A framework for cross-scale integration for predictive modeling of species' ranges

Background/Question/Methods

Predictive models of species ranges have seen wide application in ecology and biogeography, particularly with respect to forecasting the effects of climate change on biodiversity. Such models are generally using only a small subset of the total available information about a species; for example, a broad-scale model might use climate variables to predict presence or absence, but ignore what is known about smaller-scale processes such as the effect of local climate on growth, fecundity, and dispersal rates. These approaches can produce unreliable predictions if the factors included in the model do not reflect the true processes driving species distribution; potential failures include biased predictions and underestimates of prediction uncertainty. We developed a highly flexible framework for integrating models at multiple scales using hierarchical Bayesian methods. The resulting model produces probabilistic estimates of species presence with uncertainty that propagates through all models. These predictions reflect all of the information used as input for the original submodels. Although scaling is required to reconcile models at different scales, the scaling need only apply to the spatial domain (geographical or ecological) of the individual submodel, rather than the entire domain of prediction, resulting in much simpler scaling than in traditional approaches.

Results/Conclusions

We used a virtual ecologist approach to perform preliminary simulations combining a broad-scale climate SDM (i.e., relating presence/absence to large-scale climatic variables) and a mechanistic model linking population growth rate to one of the climate variables used in the SDM. The predictions of the integrated model were similar to the naive SDM, but were less biased when extrapolating to climatic regimes not represented in the original data. For regions in ecological space that were covered by both datasets, we observed a decrease in model uncertainty, reflecting the largely concordant predictions of the two submodels. In contrast, changes in uncertainty were more irregular for regions not covered by both submodels, with both increases and decreases relative to the naive model. In general, these results demonstrate that even relatively simple applications of this framework can result in significantly improved predictions, particularly when the domain of prediction is unavailable for data collection (e.g., when projecting to future climate). The framework also affected uncertainty, producing more precise predictions where submodels were in agreement and more variable predictions where submodels disagreed or did not overlap; these uncertainty estimates can be used to identify gaps in current knowledge and prioritize future data collection.