Arwa Alturki and Ouiem Bchir, King Saud University, Saudi Arabia

Spectroscopy or hyperspectral imaging consists in the acquisition, analysis, and extraction of the spectral information measured on a specific region or object using an airborne or satellite device. Hyperspectral imaging has become an active field of research recently. One way of analysing such data is through clustering. However, due to the high dimensionality of the data and the small distance between the different material signatures, clustering such a data is a challenging task.In this paper, we empirically compared five clustering techniques in different hyperspectral data sets. The considered clustering techniques are K-means, K-medoids, fuzzy Cmeans, hierarchical, and density-based spatial clustering of applications with noise. Four data sets are used to achieve this purpose which is Botswana, Kennedy space centre, Pavia, and Pavia University. Beside the accuracy, we adopted four more similarity measures: Rand statistics, Jaccard coefficient, Fowlkes-Mallows index, and Hubert index. According to accuracy, we found that fuzzy C-means clustering is doing better on Botswana and Pavia data sets, K-means and K-medoids are giving better results on Kennedy space centre data set, and for Pavia University the hierarchical clustering is better.

Image Processing, Hyperspectral Imaging, Imaging Spectroscopy, Clustering, FCM, K-means, K-medoids, hierarchical, DBSCAN