Continental species distribution and biodiversity predictions depend on modeling grain
As global change accelerates, accurate predictions of species distributions and biodiversity patterns are critical to prevent population declines and biodiversity loss. However, at continental and global scales, these predictions are often derived from species distribution models (SDMs) fit at coarse spatial grains uninformed by ecological processes. Coarse-grain models may systematically bias predictions of distributions and biodiversity if they are consistently over- or under-estimating area with suitable habitat, and this bias may intensify in regions with heterogenous landscapes or with poor data coverage. To test this, we fit presence-absence SDMs characterizing both the summer and winter distributions of 572 North American bird species - nearly the entire avian diversity of the US and Canada - across five spatial grains from 1 to 50 km, using observations from the eBird citizen science initiative. We find that across both seasons, models fit at 1 km performed better under cross-validation than those at coarser scales and more accurately predicted species presences and absences at local sites. Coarser-grain models, including models fit at 3 km, consistently under-predicted range area relative to 1 km models, suggesting that coarse-grain estimates of distributions could be missing important habitat. This bias intensified during summer (83% of species) when many birds have smaller operational scales via localized home ranges and greater habitat specificity while breeding. Biases were greatest in heterogenous desert and scrubland regions and lowest in more homogenous boreal forest and taiga-dominated regions. When aggregating distributions to produce continental biodiversity predictions, coarse-grain models overpredicted diversity in the west and underpredicted it in the great plains, prairie pothole region and boreal/taiga zones. The modern availability of high-performance computing and high-resolution observational and environmental data provides opportunities to improve continental predictions of species distributions and biodiversity.