Publication

Abstract : Hyperspectral image classification has become an important research topic in remote sensing, providing fine-grained spectral information for land cover discrimination. In precision agriculture, high spectral dimensionality, interclass similarity, and spatial heterogeneity factors produce weak boundary discrimination and feature propagation across similar spectral values, leading to misclassification. Traditional convolution-based deep learning methods are limited in this setting because they are not well suited to model the non-Euclidean structural relationships inherent in hyperspectral data. Although graph convolutional networks (GCNs) can model non-Euclidean spectral–spatial relationships, graph aggregation still propagates information through misleading edges, causing oversmoothing near class-transition regions. To address this issue, an edge-priority mechanism assigns a structural priority score to each graph edge before attention normalization, thereby highlighting informative spectral–spatial connections. Accordingly, a dual-branch multihead attention-based graph convolutional fusion network (DB-MHAGCFN) is proposed to learn the local spatial structure and global spectral dependencies. The dual-branch design combines convolutional feature extraction for local spatial continuity with GCN-based learning for long-range spectral–spatial dependence modeling, enabling complementary feature interaction between local and nonlocal representations. The scalar edge-priority score regulates the internode information flow to capture heterogeneous relational dependencies without introducing substantial additional optimization complexity. In addition, the aggregation of multihead attention scores enables the proposed model to capture diverse relational patterns across multiple scales. The proposed DB-MHAGCFN improves edge preservation and reduces fragmented predictions by maintaining sharper class transition and stronger region coherence for classification. The proposed model was experimentally evaluated using various benchmark datasets. Region coherence index (Rci) was used to quantify the boundary preservation and regional consistency, to assess the compactness of the predicted class region. The experimental results demonstrate that DB-MHAGCFN consistently outperforms the comparison models across the evaluation metrics.