Spatial and temporal components of environmental predictors in species distribution modeling: a case study using Bell's Vireo (Vireo bellii)

Abstract

The largest threats to biodiversity are global climate change and habitat loss, both of which are global concerns due to decreases in species’ populations. Understanding species’ responses to both threats is needed and a common practice used is species distribution modeling (SDM). SDM is a predictive modeling technique, which incorporates environmental conditions associated with species presence locations to derive species-environment relationships that are used to predict geographic locations of species across space and time. As the nature of SDM is both spatial and temporal, the scale of data used affects model performance and predictions. Specifically, the grain and extent of predictor variables influences model performance and hence, model interpretation. I set out to address spatial and temporal scale concerns in SDM using Bell’s Vireo (Vireo bellii), a Neotropical migratory songbird, as a case study. Bell’s Vireo is a species of concern that has shown declining trends across its range, where it inhabits threatened landscapes such as riparian and shrubland-grassland ecotones. Here I describe the use of Bell’s Vireo presence locations to address the role of extent, effects of resampling and grain size, as well as the temporal aspects of environmental predictors in SDM.

First, I compared model performance and potential distributions across three study area extents under eight variable selection techniques and five species’ occurrence data compilations. Overall, I was able to show interactions among these model components, specifically that data quality influenced model performance more than study extent size and variable subset. Second, to investigate the effects of grain size manipulation on SDM, I compared twelve grain sizes resampled using three upscaling techniques. My results showed that model performance in terms AUC was influenced by resampling method, but not grain size, whereas the model performance metric, omission error, was not influenced by resampling technique or grain size, whereas prediction of potentially suitable area was influenced by both resampling and grain size. Last, when investigating temporal effects on SDM performance, I found that more temporally explicit variables, such as seasonal variables, did not necessarily improve model performance although it did increase proportions of suitable area compared to annual variables.