高光谱图像分类作为实现高光谱遥感图像众多应用的先决条件, 一直是遥感领域的研究热点与难点。近些年, 各种深度学习网络模型方兴未艾并逐步应用于高光谱图像分类领域, 但是高光谱图像数据集往往存在着标记样本稀少、类别分布不均的问题, 导致网络在训练样本有限时容易出现过拟合现象, 取得较低的分类精度。为了解决以上问题, 本文深入调研了生成对抗网络并对其进行了改进, 同时与多任务学习结合, 提出了基于多任务生成对抗网络的高光谱图像分类方法, 通过有效训练生成器和判别器实现对于高光谱数据特征的深度挖掘, 其中生成器生成大量的虚拟样本以扩充训练集, 判别器依靠输入源判别任务、度量任务和地物分类任务的信息交互实现分类性能的提升; 所提出的方法在两个标准数据集上验证了各模块设计的有效性, 并证明该方法能够依靠样本扩充以及多任务学习架构表现出相对优异的分类性能, 并且对训练样本的数量表现出较低的依赖性。

As a prerequisite to realize many applications of hyperspectral images, hyperspectral image classification has been a hot and difficult research area in remote sensing. In recent years, various deep learning network models are in the ascendant and gradually applied to hyperspectral image classification. However, hyperspectral image datasets often have the problems of sparse labeled samples and uneven class distribution, which lead to the overfitting phenomenon of networks when the training samples are limited and achieve low classification accuracy. In order to solve the above problems, this paper deeply investigates generative adversarial networks and improves them, while combining them with multi-task learning to propose hyperspectral classification methods based on multi-task generative adversarial networks. The algorithm achieves deep mining of hyperspectral data features by efficiently training the generator and discriminator, in which the generator produces a large number of virtual samples to expand the training set, and the discriminator relies on the information interaction of the input source discrimination task, the metric task and the ground object classification task to improve the classification performance. The proposed method has validated the effectiveness of each module on two standard datasets, and demonstrated that the method can exhibit relatively excellent classification performance through sample expansion and multi-task learning architecture, with low dependence on the number of training samples.