Hyperspectral band selection using structural information via hierarchical clustering

Abstract

The techniques of unsupervised band selection are important in the processing of hyperspectral images, such as dimension reduction and abstraction of meaningful information. These techniques aim to select the most informative bands from the original dataset using some similarity metrics and search strategies. For this purpose, the structural information of the images can be utilized in many applications, such as the compact representation and classification of hyperspectral images. Hierarchical clustering method was chosen as the searching strategy and interest points, such as clear edges, blobs, and boundaries of objects, were used as the structural information during the selection of representative bands. At the same time, a similarity criterion was developed to generate a representative band subset from the hyperspectral bands. Finally, the representative band subsets obtained from the proposed methods were classified by the k-nearest neighborhood algorithm and the classification accuracy results were taken as performance criteria. The performances of the proposed methods were compared with the Walumi and Waludi techniques that are common in the literature. Experimental results show that the proposed methods give better results than the others.