Noise perturbation augmentation based dual-branch alignment network for cross-domain hyperspectral image classification

The acquisition of accurately labeled data for hyperspectral image classification is frequently associated with substantial costs and significant time investment, which considerably restricts the practical implementation of this technology. This study presents an innovative framework termed the Noise Perturbation Augmentation based Dual-branch Alignment Network (NPAnet), specifically designed for cross-domain hyperspectral image classification. This framework enables accurate classification without the requirement for labeled samples from the target domain. To address the performance degradation that arises from discrepancies in feature space during cross-domain scenarios, we propose three complementary modules. First, to enhance the learning of semantic features, we introduce Noise Perturbation Augmentation (NPA), which reconceptualizes adversarial perturbations as structured, task-relevant information through the application of positive excitation noise. Second, to reduce distributional shifts while maintaining feature consistency, we propose Dual-Branch Feature Alignment (DBFA), which implements dual alignment across spatial-spectral and graph-structural domains. This approach aims to minimize distributional shifts while ensuring the consistency of features. Lastly, to enhance generalization within the target domain, we introduce the Adaptive Pseudo-Label Refinement (APLR) module, which incrementally expands the training set for the target domain by dynamically selecting high-confidence pseudo-labels and incorporating category-weighted adjustments to improve generalization. Experimental evaluations conducted on three benchmark hyperspectral datasets demonstrate superior classification performance across various spectral configurations, thereby validating the effectiveness and practical applicability of the proposed methodology.