Cross-Architecture Contrastive Learning for Few-Shot Hyperspectral Image Classification

Hyperspectral image classification is a fundamental task in remote sensing with broad applications in precision agriculture, environmental monitoring, and related fields. However, existing models encounter significant difficulties when labeled samples are scarce. Although contrastive learning has mitigated the dependence on labeled data, most current methods remain limited to a single network architecture, thus failing to exploit the complementary strengths among heterogeneous representations. To address this issue, we propose a cross-architecture contrastive learning (CACL) framework for few-shot hyperspectral image classification, which integrates convolutional neural networks (CNNs) and graph convolutional networks (GCNs) through a parallel dual-branch design. The CNN branch captures finegrained local spectral–spatial representations, while the GCN branch models global graph-structural dependencies. Positive sample pairs are constructed by aligning these heterogeneous features, and their consistency is maximized within a shared embedding space to achieve deep cross-architecture fusion. Moreover, a spectral frequency-domain enhancement module (SFEM) is introduced into the CNN branch to refine local features via frequency-domain analysis, and an adaptive multiscale graph convolution module (AMGCM) is incorporated into the GCN branch to capture richer hierarchical spatial features. Experiments on the WHU-Hi-HanChuan, Indian Pines, and University of Pavia datasets demonstrate that our method achieves overall classification accuracies of 95.29%, 96.73%, and 99.17% respectively with extremely limited training samples, significantly outperforming multiple mainstream baseline models.