Hyperspectral image classification via within class similarity for limited training samples problem

In hyperspectral image classification, finding the best criterion for separating classes and assign a accurate label to pixels is a major challenge. In traditional classification methods, as SVM, SAM, KNN the uniform criterion is considered for separating the classes as margin, angle, distance, etc. In classification process the distance between unlabeled pixel and a class is calculated. The minimum distance between unlabeled pixel and each class is the main criterion for labeling the pixel. In this paper a simple method based on the within class similarity is proposed for hyperspectral images classification that is proper for limited training samples problem. At first, the best specification of a class is explored based on the training samples, although this specification for each class can be different. Later, unlabeled pixel is added to the training samples of either class for recalculation of each one. Now this specification is compared with the former specification. The unlabeled pixel belongs to the class with minimum difference between former and later specification. Experimental results outcomes, based on the hyperspectral images, represent the effectiveness of this method.