Answers are grounded in this paper and cite their sources.
Convolutional neural networks (CNNs) and visual Transformers have become the dominant approaches for hyperspectral image (HSI) classification due to their powerful feature representation capabilities. However, CNNs are limited in capturing long-range dependencies, while Transformers suffer from quadratic computational complexity when modeling spectral relationships in high-dimensional data. The selective state space model has linear complexity but still faces challenges of long-range decay and insufficient spatial–spectral modeling. To address these limitations, we propose a Spatial–Spectral Mamba–Convolution Fusion Network (SSMFMamba) for HSI classification. The proposed network effectively combines the advantages of both Mamba and CNN paradigms in global and local representation, respectively, enabling efficient spatial spectral feature learning. A Spatial-Guided Neighborhood Scanning Module (SGNSM) is further designed to reconstruct the spatial domain based on Euclidean distance to alleviate long range decay. Then, SGNSM performs unidirectional scanning to effectively capture global dependencies and reduce redundancy. Meanwhile, a Spatial–Spectral Convolutional Module (SSCM) is designed to extract local structural information by combining 2 D and 3-D convolutions. To adaptively balance global–local information, a Multi-Attention Fusion Module (MAFM) subsequently fuses the outputs from SGNSM and SSCM. Finally, a Spatially Enhanced Spectral Scanning Module (SESSM) enhances spectral representations by performing spatial attention-guided fusion followed by bidirectional spectral scanning, enabling more effective spectral interaction modeling. To validate the effectiveness of SSMFMamba, experiments are conducted on three benchmark datasets including Houston2013 (HS2013), WHU-Hi-LongKou (LK), and Pavia University (PU). The results indicates that SSMFMamba shows superiority over the state-of-the-art hyperspectral classification networks, which include CNNs, Transformers, and hybrid CNN-Transformer models.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: