Stratification of landsat thematic mapper data, based on regional landscape patterns, to improve land-cover classification accuracy of large study areas
Regionally based landscape characterization provides a means to identify the variation that may exist across large areas. This variation, however, may result in increased confusion among spectral classes during image classification. The following research demonstrates the utility of pre-classification stratification of large study areas into spectrally consistent subareas or strata to deal with this spectral variability. The study focuses on classification of the non-urban component of Landsat Thematic Mapper (TM) imagery using hybrid guided clustering methods for classification of individual strata. Imagery covering the Fox/Wolf River Basin, Wise. was stratified with a regional ecoregion map for Michigan, Minnesota and Wisconsin and a surficial geology map of Wisconsin. The accuracies of the stratified classifications were compared to that of a non-stratified classification on the basis of KHAT values and Z-scores. In addition, the utility of hybrid guided clustering methods was assessed for this mixed agricultural/forested area. Results indicate that stratification can be used to significantly improve land-cover classification accuracy if land-cover classes are not overgeneralized and if the landscape is not overstratified. Results also indicate that hybrid guided clustering provides a more consistent and logical training data evaluation process than might normally occur in pure supervised or unsupervised classification.
Charlotte, North Carolina