基于混合空洞卷积的高光谱图像分类方法OA

Hyperspectral image classification is often hindered by high-dimensional redundancy and complex spatial structures,leaving room for improvement in both accuracy and efficiency.To address the issue of insufficient receptive fields in the convolutional branches of convolutional neural network-graph convolutional network(CNN-GCN)fusion frameworks,an improved method based on hybrid dilated convolution,termed DC-AMGCFN,is proposed.This method introduces dilated convolutions in the final layers of the large-and medium-scale branches to expand the receptive field,and integrates multi-hop graph convolutions with cross-branch attention mechanisms to achieve spectral-spatial feature fusion.Experimental results on three benchmark datasets:Indian Pines,Pavia University,and Salinas,demonstrate that the overall classification accuracies reach 92.70%,95.18%,and 97.30%,respectively.Ablation studies confirm that a moderate dilation configuration significantly optimizes feature extraction.The research indicates that hybrid dilated convolution can achieve a favorable balance between accuracy and efficiency,providing a lightweight,effective,and reusable improvement strategy for CNN-GCN fusion.