Kmeans-CM Algorithm With Spectral Angle Mapper for Hyperspectral Image Classification

Abstract

Hyperspectral images possess the characteristics of high dimensionality, which causes “dimensional disaster” and low classification accuracy, in response to the problems, based on traditional k-means algorithm and considering the importance of different bands for classification, and also combining both intra-class and inter-class information, a Kmeans-CM (K-means with correlation coefficient and maximize inter-class distance) algorithm with spectral angle mapper for hyperspectral image classification is proposed. First, weights of bands are defined by introducing coefficient of variation and spectral angle mapping, which measures the importance of bands for classification. Second, in order to intensify correlation between pixels in the same category, correlation coefficient is introduced to reset intra-class distance. Then, inter-class information is introduced for clustering by maximizing the distance between class centers and global center to reduce effect of local optimum of clustering effect. Finally, the K-means clustering objective function is redefined according to the band weights and intra-class, inter-class information, and also solving optimally. The overall classification accuracy of the method reached 84.47%, 90.08% and 80.45% on classical hyperspectral data sets Pavia University, Salinas and Botswana respectively, and the results show that the proposed algorithm owns good classification performance.