A GIS modeling method applied to predicting forest songbird habitat

We have developed an approach for using “presence” data to construct habitat models. Presence data are those that indicate locations where the target organism is observed to occur, but that cannot be used to define locations where the organism does not occur. Surveys of highly mobile vertebrates often yield these kinds of data. Models developed through our approach yield predictions of the amount and the spatial distribution of good-quality habitat for the target species. This approach was developed primarily for use in a GIS context; thus, the models are spatially explicit and have the potential to be applied over large areas. Our method consists of two primary steps. In the first step, we identify an optimal range of values for each habitat variable to be used as a predictor in the model. To find these ranges, we employ the concept of maximizing the difference between cumulative distribution functions of (1) the values of a habitat variable at the observed presence locations of the target organism, and (2) the values of that habitat variable for all locations across a study area. In the second step, multivariate models of good habitat are constructed by combining these ranges of values, using the Boolean operators “and” and “or.” We use an approach similar to forward stepwise regression to select the best overall model.

We demonstrate the use of this method by developing species-specific habitat models for nine forest-breeding songbirds (e.g., Cerulean Warbler, Scarlet Tanager, Wood Thrush) studied in southern Ohio. These models are based on species’ microhabitat preferences for moisture and vegetation characteristics that can be predicted primarily through the use of abiotic variables. We use slope, land surface morphology, land surface curvature, water flow accumulation downhill, and an integrated moisture index, in conjunction with a land-cover classification that identifies forest/nonforest, to develop these models.

The performance of these models was evaluated with an independent data set. Our tests showed that the models performed better than random at identifying where the birds occurred and provided useful information for predicting the amount and spatial distribution of good habitat for the birds we studied. In addition, we generally found positive correlations between the amount of habitat, as predicted by the models, and the number of territories within a given area. This added component provides the possibility, ultimately, of being able to estimate population sizes. Our models represent useful tools for resource managers who are interested in assessing the impacts of alternative management plans that could alter or remove habitat for these birds.