Associate Professor Wright State University Dayton, OH, United States
Background/Question/Methods
Species distribution models (SDM) suggest that bird distributions are well modelled by climate and that land cover variables are of minor importance. This conclusion seems ecologically puzzling, since many bird species are known to be tied to specific habitats, but are well insulated and seemingly not terribly sensitive to climate. Evidence is accumulating that the putative good performance of climate variables in modeling species distributions may depend on their spatial structure being able to describe the distributions, rather than functionally modeling them. I previously showed that truly predicting onto independent new areas dramatically reduces the performance of SDMs. Here I investigated how the relative importance of climate variables vs. land cover variables changes with a) restricting modeling to within range locations only; b) including spatial predictors that account for spatial structure in species distributions; and c) predicting onto spatially independent test data. For data, I used 79 species from the North American Breeding Bird Survey, WorldClim climate data, and USGS land cover maps. The distribution models and statistical analyses rely on Random Forests, Boruta for variable selection, and spatially independent leave one out cross validation for model evaluation.
Results/Conclusions
In the default random forest SDM models, climate variables regularly outperformed land use variables. There was, however, considerable variation in the distribution of importance of these two categories of variables among species, which makes sense, given the wide variety in ecological niches among bird species. The variable importance was robust to the specific selection of explanatory variables included in the model. Modeling only within the range of a species on average slightly increased the relative importance of land cover vs. climate variables. Similarly, including Moran’s Eigenvector Maps to account for spatial structure in bird distributions on average increased the relative importance of land cover variables. Again, some of these shifts were idiosyncratic to species. And finally, a true evaluation of models on fully independent locations (aka transferability) on average also showed a relative increase in importance of landcover variables. All these shifts varied widely among species. The take home message is to better explain what drives species distributions and to link them with our understanding of the species’ ecology, we need to be more thoughtful in how we build and evaluate our models.
