Superpixel-based feature learning for joint sparse representation of hyperspectral images

Hyper-Spectral Images (HSI) have high dimensional data and low number of training samples. Hence classification of these images is an ill posed problem. Existence of inescapable noise makes it more difficult to distinguish between members of each classes. To overcome this problem extracting both spectral and spatial features in a more effective method can raise the accuracy of classifier. For classification of HSIs one appropriate method is endmember extraction. On the other hand applying sparse representation is a hot topic and high performance in this field. This paper presents a novel superpixel-based method for classification of hyperspectral images. The method is called S3EJSR which uses Semi-Supervised Shroedinger Eigenmaps (SSSE) to extract spatial-spectral features and create superpixels. Next a joint sparse representation (JSR)is applied for endmember extraction and determining the category of pixel is based on a learned dictionary. Finally the classification is accomplished on The AVIRIS Indian Pines dataset and accuracy of this method is determined by SVM classifier. The results show that, compared with the same methods, the proposed classification method has better performance.