SSMM: Semi-supervised manifold method with spatial-spectral self-training and regularized metric constraints for hyperspectral image dimensionality reduction

Abstract

Manifold learning is an important technique for dimensionality reduction in hyperspectral images. It maps data from high dimensions to low dimensions to eliminate redundant information. However, the existing manifold learning methods cannot effectively solve the problem of lacking label information and ignore the negative impact of dimensionality reduction on sample division. To address these, we propose a semi-supervised manifold method with spatial-spectral self-training and regularized metric constraints (SSMM) for hyperspectral image dimensionality reduction. The spatial-spectral self-training module is proposed, which learns pseudo-labels by jointly training the spatial and spectral information. This module first locates the spatial neighbors of the labeledit can adapt to different data distributions and feature samples and then sets an adaptive threshold based on the spectral features of labeled samples to filter spatial neighbors, so as to obtain the spatial-spectral neighbors as pseudo-labeled samples. In addition, to divide the sample categories while dimensionality reduction, low-dimensional manifold embedding is constructed and the metric constraint is imposed on the manifolds. Specifically, the Gaussian kernel function based on Mahalanobis distance is used to map the data into a more discriminative low-dimensional manifold embedding. At the same time, the regularized distance metric constraint is imposed on the manifold, so that samples of the same class are clustered and different classes are mutually exclusive. SSMM conducts various forms of experiments on the Houston 2013, Indian Pines, and Washington DC datasets. In the dimensionality reduction experiments, the overall accuracy of SSMM in any dimension is higher than that of other algorithms. In the classification experiments, the KAPPA coefficient of SSMM on the three data sets is improved by 1.41%, 0.61%, and 0.27% respectively. The feature extraction experiments show superior clustering performance. These experimental results demonstrate that SSMM not only effectively solves the problem of insufficient label information, but also significantly improves the classification accuracy of hyperspectral images after dimensionality reduction, which is superior to the existing manifold learning methods.