Multigranularity Manifold Hybrid Adaptive Near-Neighbor Preserved Clustering for Hyperspectral Image Band Selection

Band selection is an effective method to reduce the dimensionality of hyperspectral images (HSIs). However, most existing band selection methods usually consider each band as an overall feature and neglect to analyze the intrinsic discriminative components of each band. Worse still, these methods also assume that the spatial features of bands in HSIs are distributed only in Euclidean or non-Euclidean space, which may lead to the limited discriminability of the selected bands. To address these problems, a novel multi-granularity manifold hybrid adaptive near-neighbor preserved (MMANP) clustering method is proposed for band selection. Specifically, we first transform the problem of analyzing the intrinsic discriminative components of bands into a feature decoupling problem and propose a fine-grained manifold collaborative (FGMC) feature decoupling model, which can obtain spatially discriminative features and important features with the fine-grained manifold structures in each band. Then, the hybrid adaptive local neighborhood (HALN) block clustering model is proposed to interact with the spatial effective information of the obtained fine-grained important features of each band in the Euclidean and non-Euclidean spaces, which helps to improve the HSI’s adaptive clustering of similar bands. Next, multi-view cross-granularity learning (MVCL) feature selection model is proposed to harmoniously integrate information at different levels of granularity in HSIs into a unified framework to perform adaptive band selection for each cluster from the perspectives of both the different subsets and the whole HSI. Extensive experimental results on four datasets demonstrate that MMANP outperforms other state-of-the-art comparative methods.