Group-Aware Spectral Perturbation Consistency Network for Cross-Domain Few-Shot Hyperspectral Image Classification

Abstract

Cross-domain few-shot hyperspectral image classification is challenged by scarce target-domain annotations, sensor-induced domain shifts, and spectral abnormalities, including band loss, noise, and spectral misalignment. To address these issues, we propose a Group-aware Spectral Perturbation Consistency Network (G-SPCNet) and evaluate both Clean classification and Robust perturbation stability under a unified protocol. G-SPCNet divides the spectrum into contiguous band groups and employs lightweight convolutions and bidirectional spectral-context modeling to construct grouped class prototypes. During source-domain meta-training, contiguous band masking, random band dropout, noisy-band injection, and spectral-axis shifting are applied online to form clean and perturbed views. Classification and predictive-distribution consistency objectives improve global robustness, while group-level prediction consistency and clean-prototype anchoring stabilize local decisions and prototype structures. In five-way five-shot experiments using LongKou as the source domain and Indian Pines, PaviaU, Salinas, and Chikusei as target domains, G-SPCNet achieved average Clean OA/AA, Robust OA/AA, and Robust Kappa values of 87.02%, 79.47%, and 74.34%, respectively. It outperformed seven comparison methods in the main Robust evaluation and maintained the highest average OA among the five evaluated methods at all tested support-set sizes in the shot-sensitivity analysis and at all tested severity levels in the mixed perturbation analysis. With only 0.153 M parameters, G-SPCNet provides a lightweight accuracy–robustness trade-off for cross-domain hyperspectral image classification.