تحلیل ممیز غیرپارامتریک بهبودیافته برای دسته‌بندی تصاویر ابرطیفی با نمونه آموزشی محدود

Feature extraction performs an important role in improving hyperspectral image classification. Compared with parametric methods, nonparametric feature extraction methods have better performance when classes have no normal distribution. Besides, these methods can extract more features than what parametric feature extraction methods do. Nonparametric feature extraction methods use nonparametric scatter matrices to compute transformation matrix. Nonparametric Discriminant Analysis (NDA) is one of the nonparametric feature extraction methods in which, to form nonparametric scatter matrices, local means of samples and weight function are used. Local mean is calculated by k nearest neighbors of each sample and weight function emphasizes on boundary samples in between class scatter matrix formation. In this paper, modified NDA (MNDA) is proposed to improve NDA. In MNDA, the number of neighboring samples, when measuring local mean, are determined considering position of each sample in feature space. MNDA uses new weight functions in scatter matrix formation. Suggested weight functions emphasizes on boundary samples in between class scatter matrix formation and focus on samples close to class mean in within class scatter matrix formation. Moreover, within class scatter matrix is regularized to avoid singularity. Experimental results on Indian Pines and Salinas images show that MNDA has better performance compared to other parametric, nonparametric feature extraction methods. For Indian Pines data set, the maximum average classification accuracy is 80.34%, which is obtained by 18 training samples, support vector machine (SVM) classifier and 10 extracted features achieved by MNDA method. For Salinas data set, the maximum average classification accuracy is 94.31%, which is obtained by 18 training samples, SVM classifier and 9 extracted features achieved by MNDA method. Experiments show that using suggested weight functions and regularized within class scatter matrix, the proposed method obtained better results in hyperspectral image classification with limited training samples.