|dc.description.abstract||Invasive species can cause a wide range of damages from destruction of indigenous and productive ecosystems to introduction of vectors to human and animal diseases. In many countries, measures taken to prevent the establishment of invasive species are known to significantly reduce the potential damage that might be caused. As part of those measures, species distribution models (SDMs) are used to predict suitable habitats for highly invasive species so that appropriate strategies to prevent their establishment and further spread can be designed. When species distribution models (SDMs) are used for practical applications, accounting for their uncertainty becomes a priority. However, despite their wide use, reporting the uncertainty of SDM predictions is not well practiced.
The primary aim of the research in this thesis was to identify and quantify uncertainty associated with model predictions of species distributions. The major research question was, why do different models give dissimilar predictions for the same species and/or location? Discrepancy among model results is one of the major issues that affects the perception of their reliability and their capacity to inform policy decisions. In this thesis, the effect of factors considered to influence model performance and drive uncertainty in model predictions, was investigated. The particular factors were, 1) pseudo absence selection, 2) the individual and combined effect of predictor data, dimension reduction methods, and model types on model performance, and, 3) variation within the occurrence data for a given species. Following these investigations, improved procedures developed in this research were used to, 1) investigate the use of a simple mechanistic model to enhance results of correlative species distribution models in a hybrid approach and 2) improve a dispersal model that can be used to research the potential spread of an invasive species once it has established in a new habitat.
A multi-factor study to investigate the effect of pseudo-absence selection on model performance showed that not only pseudo-absences affect individual models but also consensus among model predictions. To improve individual model performance as well as model consensus, an improved pseudo-absence selection method was developed that balances the geographic and environmental space for selecting pseudo-absences.
The investigation of the individual and combined effect of predictor data, dimension reduction methods and model types on model performance, showed that the type of model is a major factor that affects model performance. The results of this research showed that the combination of appropriate explanatory variables and dimension reduction could increase individual model performance as well as model consensus. Additionally, novel indices that can be used to assess internal characteristics of the environmental predictors and data-pre-processing methods for optimized model performance, were developed.
Another important factor that contributes to model uncertainty is the reliability of the species occurrence data. While the precision of geographical references used for such data and its effect on model predictions and associated uncertainty, has been well studied, however, variation within the occurrence or presence data for a given species has been less investigated. Two case-studies were used to determine the effects of local adaptation within a species, on model predictions. It was found that apparent local adaptations resulting in ecotypes within a species could affect model predictions. As a result, methods are proposed to detect the effect of within presence data variation and an appropriate method to model potential distributions of species with such variable data, is illustrated.
Following improved procedures proposed in this research, the use of a simple mechanistic model to enhance results from correlative species distribution models was investigated. While a well parameterised mechanistic model for species distribution modelling is the ideal, such models need detailed biological data that are most often not available, especially for many invasive insects. In this study, a simple generalized mechanistic model was used to complement correlative distribution predictions. The resulting predictions from the hybrid model were shown to facilitate the identification of under- or over-predicted areas by correlative models such that its use resulted in improved overall prediction.
The enhanced protocols developed in this thesis were finally used to improve a dispersal model that can be used to project the spread of an invasive species once it has established in a new habitat by the integration of multiple scale suitability layers to represent a realistic landscape over which the dispersal of a given species can be studied. Selective landscape recoding was used to customize the landscape based on specific species-landscape interactions, to improve dispersal rate estimation and dispersal pattern determination.
This thesis presents novel methods that can be implemented to significantly increase model consensus for species distribution predictions. More important, however, the research highlights the need for implementing multi-model and multi-scenario modelling frameworks to reduce model uncertainty that can result from inappropriate use of modelling components. The findings in this thesis form the basis for research aimed at further improvement of species distribution models to provide more reliable tools for applications in invasive species management, biodiversity protection, environmental sustainability and climate change management.||en