Abstract
The purpose of this study was to compare the area classification accuracy of each of the following options of image classification: 1. a pixel-by-pixel maximum likelihood gaussian classifier. 2. a sample classifier based on B-distance (derived from the Bhattacharyya distance). 3. a sample classifier based on the generalized maximum likelihood approach. 4. the pixel-by-pixel "single-cell signature acquisition" option of the Image-100 System. 5. same as option 1, but using the following simple decision rule for classification: if the percentage of pixels classified into the same class, within a given test field, exceeded a threshold value of 60%, they were all classified into the same class. 6. same as option 4, but using the decision rule given in option 5.
LANDSAT multispectral scanner data of the following three test sites of the state of São Paulo, Brazil, were classified using each of the above six options: 1. São José dos Campos 2. Cachoeira Paulista 3. Jardinópolis.
Considering both the errors of omission as well as commission, the sample classifier (option 2) yielded better classification accuracy, as compared to the maximum likelihood gaussian classifier (option 1) as well as single cell (option 4). Options 5 and 6 considerably improved the classification accuracy of options 1 and 4 respectively.
Date of this Version
1979