Scale is a vital component to consider in any ecological study, and spatial grain size is one of its key facets. Species distribution modeling is an example of ecological research in which scale plays an important role, and it has become an indispensable tool for conservation planning. It is widely accepted that the grain size of predictor variables is an important parameter to consider when modeling a species’ distribution or habitat needs, but studies typically do not explicitly consider the effects of varying the grain size of these predictors on species distribution modeling. Explicit investigations into these effects have usually been done with two grain sizes and have used empirical data. Here we simulated four virtual species whose responses to three environmental predictors varied based on niche breadth (specialist vs. generalist) and the grain size (25 m vs. 200 m) of the predictors across a real landscape. We then created species distribution models using MaxEnt and Generalized Linear Modeling (GLM) techniques at seven different grain sizes from 25 m to 1600 m for each of the four virtual species to determine the effects of grain size on model accuracy and predictions.
Results/Conclusions
Models of a species built at grain sizes further from the one at which that species was simulated generally performed worse, as determined by the AUC statistic, Pearson’s correlations of predicted suitability with the true suitability, and binary distribution maps determined from the total area with suitability above the maximum True Skill Statistic (TSS) threshold. This pattern did not hold across all the species, however, with the broad-scale habitat specialist species more accurately modeled at smaller grain sizes than the “correct” grain size of 200 m at which it was simulated. Variable effects on the MaxEnt models also varied with grain size, with elevation increasing in importance and effect as the model grain size distance from the “correct” grain size increased, while aspect lost importance. The habitat specialist species were more accurately modeled than the generalist species, and the MaxEnt models generally performed better than the GLMs but were more sensitive to grain size. Our results have implications for species distribution modeling and conservation planning, because poorly chosen grain sizes will reduce predictive accuracy of habitat suitability and potentially substantially overestimate areas of presence. We suggest more studies include an analysis of grain size as part of their protocol.