Abstract: It is difficult to capture long-range dependencies when using deep learning methods based on convolutional neural networks, and there is a lack of attention to local information in networks represented by codecs. In addition, the limited target reference values cannot meet the demand for a large number of labeled samples during the training process. This study investigated a ternary spectral discrimination network that combined local feature extraction with global spectral attention to simultaneously focus on local information and establish long-range dependencies, thereby comprehensively capturing important features in hyperspectral data and promoting the separation of targets and backgrounds. To address the issue of insufficient samples, we obtained background samples through subset partitioning and combined them with the prior objectives to synthesize a sufficient number of samples. The experimental results on six datasets showed that the proposed method could effectively detect targets and suppress backgrounds, outperforming other experimental methods, with an average accuracy of 99.6%. This demonstrated the effectiveness of the proposed method and provided a reliable solution for future hyperspectral data processing.