model still subject to debate. A flexible modelling approach was developed, based on the use of beta functions, which can produce unimodal curves of varying skewness. The models allow \competition\" between species to introduced in the form of linear interaction coefficients thus producing shouldered bimodal or multimodal response curves of the types often observed in direct gradient analyses. Both stand abundance (the sum of the abundances of all species present in a sample) and alpha diversity (the number of species per sample) may be varied along gradients in a flexible manner. Quantitative noise may be introduced in the form of random departures from the expected abundance. Probability of occurrence profiles may optionally be used to introduce qualitative noise. The models allow both unidimensional compositional gradients (coenoclines) and two-dimensional patterns (coenoplanes) to be produced. Using such models the relative performance of the ordination techniques Detrended Correspondence Analysis (DCA) \"local\" Non-metric Multidimensional Scaling (LNMDS) Principal Co-ordinates Analysis (PCoA) Principal Components Analysis (PCA) and Gaussian Ordination (GO) was assessed. Ordination configurations were rotated to best fit with the original arrangement of samples along the simulated gradient(s) using Procrustean ana1ysis and the RMS displacement error was used as a measure of ordination efficacy. For ordinations of coenoclines the Kendall rank correlation coefficient was used to assess the accuracy with which the rank order of samples was recovered. Experiments were performed on the effects of variation in several properties of the models with particular attention being devoted to the robustness of methods to variations in the shape of response curves the introduction of noise and the arrangement of samples. Experiments with coenocline models confirmed the ineffectiveness of PCA except when the beta diversity of the gradient is low (no more than 4Z). PCoA operating on a percentage similarity matrix was similarly affected. With models containing only unimodal curves DCA was generally the most successful of the methods compared. GO was able to improve significantly on the DCA solutions only when beta diversity was less than 6Z quantitative noise levels were low and the curves were not grossly skewed. When models included shouldered and bimodal curves LNMDS was consistently superior to DCA and GO. PCA and PCoA ordinations of simu1ated coenoplanes were poor unless the beta diversity of both gradients was low. With 11 rectangu1ar 11 coenoplanes curvilinear distortion of the longer gradient often obscured variation related to the secondary gradient. LNMDS consistently improved on DCA solutions for all the coenoplane models examined. The DCA results even with noiseless data from some of the symmetrical unimodal models were occasionally very poor. Neither LNMDS nor DCA gave consistently satisfactory results when coenoplanes were sampled with cross-shaped or T-shaped sampling patterns which simulate field situations where one gradient is expressed only in a restricted region of another gradient. The present study suggests a preference for LNMDS as a \"general purpose\" approach to indirect gradient analysis. However practical limitations of LNMDS may make DCA the method of choice when the number of samples is large. Significant advances in the methodology of indirect gradient analysis are unlikely to be made until much more observational evidence concerning the nature of vegetational response to environmental gradients is available. Subalpine and alpine vegetation on the Mt. Field Plateau a dolerite capped horst in south-central Tasmania was intensively sampled using a stratified systematic arrangement of 438 100 sq. m quadrats. In each sample cover estimates were made for all vascular plant species and various site and community characteristics were recorded. A direct gradient analysis was performed using altitude and substrate drainage (estimated on a five point scale) as axes. The response surfaces for mean cover of 111 species were predominantly unimodal (80%) although only 45% appeared to be symmetrical. For the unimodal species the distribution of modes over the ecoplane was consistent with random expectation. The frequency distribution of response surface maxima was approximately 1ograndom for non-woody species but the distribution for woody species did not fit lograndom or lognormal hypotheses. Mean values of tallest stratum total commun1ty characteristics cover and alpha diversity such as height displayed systematic trends in re1ation to altitude and drainage. The pattern of alpha diversity was further clarified by the division of species into growth-form groups. Beta (between-habitat) diversity along the drainage gradient showed a consistent decline with increasing elevation. The potential value of Generalised Linear Modelling (GLM) as a more rigorous approach to the fitting of response surfaces was examined for a selected set of 18 dominant species. An initial qualitative model was fitted in order to identify samples where the expected probability of occurrence of a species was low. Such samples were then omitted when fitting a predictive model for percentage cover. For all but four of the species probability of occurrence was adequately fitted by a symmetrical bell-shaped surface. The quantitative (percentage cover) models for most species had low adjusted r-squared values indicating a relatively large amount of 11 noise\" variation. The factors which may have contributed to this noise variation are discussed. Because of the large number of samples~ which made LNMDS impracticable DCA was adopted as a primary approach for indirect gradient analysis. The first two axes of a DCA ordination based on cover class data defined a coenop1ane to which altitude and drainage were strongly related. This result supports the choice of these two complex-gradients as axes for direct gradient analysis. The third axis was a spurious \"interaction~~ axis resulting from the failure in DCA to detrend the third and subsequent axes with respect to combinations of the previously extracted axes. The fourth axis represented a compositional gradient among subalpine forests on well drained sites which was tentatively related to firing history. DCA based on presence-absence data gave very similar results~ indicating that at these levels of beta diversity most of the relevant information concerning the positions of samples on the major underlying gradients is contained in the qua1itative component of the data. A DCA ordination of percentage cover data (i.e. midpoints of cover classes) produced a distorted reconstruction of the altitude X drainage coenoplane with one corner projected into a third dimension. The origin of this type of distortion which was also noted with some of the coenoplane models in Part 1 is not clear. A compound data set was produced by forming the centroids of 60 minimum Sum-of-Squares clusters. Ordinations of this compound data set by various techniques including DCA local and global NMOS PCoA and PCA were compared. The ineffectiveness of PCoA and PCA was readily apparent. Curvilinear distortion of the altitude X drainage coenoplane prevented adequate recovery of the tertiary \"firing history\" gradient. The NMDS solutions based on local and global criteria were very similar to each other and they differed from the DCA result only in local details"