Weakly Supervised Hyperspectral Image Classification With Few Samples Based on Intradomain Sample Expansion

Insufficient sample is a common problem in hyperspectral image (HSI) classification and an important factor causing to low accuracy. Adding weak supervision mechanism to model training is an effective way to solve this problem. In this paper, we propose a weakly supervised HSI classification method with few samples based on intradomain sample expansion, which is anchored on the spatial correlation between samples. Firstly, to reduce the negative effects of spectral mixing on pseudolabel generation and classifier training, we introduce the continuous side window filter (SWF) to smooth the original hyperspectral images band by band. Second, in order to better balance the correctness and representativeness of the generated pseudolabels, a novel method for sample selection and pseudo-label generation is proposed in this paper. The method contains two branches, the segmentation graph based and the neighborhood relationship based. In the branch based on segmentation graphs, the polygon segmentation graph generated by the false color image is used to select the expansion samples. As a complementary branch, the branch based on neighborhood relationship exploits the neighborhood relationship and spectral similarity between samples to further select samples. Finally, in the classification stage, this paper uses the broad learning system as a classifier to obtain the classification graph. Experiments on three publicly available datasets show that the method in this paper can effectively achieve sample expansion and improve the classification accuracy and stability of hyperspectral images in the case of insufficient labeled samples.