Locality-preserving discriminant analysis and Gaussian mixture models for spectral-spatial classification of hyperspectral imagery

Traditional hyperspectral image classification typically uses raw spectral signatures or simple spatial characteristics such as textural features without considering the correlation between spectral and spatial information. In this paper, we propose a spectral-spatial hyperspectral image classification based on a structured multi-modal statistical model. A 3D wavelet transform is employed to extract relevant features from every pixel and its neighboring pixels; these features quantify local orientation and scale characteristics. Local Fisher's discriminant analysis is then used to project this high-dimensional wavelet coefficient space onto a lower-dimensional subspace while preserving the multi-modal structure of the statistical distributions. The proposed classification framework then employs a Gaussian mixture model classifier in this feature subspace. Experimental results at hyperspectral image-classification tasks show that the proposed approach substantially outperforms traditional methods.