Recently, the hypergraph neural network (HGNN) has drawn increasing attention in modeling complex high-order correlations. Compared to simple graph neural networks, HGNNs exhibit more powerful representational ability. There are two limitations in the application of hypergraph theory to hyperspectral image (HSI) classification. One is the inadequate explicit representation of semantic information contained in HSIs. Another is the loss of pixel-level spectral–spatial information. Thus, an enhanced HGNN (EHGNN) is proposed to promote the application of hypergraph theory to HSI classification. Specifically, two important enhancements are introduced: 1) the concept of a key hypergraph, providing more rich semantic information and improving the interpretability for complex distribution structures and 2) the integration of the convolutional neural network (CNN) and HGNN architectures into an end-to-end framework, the loss of spectral–spatial information at the pixel level is effectively reduced. Through these two enhancements, EHGNN exhibits a 4% improvement in overall accuracy (OA) on the Pavia University dataset and a 2% improvement in OA on the Xuzhou dataset compared to HGNN. Furthermore, the test results on two HSI datasets demonstrate that our EHGNN achieves competitive performance compared to other state-of-the-art methods.

中文翻译：

---

EHGNN：用于高光谱图像分类的增强型超图神经网络

最近，超图神经网络（HGNN）在复杂高阶相关性建模方面引起了越来越多的关注。与简单的图神经网络相比，HGNN 表现出更强大的表示能力。超图理论在高光谱图像（HSI）分类中的应用存在两个局限性。一是 HSI 中包含的语义信息的显式表示不充分。另一个是像素级光谱空间信息的丢失。因此，提出了增强型HGNN（EHGNN）来促进超图理论在HSI分类中的应用。具体来说，引入了两个重要的增强功能：1）关键超图的概念，提供更丰富的语义信息并提高复杂分布结构的可解释性；2）将卷积神经网络（CNN）和HGNN架构集成到终端中端到端框架，有效减少了像素级光谱空间信息的损失。通过这两项增强，与 HGNN 相比，EHGNN 在帕维亚大学数据集上的整体准确率 (OA) 提高了 4%，在徐州数据集上的 OA 提高了 2%。此外，两个 HSI 数据集的测试结果表明，与其他最先进的方法相比，我们的 EHGNN 实现了具有竞争力的性能。