Mapping subalpine forest types using networks of nearest neighbour classifiers

Predictive vegetation mapping systems use a variety of remote sensing and other geospatial data in a mixture of empirical and biophysical models to predict vegetation characteristics. These systems are important tools for understanding and monitoring spatial patterns over the landscape. We have developed a system that partitions the geospatial data into resource gradients, indirect gradients and remotely sensed reflectance. These data are input to nearest neighbour classifiers and the resulting likelihoods are combined used Dempster's rule. We show that this system yields a more accurate classification than if the data were input to a single classifier.