Spatially Adaptive Multi-Frequency Graph Representation and Evolution for Hyperspectral Image Clustering

Hyperspectral image clustering aims to discover land-cover category structures in an unsupervised manner. Graph-based methods have shown strong potential for this task by modeling spatial topological relationships and non-Euclidean correlations among samples. However, existing graph-based clustering methods often rely on smooth propagation or introduce frequency responses through graph operators and globally shared weights, which may be insufficient to capture the spatially varying frequency demands of hyperspectral scenes. In particular, homogeneous regions usually require low-frequency consistency enhancement, whereas boundaries, mixed pixels, and fine-grained structures demand stronger high-frequency discrimination. To address this issue, we propose a spatially adaptive multi-frequency graph representation and evolution framework (SAMFGE) for hyperspectral image clustering. Specifically, we design a collaborative multi-frequency representation module to capture low-frequency homogeneity and high-frequency heterogeneity. A node- and channel-aware frequency coordination mechanism is then introduced to adaptively assign frequency preferences according to local structural properties and discriminative spectral-spatial responses. Furthermore, the learned multi-frequency cues are exploited as interpretable evidence to guide graph evolution, enabling representation learning and topology refinement to co-evolve in a frequency-aware manner. Experimental results on multiple public hyperspectral datasets demonstrate the effectiveness of the proposed framework.