A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering, School of Engineering and Technology

Thesis (M.Eng.) - Asian Institute of Technology, 1992

This study combines the pyramid segmentation techniques with binary decision tree algorithm to create accurate and more efficient methods for supervised classification of multispectral images. First, three basic discriminant methods were implemented, namely, maximum likelihood function (MLD), best linear discriminant function (BLD) and then binary decision tree (BDT) is introduced, where the BLD is utilized as a decision rule . Experiments indicate that the decision tree method significantly improves upon the efficiency of classification and also gives satisfactory classification accuracy. Based on these three discriminant functions, pyramid segmentation techniques are applied as the preprocessing procedure. The performance of classification is also obviously improved. Two new pyramid segmentation techniques homogeneous nonoverlap pyramid and efficient overlap pyramid method - are introduced in this study. Both of them much reduce the preprocessing time compared to the conventional overlap pyramid segmentation method and also produce more similar segmented data to the original image. In addition, they can be used as real-time application combined with MLD and BLD, to reduce the storage space and increase the processing speed. All proposed methods have been applied to the MOS-1 (Marine Observ ation Satellite) multispectral images of Bangkok and Chiangmai are a, and good results have been obtained. It is shown that pyramid segmentation coii1bined with binary decision tree classifier is an attractive method for the classification of multispectral images.