Dynamic Adaptive Iterative Generative Adversarial Network for Hyperspectral Image Classification With Class Imbalance

Small-sample learning improves the problem of limited labeled samples in hyperspectral image (HSI) classification to a greater extent, but still suffers from the severe problem of class imbalance, where minority classes are poorly learned and classified, while majority classes are prone to more easily overfitting. Therefore, this article proposes a dynamic adaptive iterative generative adversarial network (DAI-GAN), which combines dynamic balancing generation with feature adaptive optimisation module to achieve a joint enhancement of the quality and diversity of generated samples on the basis of high-quality pre-expansion of data. The proposed local–global quality-aware pre-expansion network effectively fuses local and global information through dynamic modeling, and expands initial training data at multiple levels based on a quality-aware mechanism. Subsequently, dynamic adaptive generative balancing model generates and optimizes diversified samples in iterative training, and introduces dynamic class imbalance loss constraints to dynamically adjust the focus on minority classes to further optimize the quality and distributional balance of the generated samples, which ultimately alleviates the class imbalance problem. The experimental results show that DAI-GAN significantly improves the classification performance on three classical datasets with large differences in class balance, respectively. Compared with the existing methods, DAI-GAN improves AA and OA by 2% –3%, and reduces SDUA by 2% –3% on all three datasets, which proves that it effectively tackling the challenges of small-sample and class imbalance in HSI classification.