Improvement of Clustering for Hyperspectral Images using Spectral Information Divergence

K-Means is one of the most frequently used unsupervised classification approaches for remotely sensed image analysis. In standard K-Means version, the Euclidean distance (ED) has used to estimate the dissimilarity between an unknown vector data and the cluster center. Since, this measure is very sensitive to topographic and environmental effects on spectral observations, we have proposed to replace it with a new one for goal of hyperspectral image clustering. The Spectral Information Divergence (SID) is a stochastic measure that is a more reliable dissimilarity measure when compared to ED as a deterministic measure. Where the ED measure the spectral distance between vector data and the clusters, SID models the probability distributions for vector data and clusters by normalizing their spectral signatures and measures the distances between them. This idea has applied to develop an enhanced clustering framework. The experimental results on three real hyperspectral images collected by HyMap, HYDICE and Hyperion sensors show that the proposed method improves classification results. In the manner that the Kappa coefficient of the classification results of three hyperspectral imagery datasets increased by about 7%, 56% and 10%, respectively.