Abstract:
Clustering of hyperspectral images is a challenging task due to the high spectral resolution and the presence of elaborate spatial structures in the data. In this study, a new clustering framework for hyperspectral imagery is proposed based on the concept of nearest neighbor. The framework comprises three major steps. In the first step, hyperspectral image segmentation is performed using the unsupervised k-means method. In the second step, the segmentation map obtained from the previous step is considered as a cluster map and is refined iteratively by utilizing the mutual nearest neighbor (MNN) information. Finally, clusters are merged repeatedly with their first nearest neighbor (1-NN), until k-clusters are obtained Experiments on two widely used hyperspectral images demonstrate that the proposed framework has a high potential to attain better clustering performance most of the time.