Structured Multi-Evidence Learning for Hyperspectral Image Clustering

Abstract

The hyperspectral images (HSIs) record subtle spectral responses of land-cover materials over a large number of contiguous bands, providing rich information for unsupervised land-cover analysis. However, local environmental variations, spectral ambiguity, and band redundancy often make the underlying clustering structure unreliable. The existing HSI clustering methods usually start from a single input evidence source, such as spectral features or current embeddings, and then improve clustering stability through structural priors, graph relations, pseudo labels, or augmented views. Although useful, these signals are still largely derived from the same base evidence source and may inherit its bias when the induced clustering geometry is unreliable. To address this issue, we propose a structured multi-evidence learning (SMEL) framework for lightweight hyperspectral clustering. Instead of merely deriving constraints or perturbed views from the same base evidence source, the proposed method explicitly organizes complementary evidence from spectral characteristics, region context, and region prototype relations before assignment learning, allowing different spectral–spatial cues to compensate for each other. Benefiting from this structured evidence organization, the proposed lightweight multi-evidence assignment learning module establishes semantic consensus among complementary evidence sources without relying on complex deep encoders or elaborate handcrafted priors. Experiments on several benchmark hyperspectral datasets show that the proposed structured multi-evidence framework achieves stable and competitive clustering performance with only a few thousand learnable parameters. Our code is available at https://github.com/NuclearLemon/SMEL.