In order to take full advantage of high spectral information and to reduce the decline of classification accuracy
resulted from data redundancy, a dimensionality reduction algorithm called block non-negative sparsity reconstruction
embedding is proposed. Firstly, an ordinary over-complete dictionary is converted into an over-complete block dictionary.
Then, a block non-negative sparsity reconstruction weight matrix is obtained through computing the minimum reconstruction
error of the sample corresponding to each over-complete block dictionary. Finally, in the phase of low-dimensional
embedding, the global optimum hyperspectral data in a low-dimensional subspace can be obtained by minimizing the local
and maximizing the non-local non-negative sparse information of the hyperspectral data simultaneously. Experimental results
of three groups of hyperspectral data validate the feasibility and effectiveness of the proposed algorithm.