Hyperspectral Unmixing by Convolutional Auto-encoder with Deep Subspace Clustering and Candidate Pixel Selection

The hyperspectral unmixing process is central to identifying the objects in a ground scene and quantifying their fractional abundance in each pixel. However, the high spatial resolution and the intrinsic interactions pose a challenge in object identification from hyperspectral images by non-linear unmixing. Typically the data points are generated due to non-linear, intimate mixing from non-linear subspaces. In this work, we propose a deep subspace clustering framework to identify the underlying non-linear subspaces in the initial stage and perform non-linear unmixing on the local clusters. To this aim, we proposed a deep auto-encoder network with additional total variation and spatial consistency regularization to determine the underlying non-linear mixing process from each cluster separately. Next, we identify the pixels which contain a dominant source from the latent representation obtained after the encoding stage. Subsequently, we carried out unmixing on local clusters using a linear algebraic based on the low-rank structure of the data. A detailed comparative analysis of the unmixing algorithms on three real hyperspectral images exhibits that our proposed algorithm achieves improved performance.