Cross-Domain Self-Training Network With Prototype Rectification for Few-Shot Hyperspectral Image Classification

In recent years, the classification research of hyperspectral image (HSI) with scarce labeled samples has made significant progress, demonstrating strong performance when labeled samples are limited. However, current methods often overlook the importance of class prototypes. Additionally, while Conditional Domain Adversarial Adaptation (CDAA) is employed to mitigate domain shift, it fails to account for the varying degrees of data distribution discrepancies across different classes, leading to suboptimal alignment. To address these issues, a Cross-domain Self-Training Network with Prototype Rectification (CSNPR) for few-shot HSI classification is proposed. First, a Three-dimensional Ghost Attention Network with Brownian distance covariance (TGAN-B) is introduced to improve the representation learning of HSI, generating more effective class prototypes. Next, an adaptive weighted conditional domain adversarial adaptation (ACDAA) strategy is proposed, which assigns weights based on the domain discrepancy for better alignment. Finally, a Pseudo Label Selection (PLS) module is developed to select the most confident pseudo-labeled samples from each class in the target domain, thereby enriching the training set. The experimental findings derived from the analysis of three publicly available HSI datasets, prove that the CSNPR method outperforms existing state-of-the-art (SOTA) methods, achieving superior and stable performance even with sparse labeled samples. Our source code is available at https://github.com/Yzh190-hue/CSNPR.