DenseMixerMamba: Residual Mixing for Spectral–Spatial Hyperspectral Image Classification

The classification of hyperspectral images (HSIs) is a fundamental but challenging topic in the field of surface monitoring. In recent years, deep learning models like RNN, CNN, Transformers, and Mamba have demonstrated their remarkably and promising processing capabilities to confront high-dimensional and complicated HSI data. However, how to capture the refined spatial and spectral information from high-dimensional data continues to pose efficiency problems, which are compounded by the increased complexity of deep models. CNNs and Transformers employ thousands of parameters, imposing a high computational load and memory consumption. This paper addressed the parameter reduction problem by proposing an end-to-end DenseMixerMamba network for HSI classification, which converges Mamba, attention mechanisms, and dense architectures. Specifically, channel attention is first introduced to capture valuable channels and stimulate subsequent feature representation quickly. Then, to enhance the flow of information between layers and provide more stable training, a novel encoder based on the Mamba Vision Mixer (MVM) block and dense connectivity is developed to capture spatial and spectral patterns without increasing computational effort. Furthermore, a dual-branch architecture is designed to support the MVM block, which maintains both local and global information for final representation. Numerous experiments on four real datasets exhibit the more impressive classification accuracy and generalization performance of DenseMixerMamba over other state-of-the-art methods.