Semi-supervised spatial–spectral classification for hyperspectral image based on three-dimensional Gabor and co-selection self-training

Abstract

Benefiting from the development of hyperspectral imaging technology, hyperspectral image (HSI) classification has become a significant research direction in remote sensing image analysis. However, labeling HSI requires sufficient domain knowledge and consumes a lot of human and material resources. In addition, HSI contains rich spectral and spatial features, and how to fully extract these two joint features remains a problem worth exploring. A semi-supervised spatial–spectral method based on three-dimensional (3D) Gabor and co-selection self-training (ST) is proposed for HSI classification. First, to fully extract the spatial–spectral features, the 3D Gabor filter is used for the raw HSI data cube to generate a Gabor feature data cube with spatial–spectral joint features. Second, the Gabor feature data cube is fed into the ST process to obtain sufficient labeled samples. In the ST process, we combine the classification results and clustering results to construct the candidate sample set. Finally, a co-selection strategy is proposed to select highly confident pseudo labels from the candidate sample set, and these selected pseudo labels are added to the labeled samples set by an iteration process. The experimental results obtained with two HSI datasets with different scenes (i.e., agricultural scenes and urban scenes) show that the proposed method has a stable classification performance in a few initial labeled samples.