首页 > , Vol. , Issue () : 1993-2002
The Hughes problem together with strong intra-band correlations and massive data seriously hinders hyperspectral processing and further applications. Dimensionality reduction using band selection can be used to conquer above problems and guarantee the application performance of hyperspectral data. Particularly, spectral clustering is a typical method for high dimensional hyperspectral data, which finds clusters of all hyperspectral bands on the connected graph and select the representatives. Unfortunately, the regular similarity measures are negatively affected by outliers or noise of hyperspectral data in measuring the similarity of different bands, and moreover, they could only represent one feature of band similarity and have respective limitations. The obtained similarity matrix accordingly could not represent the full information of band selection required and could not guarantee to obtain aimed bands from spectral clustering. Therefore, we propose a robust multi-feature spectral clustering (RMSC) method to solve the above two problems and promote the performance of hyperspectral band selection from spectral clustering. The RMSC combines multiple features of similarity measures for pairwise bands, i.e., information entropy, band correlation and band dissimilarity to construct the integrated similarity matrix. It utilizes Spectral Information Divergence (SID) to quantify the information entropy between pairwise bands. The coefficient correlation (CC) is utilized to measure the band correlations and construct the similarity matrix of band correlations. Moreover, considering the inner clustering structure of all bands, the Laplacian graph (LG) is adopted to construct a similarity matrix and show the dissimilarity between different bands. In addition, the spectral angle distance (SAD) matrix is construct to reflect the similarity from the aspects of overall differences. The RMSC regards that each similarity matrix of all four features reflect the underlying true clustering information of all bands and has low-rank property. It formulates the estimation of combined dissimilarity matrix into a low-rank and sparse decomposition problem and utilizes the Augmented Lagrangian Multiplier (ALM) to solve it. After that, it implements the regular spectral clustering on the integrated similarity matrix and select the representative bands from each cluster. Two Hyperspectral datasets are used to design four groups of experiments and testify the performance of RMSC. Five state-of-the-art methods include WaluDI, Fast Density-Peak based Clustering (FDPC), orthogonal projections based band selection (OPBS), improved sparse spectral clustering (ISSC) and SC-SID, and support vector machine (SVM) is used to quantify the classification accuracy. Experimental results show that RSMC outperforms other five band selection methods in overall classification accuracy (OCA), with shorter computational time. Moreover, the regularization parameter is insensitive to RMSC, and a small candidate could bring about high classification accuracy. RMSC is better in selecting representative bands than current spectral clustering such as ISSC and can be a good choice in hyperspectral dimensionality reduction.