INTEGRATING MACHINE-LEARNING WITH CLASSICAL WILDLIFE RESOURCE SELECTION ANALYSES

Kevin T Shoemaker; University of Nevada, Reno; kshoemaker@cabnr.unr.edu; Levi J. Heffelfinger, Nathan J. Jackson, Marcus E. Blum, Tony Wasley, Kelley M. Stewart

Resource selection functions (RSFs) are tremendously valuable tools for ecologists and resource managers because they quantify spatial patterns in use of resources by wildlife, thereby facilitating identification of critical habitat areas and characterizing habitat features that are selected or avoided. RSFs discriminate between used and available resource units based on an array of environmental features and are generally performed using logistic regression. However, logistic regression has some notable limitations, such as difficulties accommodating non-linear relationships and complex interactions. Increasingly, ecologists are turning to flexible machine-learning methods to overcome these limitations. We investigated the seasonal resource selection patterns of mule deer (Odocoileus hemionus) by pairing a logistic regression analysis with random forest (RF), a popular machine-learning algorithm. RF models detected strong non-linear relationships and complex, non-linear interactions that would otherwise have been difficult to discover and characterize. RF models exhibited improved predictive skill versus linear models and provided unique insights about resource selection patterns. We recommend that researchers harness the strengths of machine learning tools in addition to "classical" tools like logistic regression for evaluating wildlife resource selection patterns.

Wildlife Techniques and Technologies