Hyperspectral Band Selection With Dynamic Graph Enhancement and Hierarchical Feature Fusion

Graph convolutional networks (GCNs) have been widely used in hyperspectral band selection (BS) and exhibits great potential. However, the substantial redundant information inherent in hyperspectral imagery and the fixed graph structure hinder the ability of GCNs to effectively capture the complex inter-band relationships. Moreover, most existing methods rely on a single criterion during the BS phase, often leading to biased results. To address these issues, this paper proposes a dynamic graph enhancement and hierarchical feature fusion (DEHF) model for efficient and robust BS. Specifically, the model first utilizes an autoencoder (AE) to extract key latent features of ground objects, thereby enhancing the contribution of high-quality representation in the BS process. To overcome the limitations of conventional graph construction, we introduce a two-stage graph optimization strategy comprising initial graph optimization and dynamic enhanced GCN (DEGCN). This strategy adaptively updates both the graph structure and node representations, enabling more accurate modeling of the complex local-global relationships among spectral bands. In order to further enhance feature integration of the model, a hierarchical fusion mechanism is designed to fuse multi-source features extracted by the AE and DEGCN across different spatial scales, thereby emphasizing the responses of salient bands. Finally, to comprehensively evaluate band importance, a multi-dimensional priority criterion is developed by integrating the information entropy and class discriminability, which guides spectral clustering to yield a more informative and representative band subset. Experimental results on three real hyperspectral datasets demonstrate the superior performance of DEHF compared to state-of-the-art methods. The source-code is available at https://github.com/ShangXD1994/DEHF-for-hyperspectral-band-selection.