ASSFMamba2 : Adaptive Spatial‐Spectral Fusion via Mamba2 for Hyperspectral Image Classification

Hyperspectral image (HSI) classification is a fundamentally challenging task in remote sensing, requiring effective joint modeling of intricate spectral signatures and complex spatial dependencies. Although deep learning models such as convolutional neural networks (CNNs) and Transformers have achieved remarkable progress, they are inherently limited by local receptive fields or quadratic computational complexity. These limitations hinder their ability to efficiently capture long‐range dependencies at the pixel level. To overcome these issues, we propose ASSFMamba2, a novel image‐level HSI classification framework that leverages the selective state‐space model (Mamba2) for efficient long‐range modeling and an Adaptive Spatial‐Spectral Fusion (ASSF) mechanism. Specifically, the framework comprises three core components. First, the Spectral Mamba2 Block (SpeMB) partitions spectral channels into groups and applies state‐space modeling to capture long‐range spectral dependencies. Second, the Spatial Mamba2 Block (SpaMB) flattens spatial sequences and aggregates global contextual information across pixels. Third, an Adaptive Fusion Module dynamically integrates spatial and spectral features through pixel‐wise soft gating, cross‐modality interaction, and channel‐wise recalibration. These designs ensure that the most discriminative features are emphasized at every location. Extensive experiments on four benchmark HSI datasets demonstrate that ASSFMamba2 not only surpasses existing state‐of‐the‐art methods in overall accuracy, average accuracy, and kappa coefficient, but also achieves faster convergence and enhanced robustness. These results highlight the potential of Mamba2 as a new backbone for hyperspectral image analysis and confirm the effectiveness of our adaptive spatial‐spectral fusion strategy. Codes are available at https://github.com/lixingqiao/ASSFMamba2 .