面向高光谱图像分类的小样本学习网络构建与优化研究

Deep neural networks have become a research hot topic in hyperspectral image (HSI) classification due to its ability to exploit the essential features of the image. The current networks training usually depends on a large number of labeled samples to resolve the problem of sample number to network reliability. For HSI classification, accurately marking the large number of samples over depends on expert knowledge and is time-consuming. To solve this issue, this proposal will propose learning network construction and optimization with small samples, and meanwhile, establish a feature learning and classification framework with high-reliability and strong generalization capability. The main contents are: 1) We take into account the limitation that intra-class variabilities increase with small samples, and design multi-resolution deep neural networks to reduce the dependency of networks reliability on the number of labeled samples; 2) In the design of the loss function, the deep neural networks optiminzation framework with prior knowledge is established to enhance the generalization ability of the networks with small samples. 3) Evaluation metrics without ground reference map are researched, and task-driven classification accuracy assessment is proposed to fill the gaps that the classiffication accuracy with insufficient labeled samples cannot be quantitatively evaluated. The research results of this proposal will contribute to the automation process of HSI classification, and have expansive application prospect in military and civil applications.

深度神经网络因具备挖掘图像本质特征的能力，已成为高光谱图像分类的研究热点。当前网络训练通常需要大量标记样本以解决样本数量对网络可靠性的制约问题。本项目针对高光谱图像分类中精确标记大量样本过度依赖专家知识和时间代价高的问题，提出小样本学习网络构建与优化方法，建立高可靠性、强泛化能力的高光谱图像特征学习及分类框架。主要内容有：1）在网络构建过程中考虑小样本条件下类内差异性增加的局限，设计多分辨深度神经网络模型，减小网络可靠性对样本数量的依赖；2）在损失函数的设计中充分挖掘图像的空-谱先验，建立知识先验的深度神经网络优化框架，提升小样本条件下网络模型的泛化能力；3）研究小样本条件下大量无标记样本的评估方法，提出任务驱动的分类精度评估策略，以填补标记样本缺失下的分类精度无法定量评估的空白。研究成果有助于推进高光谱图像分类的自动化进程，在军事应用和民用等方面都具有广阔的应用前景。

高光谱图像因其较高的光谱维数和复杂的场景信息，通常需要大量的标记样本以获取强鲁棒和强泛化的分类模型。但是高光谱图像具有较高的标注代价，标记的样本数量受到很大限制，因此小样本条件下的高光谱图像分类方法的研究更为迫切。本项目从突破网络可靠性及泛化能力过度依赖样本数量的问题出发，提出了多视角半监督高光谱图像分类模型，充分挖掘高光谱图像空间-光谱两个不同视角特征，建立轻量化半监督分类模型，以较少的时间代价实现小样本下较高的分类精度；提出了自适应测度学习模型，以有效度量高维空间中标记样本与未标记样本的距离；探索了高光谱图像分类网络的本质特性，对高光谱图像分类结果进行解释并给出了可视化展示方法，对后续研究具有指导意义。本项目成果有助于推动高光谱图像分类网络的自动化解译水平，提升了小样本条件下高光谱图像分类网络的精度，具有重要的学术价值和应用前景。