Poster: Quantifying species tolerances and functional diversity using n-dimensional hypervolumes: a comparison of methods
Abstract
Multidimensional hypervolumes enable ecologists to visualize the functional trait space occupied by an ecological community. Previously, hypervolumes have been measured using a minimum convex hull, but convex hulls are exclusively determined by extreme points and they cannot account for possible holes in the trait space. A multivariate kernel density estimation method with hyperbox kernels was proposed to deal with high-dimensional or holey datasets, but this method produces unrealistically blocky hypervolumes. We examined two alternatives: a Gaussian kernel density estimation method and a support vector machine method. We tested these two new methods and the hyperbox method by creating hypervolumes for three New World biomes using trait data from plants and mammals. We varied the parameters for each method in order to determine sensitivity to parameter variation. The resulting hypervolumes were compared with respect to their total volume, shape, and overlap. The hyperbox hypervolumes consistently had the largest volume of the three methods. The Gaussian method proved least sensitive to variation in bandwidth, while the support vector machine is the most customizable in terms of its two parameters, but may be susceptible to overfitting.
This poster was made using version 1.4.6 of the hypervolume R package. These results should not be considered representative of the current algorithms in the package - to see current algorithm performance, refer to Blonder et al. 2017.