基于深度学习的高分辨率PolSAR影像暗目标判别

This project tries to discriminate the low backscattering object (LBO) by LBO detection, feature representation and LBO distinguish. There are three major drawbacks in current LBO classification methods. Firstly, the LBO is mainly dependent on the temporal and multi-source dataset. But the dataset is usually incompatible and cost which limits the LBO application. Secondly, the LBO distinguish features highly depend on the manual work and the number of features are too few to work well in complex scenes. Furthermore, the scene contextual information of high resolution SAR has not been used..In this project, the polarimetric features and contextual descriptor are combined to discriminate LBO. The high resolution polarimetric SAR is decomposed into dozens of feature by radar decomposition theory. The decomposed features are called feature cube which consists polarimetric and contextual information. Then, the deep learning algorithm is utilized to reduce the redundancy of feature cube and represented by more effective outputs. The feedback and Back-Propagation learning algorithms are also used to improve the LBO distinguish accuracy without the temporal and multi-source data auxiliary. This project will enhance the classification accuracy of LBO, such as cement-road, soil, water and shadow, etc.

本项目针对SAR影像普遍存在的暗目标混分现象，从目标散射特性与成因出发，开展暗目标的探测、特征学习与识别研究。现有方法在区分暗目标时过于依赖多时相、多源数据，而数据时效性与获取成本制约了该研究的开展；同时，SAR极化信息的利用程度依赖人为经验，人工指定的有效特征数量少，难以适应复杂场景；此外，SAR影像蕴含的场景信息也没有得到充分挖掘，暗目标地物环境语义没有得到充分利用。.针对以上问题，在暗目标判别中将极化与场景特征结合是一个可行的思路。本课题从高分辨率PolSAR目标分解理论出发，构建富含极化、场景信息的高维度散射特征集，采用深度学习剔除特征集冗余信息并发掘其蕴含的极化-场景特征，在反馈学习与后向传播学习的指导下完成暗目标探测与分类，实现不依赖多时相、多源数据的暗目标判别。本项目的开展有望提升SAR影像水泥路、裸土、水体、阴影等典型暗目标解译水平，具有重要的理论意义和应用价值。

为了避免传统暗目标识别方法对多时相、多源数据（光学影像+SAR影像+激光雷达影像）的依赖，本课题对高分辨率极化合成孔径雷达（PolSAR）数据构建散射特征集，在深度学习框架下实现极化特征表达与暗目标探测，经过3年的项目研究，课题团队对以下问题开展了研究：1）对暗目标散射机理进行了研究，分析了暗目标后向散射特性与成像环境的相互关系，揭示了暗目标散射特性时空多变的原因；2）对暗目标极化散射特性进行了研究，在线性拉普拉斯判别（LDLE）框架下，利用流行学习、深度学习方法对原始高纬度散射特征进行降维，提高了不同类型暗目标判别分类的精度；3）利用卷积层、池化层、激活层等构建了暗目标提取、暗目标判别的深度学习神经网络，对高分辨率PolSAR影像上的暗目标提取、分类精度达到90%；4）为了降低相干斑噪声对深度学习神经网络的判别精度，课题组在具体研究中还提出了保持散射机理的极化滤波方法；此外，为了探索项目研究成果的应用潜力，课题组还将暗目标判别结果应用于PolSAR影像极化定标与系统热噪声估计中。经过本项目的支持，课题团队实现了暗目标地物的精细化识别，为道路线识别、土壤耕地面积估算、洪水灾害范围评估等应用提供了科学支持和保障。