Collaborative Graph-Based Discriminant Analysis for Hyperspectral Imagery

In previous work, a sparse graph-based discriminant analysis was proposed for when labeled samples are available. Although an affinity-based graph itself may not necessarily enhance the disciminant power, the discriminant power can truly be improved when an affinity matrix is retrieved from labeled samples. Additionally, a sparsity-preserving graph has been demonstrated to be capable of providing performance superior to that of the commonly used k-nearest-neighbor graphs and other widely used dimensionality-reduction approaches in the literature. Deviating from the concept of sparse representation, a collaborative graph-based discriminant analysis is proposed, originating from collaborative representation among labeled samples whose solution can be nicely expressed in closed form. Experimental results demonstrate that the proposed collaborative approach can yield even better classification performance than the previous state-of-the-art sparsity-based approach with much lower computational cost.