Modified Graph-Based Algorithm for Efficient Hyperspectral Feature Extraction

Abstract

Since the Laplacian Eigenmaps (LE) algorithm suffers from a spectral uncertainty problem for the adjacency weighted matrix construction, it may not be adequate for the hyperspectral dimension reduction (DR), classification or detection process. Moreover, only local neighboring data point’s properties are conserved in the LE method. To resolve these limitations, an improved feature extraction technique called modified Laplacian Eigenmaps (MLE) for hyperspectral images is suggested in this paper. The proposed approach determines the similarity between pixel and endmember for the purpose of building a more precise weighted matrix. Then, based on the obtained weighted matrix, the DR data are derived as the Laplacian eigenvectors of the Laplacian matrix, constructed from the weighted matrix. Furthermore, the novel proposed approach focuses on maximizing the distance between no nearby neighboring points, which raises the separability among ground objects. Compared to the original LE method, experiment results, for hyperspectral images classification and detection tasks, have proved an enhanced accuracy.