Ensemble modeling for American chestnut distribution: Locating potential restoration sites in Pennsylvania

Abstract

The American chestnut (Castanea dentata Borkh.) was an economically, ecologically, and culturally important tree in eastern American hardwood forests. However, the American chestnut is currently functionally absent from these forests due to the introduction of an invasive fungus (Cryphonectria parasitica (Murr.) Barr) and causal agent of chestnut blight in the early 1900s. Field experiments are being carried out to develop a blight-resistant American chestnut tree, but range-wide restoration will require localized understanding of its current distribution and what factors contribute to suitable American chestnut habitat. While previous studies have researched species distribution of the American chestnut, it is important to understand how species distribution modeling (SDM) technique impacts model results. In this paper we create an ensemble model that combines multiple different SDM techniques to predict areas of suitable American chestnut habitat in Pennsylvania. Results indicate that model accuracy varied considerably by SDM technique – with artificial neural networks performing the worst (Area-Under-the-Curve, AUC = 0.705) and gradient boosting models performing the best (AUC = 0.877). Even though SDM technique accuracy varied, most models identified the same environmental variables as the most important: ratio of sand to clay in the soil, canopy cover, topographic convergence index, and topographic position index. This study offers insight into the best SDM techniques to use, as well as a method of combining SDMs for higher prediction confidence.