Clustering Techniques for Hyperspectral Images Using Joint Analysis Dictionary Learning

The problem of dealing with the hyper-spectral images in an unsupervised manner is a challenge as the data is very high dimensional and the another challenge is non-availability of lot of labelled data. Analysis dictionary learning is a very recent technique for representation learning in an unsupervised fashion. The existing approaches learn the clusters on raw data itself which limits in its capability to capture both local and global structures. In this work, we propose a clustering framework, where, instead of performing clustering on raw pixels, we use analysis dictionary coefficients for performing the clustering task. The subspace clustering presumes that the data belonging to the same cluster shall lie in the same subspace. This limitation is taken care by transformed coefficients as it doesn’t need any such constraint on the data. Our contribution in this paper is three fold. First, Instead of performing clustering on raw pixels, we use transformed coefficients from analysis dictionary as features which are further used to cluster the hyper-spectral images. We use K-means clustering, spectral clustering and subspace clustering (locally linear manifold clustering and sparse subspace clustering) for the same. Secondly, We instead of learning analysis features and applying K-means clustering technique in piecemeal way, we embed the clustering loss with the analysis dictionary learning formulation. This joint learning framework reaches convergence by alternate minimization. Third contribution is to learn from a joint subspace clustering framework (locally linear manifold clustering and sparse subspace clustering) with analysis dictionary for hyper-spectral image clustering task. Experimental results on three hyper-spectral imaging data sets Salinas, Indian Pines and Pavia University show that our proposed method outperforms the state-of-the-art and they are computationally efficient too.