Spatial-Spectral Classification of Hyperspectral Images Using Image Geometric Moments and Genetic Algorithm

Hyperspectral images are widely used in various fields such as agriculture, geology and mining, urban management, military, target detection. Classification is one of the most important fields of hyperspectral data processing which traditionally performed with spectral features. Various studies have shown that the use of spatial features along with spectral features can increase the accuracy of the classification. In this research, a method has been developed for the spatial-spatial classification of hyperspectral images. In this method, after a feature extraction step based on the minimum noise fraction (MNF) method, the geometric moments as spatial features are generated from first few MNF components in the different window seizes. In the next step, these features are stacked with spectral features and genetic algorithm is used to select the appropriate combination. Finally, a majority voting filter is used in order to remove the wrong pixels in each class, and increasing the homogeneity of labels and classification accuracy. Using the MNF transform to generate geometric moment features of the image, the use of genetic algorithm for selecting appropriate spectral-spatial features and post-processing step based on the majority voting filter are the innovative points of this paper. The results of the implementations on a real hyperspectral image from the semi-urban agricultural area named Indian Pines show that the proposed algorithm can reach the classification accuracy above 94% which is 40% better than conventional method.