A method to quantify and depict uncertainty in wildlife habitat suitability models using Bayesian inference and expert opinion

Abstract

Knowing the distribution of wildlife habitats across the landscape is an important component in biological conservation planning. Many conservation planning projects use wildlife habitat suitability models as the basis for predicting the distribution of habitat for terrestrial species. The predictions are typically binary GIS maps depicting the distribution of suitable versus unsuitable habitat, without indication of how strong the evidence is for these predictions across the area. There are many sources of uncertainty in these models as each data layer, with its own level of uncertainty, is incorporated into the models. Habitat suitability models are often knowledge-based and do not quantify their inherent uncertainty. Or, if the models are empirically-based, there are usually insufficient data to derive habitat distribution predictions and to test the predictions to determine the level of uncertainty associated with them. To make evident the uncertainty inherent in knowledge-based habitat suitability models, Bayesian inference procedures were used to combine expert opinions about the strength of wildlife habitat relationships with prior model parameters to create probability maps that depict the state of knowledge about the distribution of suitable habitat for terrestrial wildlife species. The Bayesian method has several advantages. One is that probability in a Bayesian framework is a direct representation of uncertainty. Thus models produced using this method are easy to understand and interpret. This method can be used on any species, regardless of the amount of empirical data available. Modeling species with deficient habitat relationship data produces appropriate results showing high levels of uncertainty. Bayesian methods allow the combination of empirical and knowledge-based evidence, so that all sources of information about species habitat may be incorporated. Bayesian models may also be updated, so that models can be improved as new information arises. The models can also incorporate landscape context and depict the associated uncertainty. With binary models, a priori decisions are made to include or reject specific habitat conditions. This tends to either over or under predict suitable habitat by including or rejecting borderline conditions. The portrayal of the results (habitat is suitable: yes or no) also implies a certainty that is unwarranted. With the Bayesian method, all possible habitat conditions are retained in the models, revealing areas of potentially suitable habitat that may have been omitted by binary models, and the certainty of the predictions is forthrightly depicted. The models derived by this method produce simple, honest, spatial depictions of what is known about the distribution of suitable wildlife habitat that can be used to support more informed decisions in species conservation planning and management.