Noise-adjusted sparsity-preserving-based dimensionality reduction for hyperspectral image classification

In this paper, we investigate the performance of a sparsity-preserving graph embedding based approach, called l1 graph, in hyperspectral image dimensionality reduction (DR), and propose noise-adjusted sparsity-preserving (NASP) based DR when training samples are unavailable. In conjunction with the state-of-the-art hyperspectral image classifier, support vector machine with composite kernels (SVM-CK), the experimental study show that NASP can significantly improve the classification accuracy, compared to other widely used DR methods.