Fuzzy-MambaHSI: Pixel-Adaptive Fuzzy State-Space Model for Hyperspectral Image Classification

Hyperspectral image classification (HSI) faces substantial class-membership uncertainty in mixed-pixel and spectrally confused scenes. The existing Mamba-based methods can model long-range spatial–spectral dependencies with linear complexity, but they usually rely on deterministic feature propagation and do not explicitly represent graded pixel membership during the forward pass. To address this limitation, this paper proposes Fuzzy-MambaHSI, a novel hyperspectral classification framework that embeds learnable fuzzy neural networks into a full-image dense-prediction Mamba backbone. The framework contains three key designs: (1) a fuzzy token perturbation module, which applies soft rule-weighted additive perturbations to tokens at the input of the spatial scan; (2) a fuzzy spatial–spectral fusion module, which replaces global scalar fusion with pixel-wise fuzzy neural network inference; and (3) a training stabilization strategy composed of linear warm-up and rule-center diversity regularization. Experiments with the same settings show that Fuzzy-MambaHSI improves the MambaHSI baseline by 0.96 and 2.92 % in overall accuracy on Pavia University and WHU-Hi-HanChuan, respectively, and yields lower variance in the agricultural scene with stronger spectral confusion. A rule activation analysis shows that different spatial fuzzy rules focus more strongly on homogeneous regions and high-frequency boundary regions, revealing observable differences in pixel membership during the forward pass.