Species distribution modelling.pdf

Abstract: Detailed and reliable information about the spatial distribution of species provides critical information for effective conservation planning. In Victoria, Australia, this responsibility is held by the Department of Sustainability and Environment, which maintains and curates a database of site records for species: the Victorian Biodiversity Atlas. But the information is provided in point form, and therefore, it does not provide an adequate view of species distributions for all management purposes. By integrating known occurrences of species with environmental GIS data layers using a machine learning algorithm, random forest, we have built species distribution models for 523 vertebrate fauna species across the whole state, providing predictive ‘maps’ that are available for use in various conservation planning activities. In this paper, we introduce and discuss the methods we developed and implemented for producing these models. Specifically, 26 explanatory variables were used for the modelling; three versions of models were built using different sets of explanatory variables; pseudo-absences were chosen by filtering random points with a profile model; model accuracy was assessed using several measures, especially lift curve-related ones; a threshold was selected for each model to transform continuous result to binary one by maximizing the sum of sensitivity and (pseudo) specificity, which was proved to be valid for presence-only data.