Structure-Preserved and Weakly Redundant Band Selection for Hyperspectral Imagery

In recent years, sparse self-representation has achieved remarkable success in hyperspectral band selection. However, the traditional sparse self-representation-based band selection methods tend to neglect the spatial distribution differences and spectral redundancy between heterogeneous regions. Consequently, the uniform band subset obtained cannot accurately express the key features of various region-specific objects. In this context, this article proposes the structure-preserved and weakly redundant (SPWR) band selection method for hyperspectral imagery (HSI). Initially, to preserve the spatial structure of HSI, heterogeneous regions are generated by superpixel segmentation. This process simulates the actual distribution of ground objects and captures the spectral feature differences from heterogeneous regions, thus adapting the sparse self-representation to diverse land cover types. Subsequently, given that the different objects between heterogeneous regions have different sensitive bands, a series of region-specific multimetric hypergraphs are constructed to more accurately express the multivariate adjacencies between bands for each region. Significantly, a new spectral similarity measure that integrates both the spectral distance and physical distance is elaborately utilized to group bands into various hypergraphs. Finally, a consensus matrix is designed to fuse multiple coefficient matrices carrying the local spatial-spectral information of HSI, thereby selecting the subset of bands for a unified characterization of HSI and achieving the complementarity of multiple regions. Extensive comparison experiments on four real-world datasets demonstrate that the proposed method SPWR can efficiently select representative bands and outperforms other comparison methods.