Pseudo-Label Guided Sparse Regression With Spectral and Spatial Regularization for Hyperspectral Band Selection

Band selection, aiming at screening representative spectral bands and eliminating redundant information, has long been a popular topic in hyperspectral imagery (HSI) processing, and has garnered growing concern owing to the advancements in sparse representation techniques. Traditional sparsity-based methods are frequently impeded by the issue of inadequate or even unattainable training samples. Moreover, these approaches may fall short in thoroughly investigating the spatial structural information and spectral contextual information. To this end, this article proposes a new unsupervised band selection scheme, namely, pseudo-label guided sparse regression with spectral and spatial regularization (PSR2BS), which embeds band selection into an unsupervised sparse regression model. In particular, a pseudo-label matrix is jointly learned to serve as a discriminative cluster indicator, during which it guides the projection matrix in selecting informative bands. To leverage the spatial information within HSI, an image is segmented into different, distinct homogeneous regions to generate representation samples, wherein the local structural information is also explored through spatial regularization. Furthermore, a spectral regularization term is introduced, making full use of prior information regarding similarity within spectral bands. To solve the proposed model, an effective and efficient iterative optimization algorithm is developed. Our extensive experiments on classification and anomaly detection across six real HSI datasets clearly demonstrate the superior performance of the proposed PSR2BS compared with state-of-the-art competitors.