An Unsupervised Feature Extraction Method for Classification of Hyperspectral Images

hyperspectral images provide huge volume of spectral information for classification of land cover classes. Feature reduction plays an important role as a pre-processing step in classification of high dimensional data. Because of limited available training samples, unsupervised feature extraction is a proper selection for reduction of feature space. We propose an unsupervised feature extraction method in this paper that is called boundary clustering based feature extraction (BCFE). In the BCFE, at first using a clustering algorithm, data is clustered. We use the K-means clustering algorithm in this paper. After clustering, by training of SVM with using the obtained clusters, boundary samples of clusters are calculated. These boundary samples are used for discriminant analysis in the proposed feature extraction method. The experimental results on two real hyperspectral images show the advantage of BCFE in comparison with the most conventional feature extraction methods such as principal component analysis (PCA) and linear discriminant analysis (LDA).