Segmentation-based clustering of hyperspectral images using local band selection

Abstract

This study addresses the problems associated with high dimensionality of hyperspectral images in reference to clustering. A new local band selection approach that takes both relevancy and redundancy among the bands into account while obtaining the multiple relevant set of bands is developed. The local band selection approach is then incorporated within a multistage clustering framework that includes three stages: segmentation, region merging, and projected clustering. At first, k-means is used to produce initial segments/regions. Then, in the region merging stage, the modified local mutually best region merging strategy is applied on the initial segments to produce the refined segmentation map. Finally, an improved projected clustering technique is used to group these segments into a fixed number of clusters. Further, the main principle of projected clustering, that different sets of point may cluster better for different subsets of dimensions, is extended to region merging by incorporating the suggested local band selection approach. The framework requires input for two major parameters, which are number of clusters (k) and number of relevant bands (l). The framework is tested over two hyperspectral images and compared with other clustering frameworks. The experimental results confirm the effectiveness of the proposed framework.