Cross-realm transferability of species distribution models–Species characteristics and prevalence matter more than modelling methods applied

Abstract

• Species characteristics and prevalence affected transferability more than SDM method applied. • Model transferability was highest on species with similar prevalence in training and transfer data. • No differences in transferability performance of four SDM methods (GLM, GAM, BRT, BART) were found. • Model transferability was highest on species growing close to surface. Species Distribution Models (SDMs) are frequently applied in ecological research, but geographic transferability of SDMs holds major uncertainties. Here, we assess the cross-realm (sea to lake) geographic transferability of four SDM methods: Generalized Linear Models (GLMs), Generalized Additive Models (GAMs), Boosted Regression Trees (BRTs), and Bayesian Additive Regression Trees (BARTs) predicting occurrences of freshwater macrophytes from brackish water sea area (Bothnian Bay) to a freshwater lake environment in Finland. We found that the SDM method applied did not affect model transferability, and majority of the variation in transferability performance was associated with species. For most species model transferability was low, but reasonably good on one third of the species modelled, which had similar prevalences in both marine and freshwater data. These were emergent species or species growing close to shoreline, which presumably share similar environmental niche in terms of growing depth and water turbidity between the two environments. Generally, models which had high interpolation performance, also had higher transferability, but this relationship was not dependent on the SDM method applied. Our results suggest that species prevalence and species-specific characteristics, such as growth form, life history traits and ecological niche, are main contributors to geographic transferability of SDMs.

Topics & Concepts