Local–Global Spectral–Spatial Dual Deep Subspace Clustering Network for Hyperspectral Images

Deep subspace clustering (DSC) has become increasingly popular for hyperspectral image (HSI) interpretation, achieving remarkable accomplishments. However, the existing methods overemphasize spatial pattern mining but ignore the importance of spectral information, and insufficiently explore the underlying mechanism of self-expression learning, which results in limited clustering accuracy. To address these issues, this article proposed an innovative local–global spectral–spatial dual DSC (LGDSC) network for HSIs, which makes full use of the spectral and spatial discriminative information and integrates the advantages of local and global expressions to boost clustering performance. First, an effective dual convolutional autoencoder (CAE) architecture was developed to fully mine complementary local-spectral and global-spatial features through a 1-D CAE and 2-D CAE, respectively, with a shared self-expression module embedded for pixel affinity learning. Then, a novel selective affinity aggregation (SAA) module was put forward to precisely fuse favorable affinities derived by differentiated features to generate a more indicative and clustering-friendly adjacency matrix. Ultimately, a valid joint loss function with biliteral constraints was designed to guide local-spectral and global-spatial feature learning in the training process. Extensive experimental results on four public datasets clearly illustrate the significant advantage of LGDSC over the other state-of-the-art (SOTA) alternatives and its promising prospects in practical applications.