Unsupervised band selection based on weighted information entropy and 3D discrete cosine transform for hyperspectral image classification

Abstract

ABSTRACTBand selection is an effective means of reducing the dimensionality of the hyperspectral image by selecting the most informative and distinctive bands. Bands are usually selected by adopting information theoretic measures, such as, the information entropy or mutual information. However, directly using these information theoretic measures for selection of informative bands is not enough. Therefore, in this paper, the Three Dimensional Discrete Cosine Transform (3D DCT)-based information entropy is used for band selection from a high-dimensional data space. DCT has excellent energy compaction properties for highly correlated images such as hyperspectral images, which reduces the complexity of the separation significantly. First, DCT is applied to extract the complementary features from the raw hyperspectral cube. The weighted information entropy is introduced for measurement of the information about the discrimination ability of each band and reduction of the redundancy among them. An efficient search strategy is then used to select the optimal bands set. The selected bands are fed to a Support Vector Machine (SVM) to carry out the classification task. The proposed band selection approach is tested using three standard hyperspectral datasets such as, Indian Pines, Pavia University and Salinas dataset. The experimental results demonstrate the promising discriminant capability of the DCT features. The proposed method achieves maximum overall classification accuracy of about 84.61% for Indian Pines, 92.83% for Pavia University, and 94.14% for Salinas dataset. When compared to the other competitive band selection methods, the proposed approach shows the significant increase in the overall classification accuracy, when the number of selected bands ranges from 20 to 50.