AN EFFICIENT INITIALIZATION METHOD FOR K-MEANS CLUSTERING OF HYPERSPECTRAL DATA
- 1Department of Geomatic Engineering, Faculty of Engineering, University of Tehran, Tehran, Iran
- 2Department of Geomatic Engineering, Faculty of Engineering, University of Bojnord, Bojnord, Iran
- 3Department of Geography, University of Ottawa, Ottawa, Canada
Keywords: K-means, Clustering Initialization methods, Unmixing, MVES, Hyperspectral data
Abstract. K-means is definitely the most frequently used partitional clustering algorithm in the remote sensing community. Unfortunately due to its gradient decent nature, this algorithm is highly sensitive to the initial placement of cluster centers. This problem deteriorates for the high-dimensional data such as hyperspectral remotely sensed imagery. To tackle this problem, in this paper, the spectral signatures of the endmembers in the image scene are extracted and used as the initial positions of the cluster centers. For this purpose, in the first step, A Neyman–Pearson detection theory based eigen-thresholding method (i.e., the HFC method) has been employed to estimate the number of endmembers in the image. Afterwards, the spectral signatures of the endmembers are obtained using the Minimum Volume Enclosing Simplex (MVES) algorithm. Eventually, these spectral signatures are used to initialize the k-means clustering algorithm. The proposed method is implemented on a hyperspectral dataset acquired by ROSIS sensor with 103 spectral bands over the Pavia University campus, Italy. For comparative evaluation, two other commonly used initialization methods (i.e., Bradley & Fayyad (BF) and Random methods) are implemented and compared. The confusion matrix, overall accuracy and Kappa coefficient are employed to assess the methods’ performance. The evaluations demonstrate that the proposed solution outperforms the other initialization methods and can be applied for unsupervised classification of hyperspectral imagery for landcover mapping.