Ensemble algorithms for ecological niche modeling from presence-background and presence-only data

Abstract

Ecological niche modeling is central to many problems in population dynamics, conservation biology, biogeography, and evolutionary ecology. Recently, methodological advances have been achieved by adopting modeling principles from statistical and machine learning. One such principle that has not been explored is that voted ensembles of weakly tuned (underfit) models often exhibit similar bias and lower variance than individual well-tuned models. A further contribution of machine learning is the one class support vector machine. One class support vector machines may be more suited to ecological niche modeling than classification algorithms because the ''boundary estimation'' property of this class of models better represents the conception of the niche as the set of environments in which populations of a species persist. This paper brings together these approaches and applies them to a data set on the distribution twospined blackfish (Gadopsis bispinosus) in southeastern Australia with the aim of improving performance and enriching the repertoire of methods for ecological niche modeling. First, a study of low-bias bagging (LOBAG) applied to conventional two-class support vector machines, an approach to presence-background modeling, found this method to be comparable to another widely used method, MAXENT. Next, I introduce a variant of LOBAG for one class classification, LOBAG-OC, to be used for ecological niche modeling with presence-only data. This method performed better than other presence-only methods (DOMAIN and BIOCLIM) and nearly as well as LOBAG and MAXENT, despite the fact that the presenceonly problem is considerably more difficult than the corresponding presence-background problem. Results of these analyses suggest that LOBAG and LOBAG-OC may provide robust, high-performing solutions to the presence-background and presence-only modeling problems and point to ways in which ecological niche modeling may be simultaneously both improved and automated. © 2014 Drake.