Domain-Invariant Progressive Enhancement for Cross-Scene Classification of Hyperspectral Images

Cross scene classification is a significant direction in hyperspectral image processing, and the differences in distribution of the source and target domains render this task extremely formidable. Domain adaptive methods require access to target data when aligning distributions between domains, while domain generalization methods allow models trained on the source domain to be immediately utilized in unknown domains, with a wider range of applicability. Currently, most domain generalization methods mainly focus on generating simulated samples related to the source domain, and using some strategies such as adversarial learning to obtain domain invariant features. Nevertheless, these methods fail to clearly tell apart domain invariant features from domain specific features, which makes the model prone to fitting specific features and difficult to learn invariant features effectively. Inspired by disentangled representation learning, we propose a novel domain invariant progressive enhancement network (DIPEN) for cross scene classification of HSI. Firstly, a domain invariant progressive enhancement module is proposed, which is based on a frequency domain disentanglement layer to gradually separate domain invariant features from shallow to deep layers. Then, a re-entanglement module is carefully designed for deep fusion of original features and domain invariant features, thereby improving generalization ability. Finally, a frequency-aware feature aggregation module is constructed for more accurately extracting features that are conducive for classification. Some experiments carried out on three public datasets indicate that the proposed method outperforms some state-of-the-art approaches with a large margin.