A semisupervised feature metric based band selection method for hyperspectral image classification

This paper presents a novel semi-supervised band selection technique for classification of the hyperspectral image. In our proposed method, a simple and efficient metric learning algorithm, i.e. relevant component analysis, is adopted for learning the whitening transformation matrix from which a feature metric is constructed for feature selection. This metric assesses both the class discrimination capability of the single band and the spectral correlation between the any two bands. The affinity propagation technique is then employed as the clustering strategy to select an effective band subset from original spectral bands. Experimental results demonstrate that the proposed method can effectively select the representative bands and reduce the band redundancy for improving the classification accuracy. In addition, the comparison with some literature band selection methods also confirms the superiority of the proposed approach.