Hyperspectral image classification via adaptive jump scanning Mamba and dynamic perturbation fusion

Hyperspectral image classification (HSIC) critically depends on effective local and global spatial-spectral representation. However, existing methods often struggle to balance computational efficiency with discriminative feature extraction. While recent Mamba-based architectures have shown strong capabilities in sequence modeling with linear complexity, they typically lack adaptive scanning strategy to capture spatial-spectral variation. Additionally, while effective fusion of spatial and spectral features is crucial for classification, existing methods operate solely on the extracted features. The single-feature perspective makes the model prone to overfitting and limits its generalization ability, especially in small-sample scenarios. To address these limitations, we propose AJSDP-Mamba, a novel framework that integrates adaptive spatial-spectral modeling with noise-based dynamic perturbation fusion. The framework consists of three core components: (1) The Adaptive Jump Spatial Scanning Mamba (AJSS-Mamba) module captures fine-grained local spatial features by employing a texture-aware jump scanning strategy, which dynamically adjusts spatial strides to preserve structural details in complex regions while reducing computational cost in homogeneous areas. (2) The Adaptive Jump Band Scanning Mamba (AJBS-Mamba) module adaptively adjusts spectral scanning strides based on local spectral variations, enabling efficient and targeted feature extraction across spectral domains. And (3) The proposed Noise-based Perturbation Ensemble with Variance-weighted Fusion(NPF) module introduces controlled random perturbations to enhance feature diversity and performs adaptive spatial spectral fusion through a variance based weighting mechanism, effectively improving fusion robustness. Extensive experiments on four widely-used HSI benchmarks demonstrate that AJSDP-Mamba achieves a higher classification performances and computational efficiency across diverse classification tasks. The code is available at https://github.com/AAAA-CS/AJSDP-Mamba.