Semi-supervised Spatial-spectral Discriminant Analysis for Hyperspectral Image Classification

In order to make full use of the spatial information embedded in the hyperspectral image to improve the classification accuracy, a semi-supervised spatial-spectral discriminant analysis (S3DA) algorithm for hyperspectral image classification is proposed. According to the spatial consistency property of hyperspectral image, the intra-class scatter matrix infered from a little labeled samples preserves the spectral similarity of the same class pixels, while the spatial local pixel scatter matrix defined by the unlabeled spatial neighbors uncovers the spatial-domain local pixel neighborhood structures and the ground objects detailed distribution. The S3DA method not only maintains the spectral-domain separability of the data set, but also preserves the spatial-domain local pixel neighborhood structure, which promotes the compactness of the same class pixels or the spatial neighbor pixels in the projected subspace and enhances the classification performance. The overall classification accuracies respectively reach 81.50% and 71.77% on the PaviaU and Indian Pines data sets. Compared with the traditional spectral methods, the proposed method can effectively improve ground objects classification accuracy.