Habitat suitability models are useful tools that can be used to predict changes in habitat in response to broad-scale factors such as land-use change and climate change.  Because organisms interact with their environments at multiple spatial scales, these models may need to address relationships at different spatial scales to adequately capture different activities such as foraging, breeding, and movement.  In general, however, models for predicting habitat suitability are built with data collected at a single spatial scale.  We used Random Forests, a classification tree approach, to build habitat suitability models at two spatial scales—one to address nesting habitat and one to address foraging habitat—for a population of red-cockaded woodpeckers (RCW; Picoides borealis) at Fort Benning, Georgia.  Our models included habitat characteristics known to be important to the RCW, including stand age, basal area, and tree species composition.  For each scale, we used locations of known presences and absences to build predictive models.  These models were then used to predict a probability of occurrence for the RCW in each cell of the landscape.  Model performance was assessed using test data not used in building the models. 

Results/Conclusions Relatively fine-scale sampling of nesting habitat resulted in a relatively accurate habitat map, correctly classifying 86% of the known presences and 81 % of absences.  Broader scale sampling of foraging habitat also resulted in a relatively accurate habitat map, correctly classifying 83% of the presences and 83% of absences.  We then combined the two models to create an ensemble model of habitat probability.  The ensemble model was highly accurate, correctly classifying 91% of known presences and 82% of absences and outperforming two other available models of RCW habitat for Fort Benning.  Multi-scale ensemble models are one important step towards filling the pressing need for accurate predictive models to forecast the potential effects of both land-use and climate change. Our analyses demonstrate a simple approach for combining multi-scale habitat model predictions to produce highly accurate classifications.