A Novel Supervised Method for Hyperspectral Image Classification with Spectral-Spatial Constraints

In this paper, a new supervised classification method, combining spectral and spatial information, is proposed. The method is based on the two following facts. First, a hyperspectral pixel can be sparsely represented by a linear combination of the dictionary consists of a few labeled samples. If any unknown hyperspectral pixel lies in the subspace spanned by some labeled-class samples, it will be classified to this labeled-class. And this is to solve a fully constrained sparse unmixing problem with the l2 regularization and the criterion of classification is relaxed to be determined by the largest value of sparse vector whose nonzero entries correspond to the weights of the labeled samples. Second, since the nearest neighbors probably belong to the same class, a spatial constraint is introduced. Alternating direction method of multipliers (ADMM) and the graph cut based method are then used to solve the spectral-spatial model. Finally, two real hyperspectral data sets are used to validate our proposed method. Experimental results show that the proposed method outperforms many of the state-of-the-art methods.