Self-Supervised Contrastive Learning Residual Network for Hyperspectral Image Classification Under Limited Labeled Samples

Abstract

Recently, deep learning methods have achieved impressive results on hyperspectral image (HSI) classification. However, there is a significant challenge in HSI classification tasks in the limited number of labeled samples. Since obtaining labels for HSIs is an expensive and time-consuming task, the number of labeled samples during the training phase is relatively small. The performance of these deep learning methods may be limited when the labeled samples is limited. To solve the small-sample HSI classification problem, a novel approach by integrating self-supervised contrastive learning with residual networks is proposed in this paper. The proposed network firstly adopts a spatial-spectral data augmentation strategy to expand the limited HSI data set and generate new positive and negative sample pairs for training. Then a spectral dimension reduction residual network is conducted to relieve the vanishing gradients and extract spatial-spectral feature on limited labeled samples. The scarcity of labeled data often hinders the effectiveness of traditional classification models. Therefore, a self-supervised contrastive learning framework is designed to focus on the relative relationship between samples to make full use of limited HSI sample information. Experimental results on 2 public HSI datasets demonstrate that the proposal can achieve better performance than existing state-of-the-art methods when training samples is limited.