Superpixel-Level Hypergraph Regularized Multiview Clustering for Hyperspectral Images

The hyperspectral image (HSI) clustering has recently attracted remarkable attention. However, only a few HSI clustering methods focus on designing methods using multi-view data, where each feature type can be regarded as a specific view of data. Moreover, they typically focus on small HSIs; the applicability to large HSIs is underexplored. In this article, we proposeSuperpixel-levelHypergraphRegularizedMulti-viewClustering (SHRMC), which contains three main stages. First, we segment an HSI into several superpixels and extract the superpixel-level feature for each feature type. Second, to fully explore the complex geometric structure, the superpixel-level hypergraph on each feature type is constructed by building the shared hyperedge between each superpixel and its neighborhoods. Third, with the resulting superpixel-level features and hypergraphs on multi-feature data, we propose a novel superpixel-based and hypergraph regularized multi-view clustering model for large HSIs, which can fully explore the geometric structure and capture the intra-view and inter-view relationships. We conducted comprehensive experiments to validate the superiority of our proposed SHRMC method on four datasets, includingIndian Pines,Salinas,Pavia University, andWHU-Hi-LongKou. Notably, the performance of our proposed method is significantly better than that of the second-best method, i.e., improving the overall accuracy by around 4.29%-9.94% on the four datasets.