A semi-automated approach to spectral training for a maximum likelihood classification is shown to maintain or improve classification accuracy while reducing analyst input five-fold. The semi-automated approach is based on spectral sampling via region growing and training set refinement via transformed-divergence based mergers and deletions. Classification accuracies at the Anderson level II/1II in northern Wisconsin were higher for the semi-automated approach in five of six combinations of imagery, analyst, and study area, at significantly.
Bibliographical noteFunding Information:
Acknowledgments This work was supported by the University of Wisconsin Institute of Environmental Studies, the College of Agriculture and Life Sciences, and by Xi Sigma Pi. We also wish to thank M. Mohamed for his contributions to this research.