抄録
A Landsat data classification method to classify land cover using modified unsupervised classification techniques is presented.
It is possible to interpret physical properties of the classification result of this method. This method combines clustering with factor analysis and typical land cover spectral reflectance properties of each band. For the research, a test area including fire damage sites was used. A fire damage condition map of this area, prepared by field surveying and aerial photo interpretation, was already available.
The adopted land cover classification procedure was as follows ; Step 1. (Identification of fire damage representa-tives)
Clustering was carried out on the randomly selected Landsat data for the area of interest. Fac-tor analysis was done using band averages of the clusters which had been gotten by the above mentioned clustering. Factor scatter diagram was plotted with the two factors which were enough to represent the total variance. The factor loading of each band on each factor was examined, and the physical properties of each factor was interpreted. The relative positions of each cluster in the diagram and the spectral curve of each cluster were examined, and the clusters were regrouped into such a smaller number of groups that could be interpreted physically. The clusters belonged to the group which was interpreted as fire damage, were taken as the representatives of fire damaged area.
Step 2. (Estimation of severity of damage) Clustering was carried out on the Landsat data which were only belonged to the selected clusters (fire damaged area) in Step 1. Three severity class groups of fire damaged area was interpreted by the same procedure as in Step 1.
Step 3. (Classification of fire damaged area) The Landsat data of the fire damaged area were classified into three fire damage severity classes by the maximum-likelihood classification method using the three groups which were gotten in Step 2 as the training data. Classification accuracy was about 80% when the classified Landsat image was compared with the existing fire damage condition map.
