Using Bayesian networks to map winter habitat for mountain goats in coastal British Columbia, Canada

The mountain goat (Oreamnos americanus) is an iconic wildlife species of western North America that inhabits steep and largely inaccessible terrain in remote areas. They are at risk from human disturbance, genetic isolation, climate change, and a variety of other stressors. Managing populations is challenging and mountain goats are particularly difficult and expensive to inventory. As a result, biologists often rely on models to estimate the species' abundance and distribution in remote areas. We used landscape characteristics evident at point locations of mountain goat visual observations, tracks, and telemetry locations, along with random locations, to learn the structure and parameters of a Bayesian network that predicted the suitability of habitats for mountain goats. We then used the model to map habitat suitability across 285,000 km(2) of potential habitat in mountain ranges of the south and central Canadian Pacific coast. Steep slopes, forest cover characteristics, and snow depth were the important drivers. Modeling the system as a Bayesian network provided several advantages over more common regression methods because input variables were heterogenous (i.e., a mix of discrete and continuous), autocorrelated, and animals exhibited non-linear responses to landscape conditions. These common characteristics of ecological data routinely violate the assumptions of parametric linear models, which are commonly used to map habitat suitability from animal observations.