基于非线性流形学习的极化SAR特征提取与匹配技术研究

项目名称： 基于非线性流形学习的极化SAR特征提取与匹配技术研究

项目编号： No.41501414

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 天文学、地球科学

项目作者： 涂尚坦

作者单位： 上海卫星工程研究所

项目金额： 20万元

中文摘要： 在极化SAR图像检测、分类、检索和识别等应用中，对目标进行特征空间中的相似性度量是这些应用面临的共同技术问题，提取出能够全面表征不同类别目标鉴别信息的极化本征特征，是解决这一技术问题的关键。本项目拟采用非线性流形学习的技术，恢复出内蕴在原始高维极化观测空间中的低维极化流形结构，并在维数约减的同时保持不同极化目标的鉴别信息，提取出目标的极化本征特征；然后针对极化本征特征进行本征维数和物理意义的挖掘，建立目标的极化本征特征集；最后推导出非线性极化流形学习的显式映射，以解决新来测试样本的本征特征提取问题，最终实现极化SAR图像在本征特征空间中的特征匹配应用。该项目的研究拟寻找并发现极化SAR图像中有利于目标相似性度量的本征特征，挖掘出其内在的物理意义，建立典型目标的极化本征特征集并形成基于本征特征的极化SAR目标特征匹配应用方案，为我国即将到来的极化SAR图像大规模应用提供技术支撑。

中文关键词： 特征提取；目标识别；维数约简；流形学习；主动微波遥感

英文摘要： In the application of polarimetric SAR(PolSAR) target detection, classification, retrieval and recognition, similarity measurement in the feature space is a common problem. The key point to solve this problem is how to extract the polarimetric intrinsic features that can represent the discriminating information of different classes of targets. In this project we will restructure the low dimensional polarimetric manifold which lies on the original high dimensional polarimetric observation space by nonlinear manifold learning, meanwhile retain the discriminating information of different classes of polarimetric targets. Then the polarimetric intrinsic features will be extracted, meanwhile the intrinsic dimension and physical meaning of these intrinsic features will be researched. Based on the above steps, the polarimetric intrinsic feature database of targets can be constructed. For solving the problem of out-of-sample, the explicit mapping of nonlinear manifold learning must be found, then we can apply the feature matching in the polarimetric intrinsic feature space. The research aims to search the intrinsic features which are beneficial to target similarity measurement, and then apply them in the polarimetric target feature matching. In the future, this work will provide technological sustentation for high-volume applications of polarimetric SAR by our country.

英文关键词： feature extraction;target recognition;dimensionality reduction;manifold learning;active microwave remote sensing