Light-Weight Self-Supervised Contrastive Learning Network For Small Sample Hyperspectral Image Classification

In recent years, the rapid development of deep learning has greatly improved the performance of hyperspectral image (HSI) classification. However, due to the high cost of obtaining high-quality HSI data and the large amount of manpower and material resources required for labeling, the small sample problem has always been a challenge. To solve it, a spatial-spectral data augmentation method is designed to expand the limited data set and generate new training samples. Simultaneously, a light-weight self-supervised contrastive learning network is conducted to extract spatial-spectral feature and focus on the relative relationship between samples to improve the sample discrimination. Experimental results on 2 public HSI data sets demonstrate that the proposal could achieve better performance than some state-of-the-art contrastive learning methods facing the challenge of small sample.