DPMSCH-Net: A Dual Prior-Guidance Multi-Scale SSM-CNN Hybrid Network with Manifold Constraints for Hyperspectral Image Classification Under Small-Sample Conditions

Hyperspectral images (HSIs) demonstrate significant potential in agricultural applications, but conventional classification methods face challenges in adequately extracting complex spatial–spectral features. While deep learning approaches provide powerful representational capabilities, standard homogeneous architectures often show limitations in effectively capturing multi-scale contexts. Additionally, extreme label scarcity increases the difficulty of identifying central targets and informative spectral bands, while the absence of explicit feature-space constraints may render models more susceptible to feature collapse and overfitting. To this end, we propose a multi-scale asymmetric architecture named the dual prior-guidance multi-scale SSM-CNN hybrid network (DPMSCH-Net) for hyperspectral image classification (HSIC) based on few-shot learning (FSL) under an extreme scenario of only five labeled samples per class. Specifically, to capture multi-scale contexts, DPMSCH-Net employs a dual Mamba block (DMB) and a hierarchical context module (HCM) in the first two high-resolution stages, and the CNN branch in the third lowest-resolution stage for enhancing spatial–spectral representation. Subsequently, to extract discriminative features from limited samples, DMB is guided by the Gaussian spatial prior, while the PCA-inspired spectral decay prior provides statistical guidance for learning discriminative spectral representations. Finally, to further optimize the feature-space structure, supervised contrastive (SupCon) regularization is introduced to impose strict manifold constraints via implicit hard sample mining. Extensive five-shot experiments on the HongHu, HanChuan, and XiongAn datasets demonstrate that DPMSCH-Net achieves an average Overall Accuracy (OA) of 88.25%, 82.90%, and 67.35%, outperforming SOTA methods by a significant margin.