Answers are grounded in this paper and cite their sources.
Leveraging the advantage of utilizing both labeled data and a large amount of unlabeled data simultaneously, superpixel-based graph convolutional networks (GCN) have garnered widespread attention in the field of hyperspectral image classification. However, existing methods based on GCN primarily rely on single adjacent relationship modeling and do not fully utilize the edge relationships between superpixels, which limits their classification performance in complex terrain scenarios. In addition, GCN mainly focuses on the global dependencies of the entire hyperspectral image, making it difficult to simultaneously consider local dependencies and rich spectral features. To alleviate these issues, in this article, a novel multiadjacency Mamba-enhanced graph convolutional network (M2GCN) for hyperspectral image classification is proposed. First, a hierarchical cascade dimensionality reduction strategy is proposed to efficiently extract features and map them to the superpixel feature space. Second, considering the superpixel edge information, a similarity-edge enhanced GCN module is designed, which explicitly encodes superpixel adjacency constraints and guides the propagation of superpixel features with structural awareness. Finally, a new spectral-Mamba guided convolutional neural networks branch is constructed to decouple the extraction process of spatial local features and spectral information, and to recalibrate the channel weights. Extensive experiments demonstrate that, compared to the state-of-the-art method, M2GCN achieves classification performance improvements of 1.63%, 1.40%, and 1.29% on the Indian Pines, Pavia, and WHU-Hi-LongKou tasks, respectively.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: