The application of tree-based methods to species distribution modelling

Abstract

Species distribution models are used increasingly in both applied and theoretical research to predict how species are distributed and to understand attributes of species’ environmental requirements. This thesis aims to explore the application of tree-based methods to species distribution modelling. Although these methods have been widely used in other fields of science they have received relatively little exposure in Biogeography and Conservation Biology. The techniques applied include CART, Bagging, Random Forests and Boosted Regression Trees. These were used with four different biodiversity databases to answer different a variety of research questions aimed at: (i) understanding how landscape structure and climate affect species distributions (ii) predicting the potential impacts of climate change on species distributions (iii) to identify areas important for biodiversity conservation. Additionally, the performance of each method was compared with the aim (iv) of making suggestions for the optimal models which should be used by future researchers. In chapter 2 Boosted Regression Trees were used to quantify the importance of wetland size and weather patterns for waterbird distribution in Britain. As well as revealing the importance of wetland size for waterbirds, , the models proved to be reasonably robust when validated. In chapter 3 this basic form of modelling was expanded, using a database containing amphibian occurrence records for Italy. Random Forests was used to quantify species-climate relationship and to predict amphibian distribution in relation to current and future climate conditions. The results revealed how amphibian distribution is largely controlled by temperature-related variables and highlighted a negative response to future climate changes in most species. In chapter 4 Bagging was used to identify areas important for biodiversity conservation. Specifically, Bagging was used predict the distribution of 232 species of Butterflies in Italy. The predicted surfaces were then used in combination with a species multispecies prioritization tool in order to identify important areas for butterfly conservation. The results iii showed that the most areas important for butterfly are located within the Alps, the mountains of central Italy and the island of Sardinia. Finally, in Chapter 5, the predictive accuracy of four modelling techniques based classification trees was compared. This was done using large scale bird distribution data from Italian Common Bird Census. The results showed that Random Forests and Boosted Regression Trees were the best performing techniques and that model performance was highly influenced by species ecological characteristics as well as by the modelling method. The results of this thesis have shown how tree-based modelling methods can be used for exploring and testing hypotheses about the factors that are important in determining species distribution and making predictions of species distribution for use in conservation contexts. The methods used represent a useful way to visualize and understand relationships between environmental parameters and species distributions and to predict species distributions with high accuracy. Whilst it is true that some tree-based methods can be used instead of statistical modelling techniques others expand the analytical opportunities by enabling analyses that are impossible or very difficult with statistical methods. Hopefully this thesis will serve a source of inspiration for ecologists willing to move away from statistical inference and the P-value dogma and concentrate on understanding the data, and using alternative techniques to predict species distribution with high accuracy.