In recent years, unsupervised domain adaptation methods based on deep learning have shown great potential in hyperspectral image cross-domain classification. However, existing methods often face two major challenges: domain distribution differences and data noise interference. To address these issues, this paper proposes a cross-domain classification framework combining confident learning and masked self-distillation (CL-MSD), which constructs an adversarial training network, leveraging the competition between dual classifiers and feature extractors to learn domain-invariant features, and employs a spatial-spectral dual-branch structure to enhance the joint modeling of local and global features. The framework also introduces a masked self-distillation (MSD) module to simulate missing data scenarios and a confident learning (CL) module to filter out low-confidence samples. Experiments on the Houston and Pavia datasets show that the proposed method achieves overall classification accuracies of 83.38% and 93.60%, respectively, demonstrating its effectiveness.