Unsupervised Dimensionality Reduction With Multifeature Structure Joint Preserving Embedding for Hyperspectral Imagery

Graph embedding is an effective method that has shown superiority in dimensionality reduction (DR) for hyperspectral imagery (HSI) due to its ability to characterize the intrinsic geometric structure of the data. However, it may ignore some feature information, and the performance of the single model may result in poor classification after DR. In this article, a novel unsupervised DR method, termed multifeature structure joint preserving embedding (MFS-PE), is proposed for hyperspectral image classification. At first, a spatial-spectral model is designed based on the cooperative representation theory, which exploits the potential spatial and spectral features. Then, a neighborhood-constrained model is constructed by implementing sample augmentation through superpixel segmentation, and superpixel labels are used in local enhancement for the spatial-spectral model. Next, a k-nearest neighbor selection method is devised in the local neighborhood-constrained model to select the most suitable neighbors. Finally, both models that can maximize the total scatter of the hyperspectral data to exploit global features are combined to produce an optimal projection for DR. MFS-PE combining multiple feature information can effectively reveal the intrinsic structure of HSIs, and experiments on three publicly available HSI datasets demonstrate that it can offer better classification results compared to the state-of-the-art DR methods.