DiffMamba: Gated Differential Spectral-Spatial Mamba for Few-Shot Hyperspectral Image Classification

Few-shot hyperspectral image (HSI) classification aims to train a model with limited labeled data for accurate pixel-level recognition. Due to the scarcity of annotations and the lack of physical inductive biases in generic models, the model tends to suffer from severe overfitting and background noise interference. To mitigate this problem, in this paper, we propose a novel dual-branch network, named DiffMamba, for robust few-shot HSI classification. Specifically, we integrate a Self-Gating Mechanism into the Mamba units to dynamically recalibrate feature responses and suppress irrelevant noise. Besides, we design a Differential Spectral-Spatial Mamba (DiffSpaMamba) module and a Gated Spatial Mamba (SpaMamba) module to capture fine-grained spectral fingerprints and long-range spatial dependencies. This helps our model enhance its focus on discriminative regions and prevent overfitting in data-scarce regimes. Experimental results conducted on three public datasets demonstrate the effectiveness and superiority of the proposed model for few-shot HSI classification.