GCEMNet: Grouped Cascaded Enhanced Mamba Network for Hyperspectral Image Classification in Spectrally Similar Scenarios
Abstract
Hyperspectral image (HSI) classification remains challenging in spectrally similar scenarios, where different ground objects may share highly overlapping spectral signatures and comparable spatial structures. In such cases, weak discriminative cues are easily obscured by redundant spectral and spatial responses, leading to confusion between categories with similar material composition or spatial patterns. To address this problem, this article proposes a grouped cascaded enhanced Mamba network (GCEMNet) for HSI classification in spectrally similar scenarios. GCEMNet follows a progressive feature learning pipeline. Specifically, the channel–spatial attention fusion (CSAF) module preserves subtle discriminative cues by jointly exploiting channel statistics and spatial structural information; the grouped cascaded Mamba fusion (GCMF) module, built upon multiple enhanced Mamba (EM) units, captures complementary local and global dependencies through grouped learning and cascaded propagation; and the cross-feature separation Transformer (CFST) reduces the influence of redundant spectral and spatial responses through correlation-aware feature modeling and cross-feature interaction. Extensive experiments on two public mineral HSI datasets and one agricultural HSI dataset demonstrate that GCEMNet achieves superior classification accuracy with moderate model complexity. Module comparison, ablation studies, visualization results, and computational complexity analysis further verify the effectiveness and practicality of the proposed modules in spectrally similar scenarios and scenarios with class imbalance.