A De-Stylized Expanded Domain Sample Generation Network for Hyperspectral Image Cross-Domain Classification

In practical scenarios, spectral shifts within the same class often arise from variations in imaging conditions and changes in ground object states across different hyperspectral image (HSI) domains, leading to significant domain discrepancies. To tackle this challenge, a de-stylized expanded domain sample generation network (DEDSGnet) is proposed. First, expanded domain (ED) samples are generated to enhance sample diversity. Then, joint feature learning is performed on multiple source domains (SDs) and ED samples, enabling the model to learn domain-invariant representations and improve generalization to unseen target domains (TDs). Specifically, content and style are disentangled, and content features are encoded and decoded to produce ED samples. The generated samples are further mapped back into the content space, where a consistency constraint is applied between their representations and those of the SDs, effectively reducing domain shifts caused by style variations. Subsequently, multilevel feature extraction and fusion are conducted on both SD and ED samples, followed by classification using a classifier. Experimental results on four public HSI datasets demonstrate that the proposed method achieves superior generalization and stability, outperforming existing state-of-the-art approaches.