Predicting species distributions in poorly-studied landscapes

Conservationists are increasingly relying on distribution models to predict where species are likely to occur, especially in poorly-surveyed but biodiverse areas. Modeling is challenging in these cases because locality data necessary for model formation are often scarce and spatially imprecise. To identify methods best suited to modeling in these conditions, we compared the success of three algorithms (Maxent, Mahalanobis Typicalities and Random Forests) at predicting distributions of eight bird and eight mammal species endemic to the eastern slopes of the central Andes. We selected study species to have a range of locality sample sizes representative of the data available for endemic species of this region and also that vary in their distribution characteristics. We found that for species that are known from moderate numbers (N = 38-94) of localities, the three methods performed similarly for species with restricted distributions but Maxent and Random Forests yielded better results for species with wider distributions. For species with small numbers of sample localities (N = 5-21), Maxent produced the most consistently successful results, followed by Random Forests and then Mahalanobis Typicalities. Because evaluation statistics for models derived from few localities can be suspect due to the poor spatial representation of the evaluation data, we corroborated these results with review by scientists familiar with the species in the field. Overall, Maxent appears to be the most capable method for modeling distributions of Andean bird and mammal species because of the consistency of results in varying conditions, although the other methods have strengths in certain situations.