Creating Individual Dispersal Hypotheses Improves Stacked Species Distribution Model Performance

Abstract

Stacked distribution models are an important step towards estimating species richness and community composition, but they frequently overpredict these metrics. Developing a priori accessible area (i.e., M) hypotheses to limit the training area based on known dispersal and biogeographic barriers is one way to limit these overpredictions. In order to test the effectiveness of Ms for improving model predictions, 293 species of hummingbird (Aves: Trochilidae) were modeled in a uniform training area and within custom M hypotheses. Locality data was drawn from the Global Biodiversity Informatics Facility, while 13 pre-determined test localities were selected from well-sampled hotspots available within the eBird database. Circles with a radius of 20 kilometers around the eBird localities were removed from the testing dataset and aggregated into a known species list. These lists were compared to published checklists when available. Niche models were thresholded to create species distributions models (SDMs) and then stacked to form presence-absence matrices (PAMs). PAMs were derived for the aforementioned testing localities and their predictions of species richness and community composition were contrasted against the known data. While unconstrained (i.e., uniformly trained) models possess egregious overpredictions, M constrained models perform well and are significantly more accurate at assessing these metrics. Analyzing the amount of overprediction against the amount of effort for each locality also suggests that areas with the most effort are the least likely to possess overpredictions, but this requires further study. Using M constrained models is an effective approach for creating near-accurate estimates of species richness and composition, and therefore a much better method for estimating species distributions. Future research should focus on additional methods for improving individual SDMs within the M training region as well as further analyzing the effects of effort on comparison results.