PCRNet: Progressive Correlation Refinement for Few-Shot Hyperspectral Image Classification

Few-shot hyperspectral image classification (FHSIC) is crucial for precise land cover identification when labeled samples are scarce, mitigating the prohibitive annotation costs associated with hyperspectral image (HSI) data. However, existing FHSIC methods primarily follow a feature refinement paradigm, which struggles to effectively model the holistic relational structure among samples. Meanwhile, the commonly used attention mechanism tends to focus solely on sample-specific local information, limiting its ability to capture global category knowledge and promote cross-domain generalization. To address these challenges, we propose an efficient Progressive Correlation Refinement Network (PCRNet). Our work pioneers a correlation refinement paradigm, shifting the objective from enhancing feature representations to explicitly reasoning upon the structured query-class relationships. Specifically, PCRNet first constructs a learnable generalized correlation matrix that encodes the inter-sample relationships. Then, we present a simple yet effective memory-guided global attention mechanism to facilitate the injection of category commonality priors into the relational structure from a holistic perspective. This is further combined with query-level adaptive modulation and symmetrical gated integration modules to achieve progressive refinement of the generalized correlation matrix. Finally, the refined correlation is projected as the classification logit, enabling accurate identification of land covers. Experiments on four benchmark HSI datasets demonstrate the state-of-the-art performance of our approach. We believe this new paradigm opens a promising direction for future FHSIC model design. Code is available at https://github.com/guoying918/PCRNet.