Hyperspectral Imagery Classification Based on Sparse Feature and Neighborhood Homogeneity

In hyperspectral image classification, it is important to make use of the rich spectral information efficiently and to use the neighborhood information appropriately to alleviate the ‘salt and pepper noise pixel’. This paper presents a new hyperspectral image classification method based on Sparse Feature and Neighborhood Homogeneity (SF-NH). The core idea of SF-NH is to use sparse feature to express the hyperspectral image, and then the classification results preliminarily obtained by the Support Vector Machine (SVM) are revised by the neighborhood homogeneity. Experimental results on two classical hyperspectral data (i.e., Indian Pines, Saunas data) show that the proposed SF-NH method can greatly improve the classification accuracy.