Improved discriminant sparsity neighborhood preserving embedding for hyperspectral image classification

Sparse manifold learning has drawn more and more attentions recently. Discriminant sparse neighborhood preserving embedding (DSNPE) has been proposed, which adds the discriminant information to sparse neighborhood preserving embedding. However, DSNPE does not investigate the inherent manifold structure of data, which may be helpful for dimensionality reduction and classification of hyperspectral image. In this paper, we proposed a new sparse manifold learning method, called improved discriminant sparse neighborhood preserving embedding (iDSNPE), for hyperspectral image classification. iDSNPE utilizes the merits of both manifold structure and sparsity property. It not only preserves the sparse reconstructive relations but also explicitly boosts the discriminating information from manifold structure of data, and the discriminating power of iDSNPE is significantly improved than DSNPE. The effectiveness of the proposed method is verified on two hyperspectral image datasets (Washington DC Mall and Urban) with promising results.