<p>In the field of hyperspectral image (HSI) classification, graph neural networks have become a mainstream method. However, graph convolutional networks (GCNs) focus on spatial features and ignore spectral information, while graph attention networks (GATs) are good at extracting spectral features but insufficient at capturing the complete spatial context between pixels. To address these limitations, we propose the enhanced graph convolution and attention fusion network (EGCAFN). Firstly, the module implements the fusion of GCN and GAT: It combines the topology-aware capability of GCN and the attention mechanism of GAT and adaptively adjusts the weights of different features through a node importance assessment mechanism to accurately capture the remote dependencies between pixels. Secondly, the depth-separable convolution and multi-scale feature fusion (DSF) module are introduced to efficiently extract multi-scale information while reducing computational complexity. Finally, the global–local attention model (GLAM) is introduced to solve the feature over-dispersion problem by integrating global and local information by combining the efficient channel attention (ECA) and spatial attention mechanisms. The overall accuracy (OA) of our EGCAFN on Indian Pines, Pavia University, Salinas and Botswana datasets is 94.59%, 96.03%, 97.04% and 99.84%, respectively. And the classification accuracy is improved on all four datasets.</p>