A Study on the Classifiers Utilizing Spatial-Contextual Information for Multispectral Image Data

Abstract

The availability of spatial-contextual information for classification of multispectral image data is discussed and the experimental evaluation of the important classifiers is presented. The contextual classifiers treated in this paper are clustered into three groups according to the way of utilizing spatial-contextual information. The first approach is preprocessing of image data such as moving average filters, median filters and edge preserving filters. The second approach is postprocessing of classified results accoring to some rules such as majority filters. The third approach combines the spectral and spatial-contextual information at the same time, and in practice compound decision methods, extended adaptive classifiers and probabilistic relaxations with forced convergence are evaluated. Especially, the last two are the modified classifiers which have been proposed by the authors. The experimental results demonstrate that the highest accuracy of classification is achieved by the classifiers belonging to the third approach and in order they are probabilistic relaxa-tion with forced convergence, extended adaptive classifiers and compound decision methods. Putting together with their expending memory spaces and CPU times, it is concluded that extended adaptive classifiers are best for practical usage.