Unsupervised feature extraction for hyperspectral imagery: A new approach based on higher-order networks

For the high-dimensional nature of hyperspectral imagery and the limitation of the small number of samples, this paper presents a new method of unsupervised feature extraction to transform hyperspectral imagery into complex networks, called the image filter visibility graph (IFVG) from the image visibility graph (IVG). Each hyperspectral image (HSI) in the image datasets is first transformed into an associative complex network series. Then, calculate the set of explanatory variable vectors through the cycle ratio in higher-order networks. Finally, HSI classification can be successfully achieved by the common classifier, experimentally validated on the classic datasets of the Salinas-A scene and the Salinas scene. Numerical experiments show that the proposed IFVG model and the support vector machine classifier, in the two datasets, have overall accuracies of 99.65%, and 98.52%, respectively. The results are better than those reported in existing literature based on the same datasets, easily extending to more fields. • The constructed high-order networks improve the accuracy of the HSI classification results. • The network features reflect the spatial and spectral information and reduce the dimension of HSI. • The proposed method is suitable for classical HSI data unsupervised feature extraction. • It does not require to manually set parameters on extracting the HSI to the complex networks.