Using existing large-area land-cover maps to classify spatially high resolution images

Abstract

This paper presents Template-Guided Classification (TGC), a technique for using the class labels of existing large-area land-cover maps to automatically classify spatially highresolution images. TGC uses land-cover images as templates to guide hierarchical clustering and labeling. To test TGC, 10-m SPOT 5 HRG images and 1-m colour orthophotos of the Vermilion River watershed, Canada were classified into forest/non-forest classes using the 25-m Earth Observation for the Sustainable Development of forests (EOSD) landcover map as a template. Although the average accuracies of the 10-m SPOT classifications were poor, the 1-m orthophoto accuracies were much higher (87% forest user's accuracy, 82% forest producers accuracy, 93% overall accuracy). TGC classification accuracies were highly variable. Further investigation is needed to determine whether TGC can be made into a robust procedure.