Beyond Sequential Scanning: A Synergistic Mamba Paradigm With Adaptive Receptive Recalibration for Hyperspectral Image Classification

State Space Models (SSMs) have emerged as a compelling paradigm for long-range dependency modeling with linear computational complexity. However, their intrinsic 1D scanning process inevitably disrupts the local topological continuity of high-dimensional image structures. This limitation is particularly pronounced in hyperspectral image (HSI) classification, where existing Mamba-based methods frequently struggle with intricate spatial-spectral coupling and the deficiency of local inductive bias, resulting in suboptimal feature representation. To surmount these bottlenecks, we propose SynerMamba, a novel synergistic Mamba-based framework designed to reconstruct local structural integrity within the sequential modeling paradigm. Specifically, a Spatial-Spectral Disentangler (SSD) is first introduced to explicitly decouple highly correlated spatial-spectral features via local dynamic graphs, establishing a structurally conscious manifold prior to SSM processing. To rectify the perception bias in sequential scanning, an Adaptive Receptive Mamba (ARM) is developed to inject multi-scale spatial priors into the selective scanning process, effectively capturing fine-grained textures. Furthermore, a Mamba Synergizer (MS) is engineered to orchestrate the sophisticated nonlinear interplay between spatial contexts and spectral signatures, transcending the representation constraints of conventional linear aggregation strategies. Extensive experiments on benchmark datasets demonstrate that SynerMamba achieves state-of-the-art classification performance while maintaining high computational efficiency.