|
Abstract
|
One of the most important processes on hyperspectral images (HSI) is classification. Extracting low-dimensional subspace structures and obtaining discriminative representations are among the significant goals in the hyperspectral image classification. The low-rank representation (LRR) method obtains low-dimensional data structures; however, it cannot accurately find the intrinsic and discriminative structure in a nonlinear structure. In this paper, the cluster-based graph regularized LRR with dictionary learning (CLRRDL) is proposed for HSI classification. The data manifold structure is combined with the LRR model in the form of manifold regularization to improve LRR. The information of segments in each cluster is used in the proposed method to obtain the graph. Also, the geodesic distance is used to express the graph similarity with the nonlinear structure. In this model, each pixel is expressed as a linear combination of dictionary components. Moreover, rather than working on the entire image, it first clusters the image by considering the pixels in a cluster to be often composed of the same materials and their linear combination to be limited to common components from the dictionary. Then, the proposed method is separately applied to each cluster. Highly discriminative features are obtained using this factor and dictionary learning. The experimental results show that the proposed method performs better compared to other state-of-the-art methods.
|