Empirical Research on Cluster Analysis of Spectral Information of Hyperspectral Remote Sensing Data

摘要

The most notable features of hyperspectral remote sensing data are high spectral resolution and high dimensionality. Analyzing the spectral information of the data with clustering requires proper spectral measurements. And clustering results validating methods satisfy high-dimensional data are required. In this paper, three clustering algorithms of K-means, K-means++, and Chameleon were used in the empirical research, based on the famous data set of IndianPine. In the experiments, four spectral measures of SBD, SID, SSD and SPM were used as spectral similarity measures, and S_Dbw was chosen as clustering validating method. As S_Dbw is a relative validating index, the algorithms with different parameter combinations had been experimented many times, verified the effectiveness of the solution composed with selected clustering methods, spectral measures, and clustering validating method. Experimental results showed that hyperspectral information enhanced the ability to distinguish ground objects; Chameleon's performance was best, and S_Dbw could effectively analyze hyperspectral data; the performance of the spectral measurements showed a certain correlation with the clustering method.