An Effective Feature Extraction Approach Based on Spectral-Gabor Space Discriminant Analysis for Hyperspectral Image

With the development of remote sensing technology, hyperspectral sensors can record reflectance of ground objects in hundreds of bands, which undoubtedly brings great benefits for hyperspectral remote sensing image that contains abundant information of the land covers. At present, some hyperspectral images are contaminated by the noise and a lack of spatial information, this phenomenon hinders the improvement of classification accuracy. To further mine the spatial features of hyperspectral image, we propose an effective feature extraction method based on Spectral-Gabor space discriminant analysis (SGDA). The framework of SGDA is roughly divided into four steps: firstly, we obtain p(i){p(i)vertical bar 1 <= i < d} principal components by PCA, where d denotes the number of features; secondly, we filter the p(i)principal components with Gabor filter on five different scales and eight different orientations and obtain Gabor spatial features; thirdly, incorporating the spectral features of original hyperspectral data into the Gabor spatial features to form the Spectral-Gabor space features F; finally, we project the fusion features to a low-dimensional subspace, and then maximize the Spectral-Gabor space between-class scatter matrix (S-b(SG)) and minimize the Spectral-Gabor space within-class scatter matrix (S-w(SG)) at the same time inspired by the idea of Fisher line discriminant analysis. Also, in the above fusion process, the proportion of spectral and spatial information can be controlled by using a penalty factor alpha. Experimental results on three hyperspectral datasets show the better classification performance of the SGDA method than the state-of-art methods on the condition of small sample size by Maximum Likelihood Classifier (MLC).