The Effect Of Training Block Size On Unsupervised Classification Of Landsat Thematic Mapper Imagery PDF Download

Remotely sensed data is an important component of land use/land cover (LULC) studies. This research utilized the vegetation-impervious surface-soil (V-I-S) model. Using Enhanced Thematic Mapper Plus (ETM+) imagery, this research compared the accuracy of supervised and unsupervised classification by analyzing three study areas in and near Little Rock, Arkansas. The first study area was a homogeneous region comprised primarily of water features. The second study area was a region of an intermediate mix of land cover classes. The third study area was a region of heterogeneous land cover composition between the four land cover classes of the V-I-S model. Upon the completion of supervised and unsupervised classification, 200 points for each area were randomly generated using a stratified random sampling approach. The land cover data associated with these points were then compared to ground truth data derived from higher-resolution imagery from the National Agriculture Imagery Program (NAIP). Based on error matrices, the homogeneous and intermediate study areas featured higher accuracy values for unsupervised classification over supervised classification. For the heterogeneous study area, supervised classification was more accurate than unsupervised classification by one percent.