GCRNet: Global Context and Coordinate Attention-Based Double-Branch Residual Network for High Spatial Resolution Hyperspectral Image Classification

Deep learning has been extensively researched in hyperspectral image classification. Compared with traditional methods, it has significantly improved accuracy and efficiency. However, further research is needed on how to extract and fuse the useful information in hyperspectral images more effectively with deep learning models. In addition, the classification effectiveness of the model under small sample conditions still needs to be improved. The models need to extract more discriminative features and reduce the misclassification of similar classes. This paper proposes a Global Context and Coordinate Attention Residual Network (GCRNet). In our model, a dual-branch structure based on two attention mechanisms is used for separate feature extraction. To evaluate the effectiveness of our model under limited data conditions and to address the challenge of data annotation burden, we trained GCRNet using only 1%, 2%, 2%, and 2% of the available samples from the Wuhan high spatial and spectral resolution dataset and the Salinas Valley dataset, respectively. Despite the small sample sizes, our proposed GCRNet achieved high accuracies of 99.56%, 99.11%, 99.38%, and 99.78% on these datasets, outperforming several existing methods. These results demonstrate GCRNet's ability to extract discriminative features and model global contexts effectively, even with limited training data, thus potentially reducing the need for extensive data labeling in practical applications.