Associate Professor McGill University, Quebec, Canada
Understanding spatial distributions of species and how they will shift with climate change is a central focus of ecology and is crucial to the success of conservation efforts. Species distribution models (SDMs) are useful tools for quantifying distributions by relating species occurrences to environmental conditions, but they are not without challenges. Two main ones arise from using open source species occurrence data such as GBIF. Spatial sampling biases in the data can confound models, and the presence-only nature of the data means that probabilistic SDM predictions cannot be interpreted as true probabilities. Many methods exist for correcting for one or both issues, but the consequences of each are not well understood. Here, we compared the consequences of several corrective methods on the predicted current distributions and climate induced distribution shifts of Canadian vertebrates. This represents an ideal system for studying this question due to the severe sampling bias that exists in Canada. We focused on two commonly used corrective methods: target background sampling and sample effort covariate correction. Additionally, we introduce a novel corrective method which combines presence-only data with nature reserve species checklists. We compared how each method affects modeled species distributions and metrics of climate induced range shifts.
SDMs lacking any corrective methods severely underestimated the distribution of Canadian vertebrate species when compared with known range maps of species. Consequently, the various corrective methods increased the current predicted range sizes by over 400% on average. Meanwhile, SDMs without bias correction predicted that vertebrate ranges would increase by >30% on average in response to climate change, but that only a small portion (~20%) of the current range would remain stable. Conversely, models incorporating sample effort covariate correction predicted an average ~10% decrease in range size, and that a greater portion (~40%) of the current range would remain stable. Evidently, bias correction greatly influences the conclusions we draw from models, and such methods need to be incorporated into studies using SDMs.