Clustering of Hyperspectral Image using Fuzzy C- Means Based on Spectral Similarity Measures

As an unsupervised Classification method, clustering aims to separate a data set into a number of clusters or groups. The main objective of this technique is that in a cluster, two vector data are sufficiently similar, and in two different clusters, two vector data are as dissimilar as possible. One of the most frequently used clustering algorithms is Fuzzy C-Means (FCM), This algorithm has been employed for different applicationsincluding remotely sensed data classification. In hyperspectral imagery, due to the topo graphic effects of ground surface and the environmental factors, the observed spectral signatures for the similar objects are not similar. As a result, for the pixels covered by these objects, different brightness values will be recorded. This phenomenon leads to the misclassification of similar pixels. FCM algorithm uses the Euclidean distance as similarity criterion for measuring of closeness within data and clusters. Euclidean distance is sensitive to the brightness variations. Consequently clustering of hyperspectral images can’t lead to acceptable results. In this paper, it has been used some more reliable similarity measures such as the spectral angle and correlation. These measures are lesssensitive to the brightness variations. The modified FCM algorithms are applied on hyperspectral imagery collected by Hyperion sensor. The image data cover an agricultural region at 30m of spatial resolution. The image has 145 useful spectral bands. The experimental test on hyperspectral image data showed that this strategy improves the clustering accuracy of hyperspectral images and increases the kappa coefficient about 10%.

Keywords:

Clustering , Fuzzy C-Means , Euclidean distance , Spectral Similarity Measures , Hyperspectral ImageComputations