Spatial-Spectral Semi-Supervised Local Discriminant Analysis for Hyperspectral Image Classification

In traditional hyperspectral image classification algorithm based on feature extraction, spectral information is usually considered while spatial information is ignored. To address this problem, a hyperspectral image classification algorithm based on semi-supervised spatial-spectral local discriminant analysis (S³ELD) and spatial-spectral nearest neighbor (SSNN) classifier is proposed in this paper. Combining the spatial consistency of hyperspectral images and on the basis that the discriminant information of the labeled samples is used to maintain the separability of the data set, we define the spatial local pixel scatter matrix to preserve the spatial-domain neighborhood structures of pixel. A similarity measure method based on the spatial-spectral distance is then proposed to discover the local manifold structure and to construct SSNN. S³ELD algorithm not only reveals the local geometric relations of the data set but also enforces the compactness of the spectral-domain same class pixels and the spatial-domain local neighbor pixels in the low-dimension embedding space. Combining SSNN to classify, the classification accuracy is further enhanced. The experiments on the PaviaU and Salinas data sets show that the overall classification accuracy of S³ELD algorithm reaches 92.51% and 96.29%, respectively. Compared with several existing algorithms, the proposed algorithm can efficiently extract the information of discriminant characteristics and obtain higher classification accuracy.