<p>Data augmentation plays a crucial role in hyperspectral image classification (HSIC) modeling, particularly when dealing with label noise. However, most existing data augmentation methods generate transformations randomly or do not adaptively generate augmentation policies with label noise resistance. These methods improve the performance of classification under the training samples with label noise limitedly. In this paper, a data augmentation network based on transformation prediction (DANTP) is proposed for robust HSIC with label noise. Different from existing methods, a data augmentation policy network (DAPN) is designed. DAPN generates augmented data by dynamically applying transformations such as spectral shift and spatial flip according to the predicted importance of spectral and spatial transformations of the input data. Then, a data augmentation prediction constraint (DAPC) mechanism is proposed to establish optimization constraints for the transformation policy and its parameters, enabling the augmented data to be more reasonable and effective. Experiments conducted on four publicly available datasets show that compared with the relative state-of-the-art models, our proposed model outperforms these models on two metrics OA and Kappa under the training samples with label noise, while achieving competitive performance on AA. This validates that DANTP is more effective.</p>