BOUNDARY BASED SUPERVISED CLASSIFICATION OF HYPERSPECTRAL IMAGES WITH LIMITED TRAINING SAMPLES
Keywords: Hyperspectral data, limited training samples, non-parametric classifier
Abstract. One of the most important and challenging problems in supervised classification of high dimensional data is limited available training samples. Using the parametric classifiers is not appropriate in this condition. Thus a new simple nonparametric supervised classifier based boundary samples of each class is proposed in this paper that need no statistic parameter for classification. Accuracy and reliability of this classifier is compared whit other non-parametric classifiers such as Parallelepiped (box), K nearest neighbours (KNN), Artifical Neural network (ANN) and SVM and also a parametric classifier that use only first order statistic, Minimum Euclidean Distance (MED), for different four datasets, AVARIS data, Pavia University, Pavia center and Salinas data. The results of experiments show that proposed classifier in despite of simplicity has appropriate and reasonable efficiency.