A novel hyperspectral image clustering method based on spectral unmixing

In this paper, a novel hyperspectral image clustering procedure, which is based upon the Fully Constrained Least Squares (FCLS) spectral unmixing method, is proposed. The proposed clustering method consists of three major steps: endmember extraction, unmixing procedure and hardening process via the winner-takes-all approach. To estimate the optimal number of endmembers, instead of using the background signal subspace identification methods, the number of endmembers is varied in a predefined interval and the commonly accepted VCA (Vertex Component Analysis) algorithm is employed to extract the endmembers´ spectra. At each iteration, the bandwise Root Mean Square Error (RMSE) between the reconstructed image, obtained from estimated fractions. and the original image is computed and the mean of all bandwise RMSEs is regarded as a measure to choose the optimum number of endmembers. Experiments conducted on the Indian Pines challenging dataset proved the superiority of proposed method over the K-Means and Fuzzy c-Means methods in terms of the widely used Adjusted Rand Index measure.