Efficient superpixel-guided global-local spectral clustering for large-scale HSI

Spectral clustering is widely employed in hyperspectral image (HSI) analysis due to its well-defined framework and outstanding performance. However, its quadratic or cubic computational complexity limits its applicability in large-scale tasks. While scalable spectral methods have been developed, they often fall short in effectiveness due to insufficient structure exploration. To address this issue, we propose an efficient superpixel-guided global–local graph clustering model (ESGLGC), which enhances clustering effectiveness while preserving scalability. Specifically, the large-scale HSI is first partitioned into superpixels using the entropy rate superpixel (ERS) algorithm, significantly reducing computational complexity. Then, a superpixel-level graph convolutional subspace learning strategy is developed to capture global structure, while a multi-level neighborhood construction scheme is designed to characterize local structures. These complementary global and local graphs are then fused into a unified graph for spectral clustering. By jointly exploiting superpixel-guided global–local structure information, ESGLGC achieves a favorable balance between clustering accuracy and computational efficiency. Extensive experiments demonstrate that ESGLGC achieves superior clustering performance across various datasets compared to state-of-the-art methods.