Spectral-spatial Classification of Hyperspectral Image Based on Locality Preserving Discriminant Analysis

In this paper, a spectral-spatial classification method for hyperspectral image based on spatial filtering and feature extraction is proposed. To extract the spatial information that contain spatially homogeneous property and distinct boundary, the original hyperspectral image is processed by an improved bilateral filter firstly. And then the proposed feature extraction algorithm called locality preserving discriminant analysis, which can explore the manifold structure and intrinsic characteristics of the hyperspectral dataset, is used to reduce the dimensionality of both the spectral and spatial features. Finally, a support vector machine with a composite kernel is used to examine the performance of the proposed methods. Experiments results on a hyperspectral dataset demonstrate the effectiveness of the proposed algorithm in the classification tasks.