Using Cluster Analysis and a Classification and Regression Tree Model to Developed Cover Types in the Sky Islands of Southeastern Arizona

The objective of this study was to develop a rule based cover type classification system for the forest and woodland vegetation in the Sky Islands of southeastern Arizona. In order to develop such a system we qualitatively and quantitatively compared a hierarchical (Ward’s) and a non-hierarchical (k-means) clustering method. Ecologically, unique groups represented by only a few plots were appropriately distinguished using k-means, while Ward’s combined these unique plots into the large mixed conifer groups. Similarly, plots dominated by more than one species were more appropriately grouped with other mixed-species plots using k-means. The two clustering methods were numerically compared using a classification and regression tree (CART) model. Groups based on the two clustering methods had similar recovery rates, but k-means groups required fewer nodes or decision rules. Based on these results we developed a detailed cover type classification system for the existing vegetation of the Sky Islands in southeastern Arizona. The final cover types were based on the original k-means clusters, with some minor modifications made using CART analysis to compensate for overlapping values. This allowed us to transform the CART output into a dichotomous identification key for 20 detailed cover types. Finally, these detailed cover types were linked to a flexible three-level hierarchical framework that allows users to aggregate or segregate forest lands as needed. The hierarchical organization of this framework is similar to the natural organization of ecosystems, which will aid our understanding of natural processes in these forest and woodlands.