The review and application of new methods for species distribution modeling

The application of species distribution models ecology and evolution is rapidly increasing. Frequently applied in a geographic information system (GIS) environment, species distribution models are being used to identify areas for conservation prioritization, identify regions susceptible to invasion by exotic species, and to test whether the niches of species are conserved in evolutionary time. Here I review recent developments in the species distribution modeling literature, compare several newer algorithms and critically evaluate standard procedures for assessing predictive model accuracy. In chapter one, I compare the accuracy of three species distribution model algorithms for modeling the probability of occurrence and density of grey-headed flying foxes (Pteropus poiliocephalus) across their range in south-eastern Australia. The results reveal how increased patterns of human/bat interaction can be attributed to a lack of native foraging resources rather than a preference for human dominated landscapes. I also demonstrate that the boosted regression tree algorithm clearly outperforms generalized linear models (GLM's) and multivariate adaptive regression splines (MARS) for predicting probability of bat occurrence and bat density. In chapter 2 I review a subset of the presence-only species distribution modeling literature to identify patterns for how the scale of environmental variables and the spatial structure of occurrence records can lead to a bias in the evaluation of predictions. The review shows how many studies manipulate the scale of environmental variables used for model predictions but don't account for the original scale when designing statistical evaluation. Using simulated data, I show how this can bias evaluation of model accuracy. In chapter three, I compare two commonly used presence-only species distribution algorithms, the genetic algorithm for rule set prediction (GARP) and maximum entropy modeling (Maxent) for predicting potentially suitable areas for the invasion of an exotic weed, spotted knapweed (Centaurea maculosal). The results show that Maxent had higher predictive accuracy within the training region. There was no difference in the predictive accuracy between the two algorithms when predictions were extrapolated to a new region. Both models showed sensitivity to the spatial bias in occurrence records with corresponding reductions in predictive accuracy when predictions were evaluated with spatially independent data.