Spectral mamba enhanced neighborhood attention network for aerial hyperspectral image classification

Hyperspectral images (HSIs) are celebrated for their rich spectral information, making them highly effective for precise land cover classification. Deep neural networks, such as vision transformers (ViTs) and state space models (Mamba), have made significant advancements in hyperspectral image classification (HSIC). However, ViTs are often limited by their quadratic computational complexity and a predominant focus on global information, which can hinder their ability to extract crucial local features essential for HSIC. While Mamba-based architectures offer linear computational complexity and impressive performance, they are constrained by their limited understanding of the spatial and spectral information in HSIs. To address these limitations, we propose a novel spectral Mamba-enhanced neighborhood attention (SMENA) hybrid network, designed to effectively leverage the strengths of various architectures. This network integrates a local spatial feature extraction (LSFE) module with a spectral Mamba (SpeM) specifically for HSIC. The bidirectional scanning mechanism in SpeM enhances its ability to capture discriminative spectral features, while the LSFE, composed of convolutional and neighborhood attention modules, hierarchically captures detailed local spatial features. Extensive experiments on four widely used public datasets demonstrate that our model achieves superior classification performance compared to other eight benchmark methods.