Answers are grounded in this paper and cite their sources.
Convolutional neural networks (CNNs) and graph convolutional networks (GCNs) have made significant advances in hyperspectral image (HSI) classification. However, the outstanding performance of these methods largely depends on a substantial amount of labeled samples, while the high labeling cost of HSIs leads to a scarcity of labeled samples. Meanwhile, using only CNN or GCN may result in information loss due to the limited range of feature perception, making it difficult to conduct sufficient feature mining. In this article, we propose a semisupervised HSI classification method called pseudosample guided multiview GCN-CNN fusion network (PS-MGCFN), which utilizes a pseudosample guided multiview GCN-CNN fusion network. Specifically, an adaptive center comparison pre-encoding criterion is first devised to learn to differentiate category-specific sample features and preliminarily filter valuable unlabeled samples. Furthermore, the pseudosample generation is performed by referencing the distribution characteristics of labeled samples. Then, a multiview GCN-CNN fusion network is developed to combine information from both short-range and long-range scenes for resulting in representative fusion features. Importantly, the collaborative training of a large number of pseudosamples with a small number of labeled samples greatly enhances the performance of the model. Extensive experiments conducted on three datasets demonstrate that the proposed PS-MGCFN exhibits superior stability and classification performance compared to existing methods with limited training samples.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: