Random Matrix Based Nonnegative Sparse Representation for Hyperspectral Image Classification

Abstract:

Considering the limitations of regular classification model using Sparse Representation (SR), this paper proposes an innovative model named Random Matrix-Nonnegative Sparse Representation (RM-NSR) to improve the classification results of hyperspectral imagery. The RM-NSR model introduces a random matrix inspired by random projection to improve the Restricted Isometry Property (RIP) condition of measurement matrix in the regular SR model. The new model also considers the non-negativity of reconstructed sparse coefficient vectors. Based on Urban and PaviaU hyperspectral datasets, three different schemes in the RM-NSR model are utilized to recovery the sparse efficient and the classification results are compared with that of the regular SR model. Experimental results show that the RM-NSR model obviously outperforms the regular SR model in the average classification accuracies (ACAs). Furthermore, the relationships between projected dimension of random matrix and the ACAs show that a greater projected dimension guarantees the improvement of ACAs by the RM-NSR model.