Certain and correlated neighboring pixels in multispectral image classification

This paper presents two image classification algorithms, which utilize spectral attributes and spatial interdependency of neighboring pixels. These classifiers, namely FKNN-mean classifier and FKNN-median classifier, incorporate the K-nearest-neighbor rule into their algorithms. Fuzzy membership functions are suggested in solving the problem of inherent ambiguity of spatial boundaries in images. Experimental results of these algorithms when compared with those generated by perpixel classification algorithms show significant improvement.<>