Mamba-DCAU: state space dual attention center-sampling U-Net for hyperspectral image classification

Abstract

Hyperspectral image classification, a key remote sensing research area, has greatly benefited from recent advances in deep learning. Although computer vision methods based on V-Mamba and U-Net have demonstrated remarkable results, existing frameworks still face dual challenges: the unidirectional scanning mechanism of the Mamba model limits its ability to jointly model spatial neighbourhood features, while the inherent high-dimensional characteristics and large-scale nature of hyperspectral images create significant computational efficiency bottlenecks in traditional whole-image input paradigms. This study introduces a novel hybrid model, Mamba-DCAU, which integrates Mamba, U-Net, and attention mechanisms to improve classification accuracy. The model synergistically optimizes local spectral-spatial feature representation and global semantic segmentation capabilities, offering a unique approach to feature extraction and segmentation. The proposed bidirectional dynamic attention mechanism constructs horizontal and vertical dual-path attention branches through feature map transposition, employing a learnable weight coefficient-based linear interpolation fusion strategy to achieve bidirectional adaptive modelling of neighbouring relationships, effectively addressing the local correlation deficiency caused by Mamba’s unidirectional scanning mechanism. A centre-sampling training method is designed to constrain gradient backpropagation through the centre pixels of U-Net outputs, establishing a mapping between local features and global parameters while maintaining the advantages of convolutional kernel weight sharing, enabling end-to-end efficient training. By combining Mamba’s global modelling capability with U-Net’s multi-scale feature extraction, feature selection is enhanced through attention mechanisms. The model was evaluated on three commonly used datasets: Indian Pines, University of Pavia, and Houston 2013. Experimental results demonstrate that it achieves outstanding classification accuracies of 98.29%, 99.82%, and 99.02%, respectively, outperforming baseline methods in terms of both accuracy and robustness. Codes are available at https://github.com/cuixiandai/MambaDCAU

Topics & Concepts