The relationship between scale and predictor variables in species distribution models applied to conservation

Abstract

Species distribution models have been used to assist decision-making in many different aspects of conservation, restoration, and environmental management. However, to apply species distribution models effectively, we need to discriminate between suitable and unsuitable environments and the models need to be developed at fine scales (i.e. covering small areas at a fine resolution). These characteristics allow more precise decision-making for heterogeneous environments in smaller areas, such as biomes. We also need to understand the potential limiting factors in relation to these models better, including the effects of sample bias in species occurrence records and the potential mismatch between the scale at which the models were built and the scale at which the predictor variables interact with species occurrence. Here we evaluate the effects of two methods used to reduce bias (geographic vs. environmental filters) and three predictor variable types (climactic, local and biotic) on model predictions. We explore these issues for the hyacinth macaw (Anodorhynchus hyacinthinus), a globally vulnerable species in the Pantanal biome of central South America. We consider broad-scale variables, local-scale habitat associations, and the interactions of the macaw with two plant species that provide its food and nesting location. Our results show that using broad-scale climate variables for local-scale models (i.e., models with a fine resolution with a small extent) can generate predictive distribution models that underpredict suitability. Using local and biotic variables generates more accurate models with predictions consistent with the known distribution of the bird species. Although not commonly used, local-scale variables strongly affect model performance by increasing accuracy, reducing omission error, and leading to more conservative predictions. On the other hand, these methods lead to variable results in relation to bias reduction, with their efficiency depending on the amount of sampling bias in the occurrence records. In conclusion, local variables and the method of bias reduction play an important role in species distribution models. Fine resolution models constructed at the local scale for small areas show the greatest skill in predicting species distribution.