卫星遥感图像的混合像元分解理论及其应用研究

项目名称： 卫星遥感图像的混合像元分解理论及其应用研究

项目编号： No.60872083

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 自动化技术、计算机技术

项目作者： 孙卫东

作者单位： 清华大学

项目金额： 30万元

中文摘要： 本项目以多光谱及高光谱卫星遥感图像为对象，就多光谱图像特征端元提取、高光谱图像端元提取、明确物理含义下的混合像元分解、端元个数估计等核心问题展开了研究。研究成果包括：1)给出了特征端元的新概念，提出了可用于多光谱图像的特征端元提取与混合像元分解方法；2)从几何凸集描述与统计描述两个侧面，提出了最小距离限制的非负矩阵分解算法、基于盲信号分解的混合像元分解算法以及分层查找的端元提取算法；3)针对非负矩阵分解中解的非唯一性问题，提出了包括端元中心距离约束、丰度值空间连续性约束与丰度值稀疏性约束在内的三种正则项约束方法，并给出了基于Nesterov方法的优化算法与线性混合像元模型物理约束的高效处理方法；4)针对端元物理含义的诠释问题，提出了单光谱吸收特征提取方法、光谱簇吸收特征提取方法以及光谱相似性度量方法，并在基础上提出了一种全新的目标光谱指导下的像元分解方法；5)针对端元个数确定问题，提出了一种基于主成分分析的快速混合像元分解方法、以及将丰度值稀疏性约束作为正则项的思路；6)在新型应用形态方面，提出了基于混合像元分解的亚像元级分类精度评估方法、薄云下光学遥感图像的无监督和半监督恢复方法。

中文关键词： 卫星遥感图像；端元提取；混合像元分解

英文摘要： Focusing on multispectral and hyperspectral remotely sensed satellite images, several key issues including feature-endmember extraction for multispectral images, endmember extraction for hyperspectral images, spectral unmixing with exact physical implication, estimation of the endmember's number have been studied in this project. The major contributions of this project including: (1) A new concept called as feature-endmember has been given, and the corresponding feature-endmember extraction and spectral unmixing method which can used even for multispectral images has been proposed; (2) From geometrical and statistical aspects, a minimum distance constrained nonnegative matrix factorization method, a spectral unmixing method based on Blind Signal Separation and a hierarchical approach for endmember extraction have been proposed; (3) For the non-uniqueness problem of nonnegative matrix factorization, endmember centroid distance, spatial smoothness and the sparsity of abundance map have been introduced as three regularization terms, an optimization framework base on Nesterov's method and a finite step terminated projection method have also been proposed; (4) As for the physical implication of the extracted endmembers, feature extraction both for single spectrum and spectra cluster data and furthermore a spectral similarity metric method have been proposed, and base on these methods, a novel target spectra guided spectral unmixing method has been proposed; (5) For the estimation of endmember's number, a fast unmixing method base on the Principal Component Analysis and a regularization method based on the abundance sparsity have been proposed; (6) As new type of applications, a new accuracy assessment method for classification, an unsupervised image restoration method based on linear imaging model and a semi-supervised image restoration method based on nonlinear imaging model for thin cloud covered optical images have been proposed respectively.

英文关键词： Satellite images; Endmember extraction; Spectral unmixing