Vortex Mamba for hyperspectral image classification

Accurate classification of hyperspectral imagery (HSI) is critical for remote sensing applications such as land-cover mapping and urban monitoring. Although transformers excel at modeling long-range dependencies in HSI, their quadratic complexity in modeling the extensive spatial contexts imposes prohibitive computational burdens. Mamba offers linear efficiency but suffers from unidirectional scanning, limiting its capacity to fully capture the 2D neighborhood structures that are crucial for local spatial relationship modeling. To overcome these limitations, we propose Vortex Mamba—a hybrid architecture where the U-Net captures multiscale contexts, the Mamba module enables efficient long-range modeling with linear complexity, and the transformer component performs dynamic feature recalibration for complementary spectro-spatial feature extraction. Our Vortex Mamba module replaces unidirectional scanning with centripetal spiral traversal—a scan that starts at the patch center and spirals outward, propagating features from the spatial center outward to establish radial hierarchical dependencies and angular continuity—meaning it preserves the sequential relationship of pixels along the circular direction of the spiral. This enables direction-agnostic global modeling by leveraging the linear computational complexity of the Mamba module during both training and inference. Furthermore, our channel-enhanced attention module performs dynamic spectral recalibration, which computes channel-wise weights from the central pixel’s feature vector, a design based on the assumption that the central pixel typically plays a dominant role in patch-based classification. To ensure rigorous evaluation and avoid data leakage, we explicitly guarantee that there is no pixel-level overlap between the training and testing patches. Each pixel belongs exclusively to either the training set or the test set, but not both. Using only 695, 1977, and 2833 labeled samples (10%) for training on the Indian Pines, Pavia University, and Houston 2013 datasets, respectively, Vortex Mamba consistently outperforms leading modern models across five independent runs, achieving notable mean overall accuracies of 98.78%, 99.51%, and 99.61%. Codes (including training scripts and experimental hyperparameters) are available at https://github.com/cuixiandai/VortexMamba.