Improvement of Unsupervised Classification for Hyperspectral Images using Gustafson-Kessel Clustering Model

C-means clustering models are one of the most widely used methods for unsupervised classification of any data. Fuzzy c-means (FCM) is one of the most well-known clustering models in which, each data may be belonged to multiple clusters with different membership degree between 0 and 1. This model has been employed for different application including remotely sensed data classification. FCM model uses Euclidean distance for clustering and assumes the same shape/distribution for all of clusters. However, this causes misclassification in data in which the classes have different shape and size. In this paper, Gustafson-Kessel clustering model is presented to overcome this problem. This model is based on using a fuzzy covariance matrix for each cluster which does not consider the same geometric shape, size and orientation for all clusters. The above models were applied for clustering of hyperspectral imagery issue of Hyperion, ROSIS and CASI sensors. The results of Gustafson-Kessel clustering model prove that the accuracy of classification increased about 12.5% for Hyperion imagery and about 8.45% for ROSIS imagery. Also, the visual test on CASI imagery show that Gustafson-Kessel clustering model has better performance.