Spatial–spectral semisupervised classification based on teaching–learning-based optimization for hyperspectral image

Abstract

For hyperspectral images, most traditional semisupervised classification algorithms ignore spatial information. An effective hyperspectral spatial–spectral semisupervised classification algorithm based on teaching–learning-based optimization (S3C-TLBO) is proposed. In the algorithm, two aspects were used to expand the training labeled set: first, the spatial neighborhood samples of the existing labeled samples with high confidence were labeled to expand the labeled training samples. Second, the margin sampling (MS) combined with improved TLBO was exploited to quickly select the unlabeled samples near the classification hyperplanes from the unlabeled sample set. In addition, spatial neighborhood information was incorporated into kernel function to train a spatial–spectral support vector machine, which was used to determine the category of the test sample comprehensively. The image noise points were removed by smoothing the classification result with neighborhood information. Two sets of experiments on two hyperspectral datasets showed that the proposed algorithm can produce greater classification accuracy and reduce noise points by organically using the combinative spatial information. In particular, the S3C-TLBO can produce better classification result in few labeled samples, which is a major difficulty in hyperspectral image classification.