Prompt-Guided Cross-Network Distillation for Open-Set Cross-Domain Few-Shot Hyperspectral Image Classification
Compared with cross-domain few-shot learning (CDFSL), open-set CDFSL (OS-CDFSL) is more challenging because it must accurately classify known classes across domains while reliably rejecting unknown samples. Under this setting, domain shift causes inconsistent spectral distributions for samples from the same class and causes unknown samples to intrude into the known feature space, resulting in spectral confusion and blurred decision boundaries. Existing methods still rely on feature alignment, which may incorrectly pull unknown-class samples toward known-class regions. Moreover, threshold-based rejection strategies are sensitive to dataset characteristics and domain shifts, limiting their generalization. To address these limitations, this article proposes a prompt-guided cross-network prototype distillation (PG-CPD) method for OS-CDFSL. First, a prompt-guided foundation adaptation (PGFA) module is developed based on parameter-efficient fine-tuning (PEFT) to reduce cross-domain spectral discrepancies. Masked image modeling (MIM) optimizes learnable spatial–spectral prompts to adapt the frozen backbone to domain-specific spectral variations without using class labels. Second, a cross-network prototype distillation (CNPD) module is introduced to transfer stable semantic structures while avoiding boundary contamination from direct feature alignment. In particular, student features are evaluated with teacher semantic prototypes for cross-network logit distillation, while dynamic graph-based nonknown anchors serve as negative prototypes to form rejection regions. Finally, a threshold-free inference strategy reformulates open-set rejection as a similarity competition, avoiding manually defined thresholds. Experiments on three hyperspectral image (HSI) datasets demonstrate that the proposed method achieves state-of-the-art performance. The code is available at https://github.com/ZhengSR-031/PG-CPD