Data augmentation-based evidential generative adversarial network for open-set hyperspectral image classification

Hyperspectral image (HSI) classification has received more and more attention due to its wide applications in the field of remote sensing. In recent years, numerous deep learning-based methods have been proposed to classify objects with known classes; however, they fail to effectively identify new samples belonging to unknown classes due to be lacking of prior knowledge about unknown categories. Although some attempts have been devoted to classify unknown samples with certain rules, less works commit to augment training data, including samples and features. Moreover, there is no work on direct classification of all classes. To solve these issues, a novel data augmentation-based evidential generative adversarial network is proposed to enhance the performance of the open-set classification of HSI. The core idea is summarized as data augmentation and direct classification. For the former, we first design an extended generator by using Kullback–Leibler divergence and bands substitution to enhance the ability of generating fake samples with authenticity and diversity. Then, feature representation is learnt by optimizing new losses based on supervised contrastive learning, which ensures the features with the same class are more clustered and the features with different classes are more scattered. In this way, it forms a clear boundary of all categories. For the latter, an evidential discriminator based on evidence deep learning is proposed to directly identify unknown objects by converting the uncertainty into the probability of the unknown classes instead of setting user-defined thresholds. Extensive experiments demonstrate that our proposed network has a competitive performance both in closed-set and open-set classification. Moreover, it outperforms the state-of-the-art open-set classification methods.