Wildlife managers commonly assume that the removal of habitat disturbance benefits wildlife, and that wildlife-habitat models adequately predict habitat use. Both assumptions are questioned in this thesis, which uses two gecko species, Gehyra variegata and Rhynchoedura ornata, as case studies. The study was conducted at Currawinya National Park (28° 50'S, 144° 29'E) in southwest Queensland, where the removal of domestic stock in 1992/93 from apparently degraded natural systems, created a suitable context in which to examine the above assumptions.
Two approaches were used to test whether the removal of disturbance (domestic grazing pressure) benefited the target species. I compared samples of the gecko fauna from the unstocked national park and the adjoining stocked grazing properties in the four years immediately following destocking. The results indicated no significant change in target species relative abundance or gecko species richness, suggesting that geckos are relatively insensitive to the removal of grazing pressure in the short to medium term. I also constructed a series of wildlife-habitat models to predict whether the target species would be affected by destocking beyond the life of the study. No G. variegata model suggested a destocking effect over the longer term. However, some of the R. ornata models suggest that the distribution of this species on Currawinya may contract if grass coverage increases following destocking. These results suggest that the removal of habitat disturbance factors is not universally beneficial to wildlife populations.
I developed a wide array of models to test whether wildlife-habitat models adequately predict habitat use and fitted them all to independent data to assess their predictive power. The development of this range of models not only tested the predictive power of wildlife-habitat models per se. but also permitted three comparisons of individual modelling approaches. Firstly, because target species were sampled in this work at low abundances, I built models with three different assumed data structures; normal, poisson, and binomial; and examined whether poisson or binomial models fitted the low count data better than the more widely used normal error models. No single assumed data structure performed best, however, results indicate that several different structures should be trialed for low abundance species. Secondly, I developed wildlife-habitat models with both GLMs (generalized linear models) and GTMs (generalized tree models), testing the idea that tree-based models may be more applicable to wildlife-habitat data. The results indicated no difference between the two approaches, but results with larger data sets (n >100) may differ. Finally, I developed models with single and dual scales of predictors, comparing the predictive power of analogous single and dual scale models. This work tested two ideas; that dual scale models will capture habitat use better than single scale models, and that tree models are more appropriate than GLMs to dual scale modelling. The results supported neither hypothesis, but reveal the critical role of sampling scale selection in the modelling process.
Despite the range of models developed, none fitted test data significantly better than its null, casting doubt on the idea that wildlife-habitat models are useful predictive tools. These results are also reflected in the published literature, where most models are not tested with independent data, and of those that are, a substantial proportion do not predict well. I construct a meta-model of distribution (presence and abundance) based wildlife-habitat models, that explains both why it is so difficult to predict future habitat use, and how to build better models. The meta-model identifies six critical assumptions in every distribution-based wildlife-habitat model. These six assumptions are linearly related, so the validity of each assumption is equally important, and it is only possible to accurately predict future habitat use where all assumptions are valid. More importantly, the meta-model demonstrates that even if all six assumptions are validated, there is still a high likelihood that model output and habitat value for the target species will not be related and that more direct measures of habitat quality than distributional data should be modelled.