Three way k-fold cross-validation of resource selection functions

A resource selection function (RSF) yields a prediction that is proportional to the probability of use of a resource unit by an organism. Because many apparently adequate models fail in new areas or time periods we developed a method for model selection and evaluation based on the model's ability to predict generally, spatially, and temporally. This work is an extension of previous work using k-fold cross-validation to evaluate models developed using presence-only study designs. A RSF model's utility is its ability to predict, so this method is applicable to any RSF model regardless of study design. The use and application of our proposed 3-way evaluation using the RSF Plot Index (RPI) statistic is illustrated using survey data of grassland birds, Landsat imagery, soil data, and a Digital Elevation Model from the Canadian Forces Base Suffield in southeastern Alberta. The sensitivity of the RPI statistic to the number and placement of bins is addressed and a method is presented to ameliorate this problem. The 3-way method provides the means to not only select the model with the best predictive power, but to understand the limitations of all models under consideration. Test results of best models using an independent field season are presented.