Learning Discriminative Sparse Representations for Hyperspectral Image Classification

In sparse representation (SR) driven hyperspectral image classification, signal-to-reconstruction rule-based classification may lack generalization performance. In order to overcome this limitation, we presents a new method for discriminative sparse representation of hyperspectral data by learning a reconstructive dictionary and a discriminative classifier in a SR model regularized with total variation (TV). The proposed method features the following components. First, we adopt a spectral unmixing by variable splitting augmented Lagrangian and TV method to guarantee the spatial homogeneity of sparse representations. Second, we embed dictionary learning in the method to enhance the representative power of sparse representations via gradient descent in a class-wise manner. Finally, we adopt a sparse multinomial logistic regression (SMLR) model and design a class-oriented optimization strategy to obtain a powerful classifier, which improves the performance of the learnt model for specific classes. The first two components are beneficial to produce discriminative sparse representations. Whereas, adopting SMLR allows for effectively modeling the discriminative information. Experimental results with both simulated and real hyperspectral data sets in a number of experimental comparisons with other related approaches demonstrate the superiority of the proposed method.