Embracing Hierarchical Classification: A Multifeature Space Hierarchical Network for Hyperspectral Image Classification

Classes in hyperspectral images (HSI) often exhibit inherent hierarchical structures, such as the family–genus–species hierarchy in tree species. Previous studies have shown that modeling these hierarchical structures can improve classification performance. However, existing deep learning methods often overlook such structures, limiting their ability to distinguish fine-grained categories. To address this issue, we propose a Multi-Feature Space Hierarchical Network (MFS-HiNet), which enhances fine-grained discrimination by modeling hierarchical relationships and guiding classification top-down. The framework consists of three key components: Hierarchical Structure Mining (HSM), Multi-Feature Space Classification Network (MFSCN), and Parameter Inheritance Strategy (PIS). Specifically, HSM automatically mines latent hierarchical relationships in HSI and constructs the hierarchy; MFSCN integrates multi-feature space information for node classification to improve the discrimination of subtle inter-class differences; and PIS leverages parent node parameters to guide the learning of child nodes, fully exploiting inter-level correlations. Experimental results on five HSI datasets demonstrate that MFS-HiNet outperforms existing methods in classification accuracy, validating the effectiveness of the framework and highlighting the potential of hierarchical classification. The source code will be made available at https://github.com/jin-yaoWHU/ MFS-HiNet.