Hyperspectral Image Clustering by Superpixel- Based Low-Rank Constrained Bipartite Graph Learning

Large-scale hyperspectral image (HSI) clustering is a challenging and attention-grabbing problem. Most existing methods are difficult to solve high dimensionality and computational complexity in HSI. In this letter, an innovative algorithm, termed superpixel-based low-rank constrained bipartite graph (SLCBG) learning is introduced, to efficiently achieve large-scale HSI clustering. Specifically, to solve the problem of massive data and high computational memory of large-scale HSI, superpixel segmentation is adopted to obtain superpixels in this letter. The spectral distance is measured using Mahalanobis distance (MD) to better obtain the overall distribution features of HSI. Second, k-means is used to generate anchors based on the superpixel data, and an anchor graph is constructed for each view to better represent the whole HSI. Finally, tensor Schatten-p norm and nuclear norm minimization constraints are introduced to search for the low-rank structure graphs within and between views in HSI, thus obtaining a consensus graph with clear connectivity components for realizing clustering. Experimental results on three HSI showed the superiority of SLCBG with OA of 49.20%, 75.48%, and 68.59%, respectively.