在其他条件不变的情况下，我们尝试将所提方法的后续粗化阶段中的GCN全部

在其他条件不变的情况下，我们尝试将所提方法的后续粗化阶段中的GCN全部换成GCNII，在训练样本比例为10%的Indian Pines数据集得到的结果如下表2所示。我们可以清楚地发现，相比于GCN，GCNII的确可以提取到更具判别性的特征。对于表2中的XX仍然比不过[19]中表II中的98.36，我们认为是由于这两篇论文存在的差异造成的，包括构图方式、特征提取方式以及目标优化的损失函数。

Under other unchanged conditions, we attempted to replace all GCNs in the subsequent coarsening stage of the proposed method with GCNII. The results obtained on the Indian Pines dataset with a training sample ratio of 10% are shown in Table 2. We can clearly observe that compared to GCN, GCNII can indeed extract more discriminative features. We believe that the XX in Table 2 is still lower than 98.36 in Table II of [19], which is due to the differences between these two papers, including the composition method, feature extraction method, and the loss function of objective optimization.

Under other conditions unchanged, we try to replace all GCN in the subsequent coarsening stage of the proposed method with GCNII, and the results obtained on Indian Pines data set with a training sample ratio of 10% are shown in Table 2 below. We can clearly find that GCNII can indeed extract more discriminative features than GCN. As for XX in Table 2, it is still not as good as 98.36 in Table II in [19]. We think it is caused by the differences between the two papers, including composition mode, feature extraction mode and loss function of objective optimization.