Predicting koala (Phascolarctos cinereus) distribution from incidental sighting data in South-East Queensland, Australia

Species distribution maps are important tools for wildlife conservation planning and management. To model koala distributions, usually, a spatially representative sample of koala populations is collected through systematic field surveys. Details of koala sightings collected by members of the public could potentially be used to develop species distribution models if appropriate analytical approaches are applied to address the inherent biases in such datasets. We developed a stepwise approach for applying bias correction techniques to estimate and map koala distributions. Using a Boosted regression tree approach, we modelled indirectly the search effort made by observers to identify or sight koalas. Land lot density (58%) and human population density (19%) had the strongest positive impact on the indirect search effort, while the distances to roads were negatively associated with the indirect search effort. To estimate the koala distribution across South-East Queensland, we then developed models describing koala habitat (environmental model), access to koala habitat (accessibility model) and the search effort (search effort model), with the latter including the indirect search effort covariate. Finally, we corrected the estimates derived from these models (bias-corrected search effort and accessibility model). Three independent statistical modelling approaches (Lasso penalty Poisson regression, Down-weighted Poisson regression, and Maximum entropy) were used to compare the five koala distribution models. Based on assessments of areas under curves, the predictive accuracy of models improved when area accessibility and search effort were included. Overall, the spatial extent of koala distributions increased in the prediction maps when models were corrected for accessibility and indirect search effort (except for Down-weighted Poisson regression)