Spatial Granule based Clustering Technique for Hyperspectral Images

Lack of labelled class information in data, huge data size and spatial autocorrelation pose challenge in the classification of remote sensing images. The concepts of clustering, dimensional reduction, granular computing and spatial neighbourhood address the above-discussed challenges. In the proposed method, the Unsupervised Spatial Granulebased Clustering algorithm, we have used Principal Component Analysis for dimensional reduction. Then Spatial Clustering using neighbourhood is applied to principal components to give spatial granules. Finally, we applied kMeans on the spatial granules to yield concept-based clusters. The proposed clustering technique produced more accurate results than other unsupervised clustering techniques. The external cluster quality measures compare the clusters produced by different clustering techniques. The highest Fscore, significant accuracy and better cluster quality parametric values on Indian Pines, Salinas-A and Kennedy Space Center images show the efficacy of the proposed method.