Collaborative Label Propagation-Based Semisupervised Linear Discriminant Analysis for Hyperspectral Images

Based on collaborative representation (CR), we present a novel semisupervised dimensionality reduction (DR) method termed collaborative label propagation-based semisupervised linear discriminant analysis (CLP-SLDA) for hyperspectral images (HSIs). The new method needs three steps to obtain the optimal projection vector. First, CLP-SLDA utilizes the collaborative label propagation (CLP) technique to obtain the weak labels of the unlabeled samples and the confidence scores of the weak labels. Second, a novel weight matrix is constructed based on the known labels, newly acquired weak labels and confidence scores of the weak labels. Third, the newly obtained weight matrix is utilized to learn the optimal transformation vectors to achieve DR for HSIs. Compared to some other state-of-the-art semisupervised DR methods, our proposed CLP-SLDA can acquire the highest classification accuracy in 9 of 13 classes for KSC and four of the nine classes for PU and achieve the best performance in AA, OA, and
$\kappa $
.