Abstract
Recently, graph-embedding framework has been developed for dimensionality reduction (DR) and classification of hyperspectral images (HSI). However, it suffers from intraclass difference and interclass similarity in complex scenarios. In this article, an unsupervised DR method called superpixelwise collaborative-representation graph embedding (SPCRGE) is proposed for the HSI classification. In SPCRGE, homogeneous regions called superpixels are generated by grouping spectral-similar and spatially adjacent pixels. Pixels in one homogeneous region come from one class with high probability. Then, Laplacian regularized superpixelwise collaborative representation (SPCR) of a query pixel, i.e., using all pixels in its superpixel to represent the pixel, is obtained by solving a generalized Sylvester equation to extract commonality and maintain individuality of the pixel to some extent. Finally, a global projection matrix to a low-dimensional space is calculated by reducing the discrepancy between SPCRs and the original spectral features, and reducing the differences between pixels from one superpixel and increasing the differences between pixels from different superpixels simultaneously. Superior classification performances on several HSI datasets demonstrate the effectiveness of the proposed SPCRGE.
Highlights
S PECTRAL information is important for discriminating different materials, since different materials possess different spectral properties
Based on the SVM classifier, experiments of superpixelwise collaborativerepresentation graph embedding (SPCRGE) on the four hyperspectral imagery (HSI) datasets are carried out in comparison with the traditional methods, principal component analysis (PCA) [17], linear discriminative analysis (LDA) [18], and other stateof-the-art methods, collaboration-competition graph preserving embedding (CCPGE) [30], BCGDA [29], orthogonal total variation component analysis (OTVCA) [35], superpixel-based linear discriminative analysis (SPLDA) [40], SuperPCA [41], and Tensor sparse and low-rank graph-based discriminant analysis (TSLGDA) [34]. Among these dimensionality reduction (DR) methods, PCA, CCPGE, OTVCA, SPLDA, SuperPCA, and the proposed SPCRGE are unsupervised without using any label
PCA, LDA, CCPGE, BCGDA do not use any spatial information while TSLGDA utilizes spatial information via tensor, OTVCA via total variation minimization, and SPLDA, SuperPCA, and SPCRGE via superpixel segmentation
Summary
S PECTRAL information is important for discriminating different materials, since different materials possess different spectral properties. Other strategy like orthogonal total variation component analysis (OTVCA) [35] learns low-dimensional features with total variation regularization based on that the learned features should be smooth in spatial space These methods consider only specified neighborhood around a query pixel and cannot fully reflect local manifold in spectral-spatial feature space. Many classification methods have been proposed, such as Hidden Markov Random Fields-SVM (HMRFSVM) [36], set-to-set distancebased spectral-spatial classification (SD-SSC) [37], superpixelbased extended random walker (SPERW) [38], and superpixel contracted graph-based learning (SGL) [39] All these methods assume that pixels within a superpixel are from one class and have achieved good classification performances.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
