摘要
本文引用独立成分分析与盲信号分离的理论,从遥感高(多)光谱数据的基本统计特征出发,对其概率密度分布作出了分类与解释,并同图像数据的背景与异常建立了联系。在此基础上,对高(多)光谱数据点阵分布的空间几何结构进行了深入的研究分析,推断出遥感高(多)光谱数据集合的高维空间属于低维几何结构-"超平面"形态,而包含蚀变信息在内的异常点群通常会游离在"超平面"之外。然后,对主成分分析(PCA)的信号-噪声模型加以引申,提出了遥感图像多元数据集合高维空间的背景-异常信号子空间可划分的概念,并给出了子空间划分的阈值估计方法。同时,探讨了遥感图像的端元数目、多波段数据集合的本征维数、主要背景地物数目三者之间的关系;通过西藏驱龙地区两种类型遥感数据的实例分析,说明了本文所讨论的光谱数据空间的低维结构以及背景-异常子空间模型在遥感高(多)光谱数据分析应用中的正确性与实用性。研究结果表明:尽管不同自然景观区的遥感图像的光谱变化复杂,而它们的光谱数据空间属于低维几何结构,以及背景-异常(含噪声)子空间的可分性是其具有共性的本质特征。在统计意义上蚀变异常在遥感高(多)光谱数据集合中是可识别的。
The article quoted a number of important theory thoughts of independent component analysis and blind signal separation, classified and explained the distribution of probability density from the basic statistical characteristics of remote sensing hyper （multi-） spectral data and, established the connection between background and anomaly of image data. On this basis, it deeply analyzesed the spatial geometry structure of hyper （multi-） spectral data lattice distribution, concluded that high-dimensional space of remote sensing hyper （multi-） spectral data belong to low-dimensional geometry structure - ＂hyperplane＂ shape, and anomaly point group which includes alteration information usually dissociates out the ＂hyperplane＂. Then, the paper extended the signal - noise model of principal component analysis （PCA）, proposed the dividable concept of background-anomaly signal sub-space of remote sensing image multi-data set high-dimensional, and gave the threshold estimated method of sub-space division. At the same time, it discussed the relationship among the pixel number of remote sensing image, the intrinsic dimension of multi-band data set, and the main background object number. Finally, through analysing two types of remote sensing data for Qulong area, Tibet, it explained the correctness and practicality of spectral data space lowdimensional structure and the application of analysing remote sensing hyper （multi-） spectral data to background - anomaly sub-space model. The results show that： although the spectral change of different landscape areas is complex, the low-dimensional geometry structure which spectral data space belongs to, and the practicality of back- ground - anomaly （including noise） sub-space are the essential characteristics which they all possess. In a statistical sense, the alteration anomaly of remote sensing hyper （multi-） spectral data set is identifiable.
出处
《地球信息科学》
CSCD
北大核心
2009年第3期282-291,共10页
Geo-information Science
基金
国家高科技研究发展计划(863)资助项目(2006AA06Z112)
关键词
高(多)光谱数据
蚀变信息检测
光谱数据空间结构
盲信号分离
主成分分析
Hypor （multi-） spectral signal separation
principal component data
alteration information detection
spectral data space structure
blind analysis
