How can different species distribution models be combined?Submitted by editor on 5 August 2015. Get the paper!
Favourability models combining different factors in mainland Spain of four representative species: Chioglossa lusitanica, Iberolacerta cyreni, Pterocles orientalis, and Galemys pyrenaicus (favourability ranges from 0 to 1). Five alternative methodological approaches for factor integration were applied: Bayesian, Akaike weight (AICw) averaging, stepwise-procedure, updating, and fuzzy intersection. The spatial resolution is 10 × 10-km.
by David Romero, Jesús Olivero, José Carlos Brito and Raimundo Real
In the context of the current biodiversity crisis, the distribution of a species, especially if threatened, must be preserved in order to conserve it. In order to know the processes that affect a species range, biogeographers are as interested in the individual role of the factors driving the species distribution as in the combined effect of several factors. However, methodological questions continue to arise regarding the way in which to combine different factors into comprehensive explanatory distribution models. Our aim was to compare the different methods available for combining species distribution models.
We used five approaches for model combination: Bayesian integration, Akaike weight averaging, stepwise variable selection, updating, and fuzzy logic. We demonstrated that different approaches to model combination give rise to disparities in the model outputs. Our conclusions were that Bayesian integration and the Akaike weight averaging should not be used unless their mathematical foundation is revised; the stepwise and updating approaches are recalibration methods that produce similar models useful if counterbalance between factors is permitted; and fuzzy logic is better when combining limiting factors that cannot be counterbalanced by more favourable factors.
Comparative assessments of five methodological integration approaches, and of three partial models according to the five indices used: sensitivity, specificity, correct classification rate (CCR), and Cohen’s kappa (all calculated using the favourability value 0.5 as the classification threshold), and the area under the curve of the receiver operating characteristic (AUC). Bars represent average values for the integrated models for the 19 vertebrate species (see the species list in Table 1) and 95% confidence intervals are represented above each bar. Under-predicted values were multiplied by ten for reasons of clarity.