Hypergraph-regularized low-rank tensor subspace clustering for hyperspectral band selection

Abstract

Band selection for hyperspectral images (HSIs) is an effective strategy to reduce data redundancy by selecting few representative bands, which boosts subsequent HSI applications. In this paper, we propose a novel hypergraph regularized low-rank tensor subspace clustering (HyGLRTSC) method for hyperspectral band selection. In our model, CANDECOMP/PARAFAC (CP) decomposition is introduced to exploit the intrinsic correlation. Orthogonal constraints are performed on the spatial modes to explore the spatial structure, and a low-rank constraint is imposed along the spectral mode to capture the global latent representation. Moreover, a hypergraph constraint is incorporated to capture the local manifold structures among bands, promoting the subspace-wise grouping effect. An efficient algorithm is also proposed to solve the non-convex optimization problem. Finally, the representative bands are selected via spectral clustering in the subspace constructed by the proposed model. Experimental results verify that our model surpasses the state-of-the-art methods.