Robust Unsupervised Hyperspectral Band Selection via Global Affinity Matrix Reconstruction

Unsupervised band selection is fundamental to alleviate the curse of dimensionality for hyperspectral imagery. Although many research works have been developed, it is still a challenging problem to improve the poor classification performance with a small set of selected bands due to the lack of discriminative information that simultaneously distinguishes multiple classes. Moreover, traditional methods utilize the predefined graph representation for different hyperspectral image (HSI) datasets that may result in degraded generalization performance. To address this issue, this article proposes an unsupervised band selection method via global affinity matrix reconstruction for HSI classification, referred to as GAMR briefly. To explore the discriminative information from HSIs, GAMR formulate a convex optimization problem to iteratively refine the randomly generated pseudolabels. Meanwhile, the similarity between band pairs is estimated by adaptively reconstructing a global affinity matrix from several local graph representations. To solve the proposed formulation, an alternating optimization algorithm is designed to search for the optimal solution. Experimental results on three HSI datasets demonstrate the effectiveness of the proposed method and the superiority over seven state-of-the-art methods.