OmniMamba: Omnidirectional Scanning Meets State Space Models for Efficient Hyperspectral Image Classification

Qiyun Zheng, Taosheng Xu, Chenglong Zhang, Peng Li, Wenwen Min, Changmiao Wang

Abstract: Accurate classification of hyperspectral images (HSI) is crucial for earth observation and agricultural production analysis, yet remains challenging due to high dimensionality, spectral variability, and limited training samples. Traditional approaches often struggle to effectively balance computational complexity with the ability to capture spatial-spectral relationships of observation targets. To address this challenge, we propose OmniMamba, a novel omnidirectional state space model that adopts a collaborative alternation strategy integrating single-scale and multi-scale feature processing with omnidirectional scanning mechanisms. Four complementary scanning patterns (row, column, zigzag, and snake) are employed in the omnidirectional scanning mechanism to transfer 2D spatial data into 1D spatially structured feature sequence, which preserves directional sensitivity while achieving global dependency modeling with only linear complexity. This avoids the quadratic complexity bottleneck inherent in self-attention mechanisms. Our collaborative alternation strategy coordinates fine-grained spectral signatures with hierarchical spatial contexts through cascaded processing stages, addressing the spectral-spatial feature fusion challenge in HSI classification. Extensive experiments conducted on four benchmark datasets validate the superiority of OmniMamba, achieving a mean overall accuracy of 99.28%, significantly outperforming existing methods. Remarkably, our model accomplishes the performance with only 246 K parameters and 0.04 GFLOPs, demonstrating dramatically low computational complexity than the state-of-the art conventional CNN and transformer-based architectures.

External IDs:doi:10.1109/jstsp.2026.3657243