Salient feature extraction method for hyperspectral image classification

Aiming at the problem of hyperspectral image classification, a salient feature extraction method is proposed. Firstly, the method uses a superpixel segmentation algorithm to divide three adjacent bands of hyperspectral image into several small regions. Then, the salient features of different regions are calculated based on the small regions. Finally, the sliding window method with a size of 3 steps is used along the spectral direction to obtain the salient features of all bands. The extracted saliency features are further combined with the spectral features, and the combined features are fed into a support vector machine for classification. The classification experiments were carried out on three hyperspectral image datasets including Pavia University, Indian Pines and Salinas. The experimental results show that compared with the traditional spatial feature extraction method and the convolutional neural network based methods, the extracted salient features can obtain higher classification accuracy. Combining salient features and spectral features can further improve classification accuracy. © 2019, Surveying and Mapping Press. All right reserved.