随着高光谱图像数据的普及与深度学习在图像处理领域的进步，如何进一步提高分类精度成为一个令人关注的问题。高光谱图像因为波段多且相互之间冗余度高，分类受到“同谱异物”或“同物异谱”的影响很大。除此之外，高光谱数据标注因为成本高，所以很多时候会遇见标注样本少的状况，这对深度学习要求大的训练样本数量相悖。 基于以上问题，本实验提出了使用生成对抗神经网络的一种——ACGAN 进行高光谱图像分类，利用生成器与判别器之间的对抗学习来学习如何处理高光谱图像的光谱信息，并研究了训练样本数与分类精度的关系。 研究发现，当训练样本占总样本数5%时，ACGAN在帕威亚大学数据集上的分类精度已经达到了96%，在印度松树数据集上分类精度达到98%；当样本数进一步提高时，分类精度没有明显改变。证明本方法可以在高光谱图像标签样本少的情况下，取得较高的分类精度。

With the popularization of hyperspectral image data and the advancement of deep learning in the field of image processing, how to further improve the classification accuracy has become a concern. Hyperspectral images are influenced by "homologous" because of the high redundancy between each other. In addition, the insufficient of annotation has seriously limited the development of deep learning in hyperspectral images. This experiment proposes to use ACGAN to classify and use the confrontation learning between the generator and the discriminator to learn how to process the spectral information of hyperspectral images, and to study the number of training samples. Relationship with classification accuracy. The study found that when training samples accounted for 5% of the total sample, ACGAN's classification accuracy on PaviaU has reached 96% while on Indian Pines has reached 98%; when the number of samples is further improved, the classification accuracy has not changed significantly. It is proved that this method can achieve high classification accuracy in the case of low-spectral image label samples.