Current approaches to modelling the environmental niche of eucalypts: implication for management of forest biodiversity
M.P. Austin *, J.A. Meyers
C.S.I.R.O. Division of Wildlife and Ecology, P.O. Box 84. Lyneham, A.C.T. 2602, Australia
Robust predictive models of the distribution of forest biota are important tools for the management of forest biodiversity. To build robust models, it is essential to understand the environmental processes which control species distribution and hence choose appropriate predictor variables. Requirements for modelling the environmental niche of plant species include: environmentally stratified survey data of the vegetation and associated environmental measurements, an understanding of ecological theory, robust statistical models and geographical representation of the models. These requirements can be satisfied in different ways, many of which are discussed. The choice of modelling technique and curve fitting function should be related to ecological theory. Prediction becomes increasingly robust and less location-specific as the predictor variables become more process-oriented and relevant to biological processes. However, the need to use predictors for which estimates are available for unsampled regions may limit the choice to less direct variables. In this context we examine the performance of two modelling techniques: Generalised Linear Modelling (GLM) and Generalised Additive Modelling (GAM). Trees are ideal to study, because their size and immobility make for ease of collecting data and they provide important habitat for fauna and understorey herbs and as such are useful for predicting the distribution of some other biota. The data set includes 8377 sites in south-eastern Australia, with presence/absence data for trees and seven environmental predictors. A detailed comparison is described for Eucalyptus cypellocarpa. The influence of ‘naughty noughts’, or zero values beyond the range of a species, can distort the response function, giving positive predictions where the species is known to be absent. The model is improved by restricting the data to a suitable range. GAM has advantages over GLM due to the flexible nature of the non-parametric smoothing function. Different response curves can produce divergent predictions of species occurrence, particularly at the limits of species distribution. Conservation evaluation often requires making predictions in unsampled areas, and so the assumption of particular shapes of response curve could lead to significant errors in the estimation of the conservation value of these areas.