Strategies for sampling pseudo-absences for species distribution models in complex mountainous terrain

Abstract

O_LIPredictions from species distribution models (SDMs) that rely on presence-only data are strongly influenced by how pseudo-absences are derived. However, which strategies to generate pseudo-absences give rise to faithful SDMs in complex mountainous terrain, and whether species-specific or generic strategies perform better remain open questions. C_LIO_LIHere, across 500 plant species, we investigated comprehensively how predictions of SDMs at a 93 m spatial resolution are influenced by pseudo-absence strategies, using the complex topography of the Swiss mountains as a model system. We used five generic (random, equal-stratified, proportional-stratified, target, density) and three species-specific (target specific, density specific and geographic specific) approaches to derive pseudo-absence data. We conducted performance tests for each of our eight strategies in combination with (a) spatial bias, generated within our occurrence dataset on sites with highest sampling density, to investigate how this common bias problem influences the performance of pseudo-absence sampling strategies, and (b) a new approach to reduce model extrapolation in environmental space by including background data from all environmental conditions of the study area. SDMs were evaluated against an independent and well-sampled dataset of true presences and absences. C_LIO_LIThe random, the density (generic), and the geographic specific (species-specific) strategies consistently performed best, even in cases of strong spatial sampling bias in the occurrence data. Including a background of environmentally stratified pseudo-absences improved predictions of species distributions towards environmental extremes, and significantly reduced spatial extrapolations of model predictions in environmental space. C_LIO_LIOur results indicate that both generic and species-specific pseudo-absence strategies allow estimating robust SDMs and we provide clear recommendations which strategies to choose in complex terrain and when presence data are prone to high sampling bias. In datasets with strong sampling bias, most pseudo-absence strategies produce extrapolation problems and we additionally recommend environmentally stratified pseudo-absences in these cases. Overall, in species rich datasets the use of complex and computationally demanding, species-specific pseudo-absence strategies may not always be justified compared to simpler generic approaches. C_LI