基于深度对抗度量网络的异源极化SAR图像分类方法研究

Polarimetric synthetic aperture radar (SAR) is a hot topic in the field of remote sensing. Heterogenous polarimetric SAR image classification is a key technology for the practical application of SAR images. However, due to factors such as sensor and imaging mode, heterogeneous polarimetric SAR images have different feature probability distributions and inconsistent classification tasks, which make the existing methods fail to achieve high accuracy for heterogeneous SAR image classification. In order to overcome the above issues, this proposal studies heterogenous polarimetric SAR image classification based on deep adversarial metric networks, which aims to realize domain adaptation of feature probability distributions and transfer learning of heterogeneous categories. To solve the problem of different feature probability distributions, a weighted adversarial neural network is proposed to adopt adversarial learning for domain adaptation of features between source domain and target domain, which can improve the classification accuracy of polarimetric SAR image in the target domain. To solve the problem of inconsistent classification tasks, a similarity metric network with polarimetric information constraint is proposed to measure the similarities between heterogeneous samples of the target domain and each category of the source domain, which aims to utilize the similarities for classification of heterogeneous samples in the target domain. This proposal will study effective deep adversarial metric networks to improve the transfer learning capability of heterogeneous polarimetric SAR images, which can provide theoretical support for the applications of large-scale heterogeneous SAR image processing.

极化合成孔径雷达(SAR)是当前遥感领域的研究热点，异源极化SAR图像分类是SAR遥感走向实际应用的关键技术。但是，受传感器、成像模式等因素的影响，异源极化SAR图像存在特征概率分布差异、分类任务不一致的问题，现有方法难以实现高精度的跨数据分类。针对上述问题，本课题拟开展基于深度对抗度量网络的异源极化SAR图像分类研究，实现异源极化SAR图像的特征分布域适应、异类迁移学习。针对异源极化SAR图像特征概率分布差异的问题，提出权重对抗神经网络，利用对抗学习策略实现源域和目标域的特征域适应，提高目标域极化SAR图像的分类精度；针对异源极化SAR图像分类任务不一致的问题，提出极化信息约束的相似度量网络，结合极化先验进行目标域异类样本与源域各类别的相似性度量，实现目标域异类样本的分类。本课题研究的深度对抗度量网络可以提高异源极化SAR图像的迁移学习能力，为大规模异源SAR图像处理的应用提供理论基础。

合成孔径雷达（SAR）图像分类是遥感智能解译的重要内容，高精度极化SAR图像分类关乎着遥感对地观测的应用发展。异源极化SAR图像分类存在特征分布差异、分类任务不一致、训练样本不足等问题，现有方法难以实现高精度的异源SAR分类。针对此，本项目开展了极化SAR图像特征提取、小样本分类、异源迁移学习等方法研究，主要研究内容包括：（1）提出了基于超像素-超图特征增强网络的极化SAR特征提取方法，基于多尺度特征学习与双边滤波的极化SAR分类方法，提升了极化SAR图像地物分类的精度。（2）针对SAR图像训练样本不足的问题，提出了基于任务驱动对比学习的小样本遥感图像分类方法，基于伪损失置信度量学习的半监督小样本分类方法，解决了监督分类和半监督分类中训练样本有限的问题。（3）面向异源SAR遥感图像的迁移学习，提出了基于深度联合分布匹配网络的异源SAR特征域适应方法，多尺度对比域适应的无监督分类方法，提升了异源遥感图像迁移学习的性能。在本项目的资助下，项目组发表学术论文15篇，其中SCI论文9篇，EI论文6篇，申请国家发明专利9项，已授权3项，培养博士研究生2名、硕士研究生7名。研究成果能够提高异源极化SAR图像的迁移学习能力，促进了极化SAR图像在国防、农业、海洋探测等领域的应用。