Evaluating the importance of wolverine habitat predictors using a machine learning method

Abstract

In the conterminous United States, wolverines (Gulo gulo) occupy semi-isolated patches of subalpine habitats at naturally low densities. Determining how to model wolverine habitat, particularly across multiple scales, can contribute greatly to wolverine conservation efforts. We used the machine-learning algorithm random forest to determine how a novel analysis approach compared to the existing literature for future wolverine conservation efforts. We also determined how well a small suite of variables explained wolverine habitat use patterns at the second- and third-order selection scale by sex. We found that the importance of habitat covariates differed slightly by sex and selection scales. Snow water equivalent, distance to high-elevation talus, and latitude-adjusted elevation were the driving selective forces for wolverines across the Greater Yellowstone Ecosystem at both selection orders but performed better at the second order. Overall, our results indicate that wolverine habitat selection is, in large part, broadly explained by high-elevation structural features, and this confirms existing data. Our results suggest that for third-order analyses, additional fine-scale habitat data are necessary.