Invariant semantic domain generalization shuffle network for cross-scene hyperspectral image classification

Abstract

Cross-scene hyperspectral image classification is currently receiving widespread attention. However, domain adaptation-based methods usually perform domain alignment by accessing specific target scenes during training and require retraining for new scenes. In contrast, domain generalization only trains using the source domain and then gradually generalizes to unseen domains. However, existing methods based on domain generalization ignore the impact of domain invariant semantics on the invariant representation of the domain. To solve the above problem, an invariant semantic domain generalization shuffle network for cross-scene hyperspectral image classification is proposed, which follows a framework on the generative adversarial network. Feature style covariance in style and content randomization generator with invariant semantic features is designed to safely extend the style and content of features without changing the domain invariant semantics. We proposed a spatial shuffling discriminator, which can reduce the impact of special spatial relationships within the domain on class semantics. In addition, we proposed a dual sampling direct adversarial contrastive learning strategy. It uses a dual sampling in two-stage training design to prevent the model from lazily entering the local nash equilibrium point. And based on dual sampling, directly adversarial contrastive learning using clearer contrastive samples is used to reduce the difficulty of network training. We conduct extensive experiments on four datasets and demonstrate that the proposed method outperforms other current domain generalization methods. The code will be open source at https://github.com/jixiangyu0501/ISDGS.