正如前面所描述，[19]和本文存在诸多的不同之处，包括动机、框架以及模

正如前面所描述，[19]和本文存在诸多的不同之处，包括动机、框架以及模块。本文搭建了一种由粗化阶段和细化阶段组成的新颖层级结构以提取HSI中丰富的多层次特征，此外它还考虑了像素级特征和超像素级特征。值得注意的是，我们在第一个粗化阶段引入关系图卷积网络（RGCN）使得相似的节点特征更近，后续粗化阶段利用GCN进行特征提取。[19]提出的自适应图结构集成深度网络（DNAGSI）可以动态学习HSI的图结构，并且通过将中心损失函数和GCNII以有效地缓解过平滑问题。

As described earlier, there are many differences between [19] and this article, including motivation, framework, and modules. This article constructs a novel hierarchical structure consisting of coarsening and refinement stages to extract rich multi-level features from HSI. In addition, it also considers pixel level and superpixel level features. It is worth noting that in the first coarsening stage, we introduce the Relationship Graph Convolutional Network (RGCN) to bring similar node features closer, and in the subsequent coarsening stages, we use GCN for feature extraction. [19] The proposed Adaptive Graph Structure Integrated Deep Network (DNAGSI) can dynamically learn the graph structure of HSI, and effectively alleviate the oversmooth problem by combining the center loss function with GCNII.

As described earlier, there are many differences between [19] and this paper, including motivation, framework and modules. In this paper, a novel hierarchical structure composed of coarsening stage and thinning stage is built to extract rich multi-level features in HSI, and pixel-level features and super-pixel-level features are also considered. It is worth noting that in the first coarsening stage, we introduce the graph convolution network (RGCN) to make similar node features closer, and in the subsequent coarsening stage, we use GCN to extract features. [19] The proposed adaptive graph structure integration depth network (DNAGSI) can dynamically learn the graph structure of HSI, and effectively alleviate the over-smoothing problem by integrating the center loss function and GCNII.