Recent advances in mamba-based architectures have demonstrated promising potential for hyperspectral image classification (HSIC), offering linear-complexity long-range dependency modeling. However, two critical challenges persist in adapting this paradigm to HSI analysis: the substantial requirement for annotated training samples, and insufficient capacity to interpret the intricate spatial-spectral features inherent in HSI data, particularly under few-shot learning scenarios. To address these limitations, we present SSupMamba, a novel self-supervised mamba framework tailored for HSIC. First, we propose a composite scanning mamba block (CSMB) that enables comprehensive global feature extraction through multi-directional selective scanning of HSI data cubes. Second, we develop a spatial-spectral masked mamba (SAEM) framework that employs randomized masking and reconstruction tasks to enhance local representation learning. Third, we establish a unified self-supervised architecture incorporating contrastive learning to maximize mutual information between multi-views while preserving intrinsic spatial-spectral characteristics. Experimental results on four public datasets demonstrate that the proposed method exhibits excellent feature extraction capabilities under few-shot conditions and outperforms several state-of-the-art HSIC methods. The code is available at: https://github.com/Winkness/SSupMamba.

The hyperspectral image classification community has long sought methods to handle scarce samples and unpredictable variations. Recent research aims to develop foundational models that overcome traditional limitations, such as inadequate representational capability, poor integration of prior knowledge, and incomplete feature extraction. However, the high cost of data collection and annotation, along with the billions of parameters involved, hinder real-world deployment. In this paper, we propose an insightful method for domain generalization using overall semantics, positional relationships, and morphological characteristics as domain constant knowledge carriers. This approach enables accurate cross-scene classification while significantly reducing deployment costs. Specifically, we design a lightweight vision-language hybrid domain generalization network (VLDG-Net) for few-shot cross-scene hyperspectral image classification. First, features from the source domain (SD) data are extracted through jointly optimized image and text encoders, with the resulting model transferable to the unseen target domain (TD) after pre-training. Second, we propose a residual hybrid backbone (RHB) as the image encoder, using residual 3D blocks (R3DB) to extract spatial-spectral features and residual 2D selective scanning blocks (RSS2DB) to model long-range dependencies. Additionally, we design an adapter for efficient prior knowledge embedding in the text encoder. Lastly, a two-stage contrastive loss (TSCL), combining vision-language alignment loss and instance discrimination loss, is proposed to create a more separable feature space. Extensive experiments on three public datasets show that our method outperforms several advanced approaches in few-shot cross-scene HSI classification, while also being efficient in training and testing.