中国科学院遥感与数字地球研究所-高光谱遥感应用技术研究室推荐内容
【2008博】基于光谱维变换的高光谱图像目标探测研究
|
【作者】 刘翔 【导师】 童庆禧；张兵；郑兰芬 【 学位年度 】 2008 【论文级别】 博士 【关键词】 高光谱遥感，目标探测，噪声评估，光谱维变换，白化空间，子空间投影 【Key words】 Hyperspctral Remote Sensing, Noise Estimation, Target Detection, Spectral Dimensionality Transform, Whitened Space, Subspace Projection 【中文摘要】 高光谱遥感是20世纪80年代逐渐发展起来的新兴遥感技术，它在空间成像的同时记录下成百个连续的光谱通道，从图像的每一个像元均能提取出一条连续的光谱曲线，代表了像元对应地面瞬时视场区域中的地物平均反射光谱信息。因此，本文研究的高光谱图像目标探测正是基于高光谱图像光谱维分析以实现特定目标的快速和自动检测。 基于高光谱图像对特定目标的探测是一个涉及光谱成像和像元光谱分析的复杂过程，本论文研究了多种光谱维去相关和光谱维投影变换方法。在系统总结前人研究成果的基础上，研究了与高光谱遥感目标探测相关的探测算法设计、种类划分及其模型评价等问题。最后针对概率密度模型和子空间模型进行了目标探测算法的讨论、设计和实验。 论文的主要研究工作和成果包括在以下五个方面： 1. 将MNF变换和NAPC变换当成两种不同的数学变换方法展开讨论，证明两者的等效性。随后通过具体实验阐明：高光谱数据样本的空间分布对基于MNF变换的特征提取结果有着显著影响。随后针对地物混杂下分类效果不佳问题，提出了一种通过优化噪声协方差矩阵评估的MNF改进算法。该算法针对高光谱图像中很难找到同类均匀地块的情况，能够更加有效的进行数据降维和噪声分离处理。 2. 为了综合评估各种高光谱图像目标探测算法性能，本文总结了应用于各个领域的探测算法性能指标，如虚警概率和检测概率、恒虚警概率、超标概率、信杂比、ROC曲线等，并在实验中依据该指标因素全面地评价了各种探测算法。 3. 本文将各种基于概率密度模型探测算法在白化空间中统一描述，针对白化空间中目标和背景分布之间决策边界选择问题，打破背景在白化空间中服从正态分布的假设，提出了基于椭圆轮廓分布的双曲线决策门限型探测算法和抛物线决策门限型探测算法，实验证明这两种探测算法在检测概率上优于ACE探测算法。 4. 将各种基于子空间投影的目标探测算法在光谱向量欧氏空间中统一描述，证明部分目标探测算法的等效性，并提出了一种基于斜子空间投影的广义化似然比(GLRT)探测算法，通过实验证明，在获取背景特征向量不能够完备表述背景子空间的时候，该探测算法有着优于GLRT探测算法的性能。 5. 本文在试图将MNF变换应用于以CEM算法为代表的概率密度模型算法和以OSP算法为代表的子空间模型算法中，发现MNF变换中的典型特征向量能够准确表述背景子空间，而非图像子空间。因此，将其用在基于背景正交子空间投影的目标探测算法有着更好的探测性能，于是提出了一种基于MNF变换的非监督OSP算法，并取得了较好的实验结果。
【Abstract】 Hyperspectral imaging is a new and growing technology with the development of airborne and spaceborne remote sensing from the 80s’ of 20th Century. Each pixel in the hyperspectral image is an observation vector and it represents a reflectance spectrum of the materials in the ground area in the Instantaneous Field of View. The Target Detection on hyperspectral imagery is a technique by which the information is obtained based on the transform of spectral dimensions. It can help the expert in many kinds of fields to find the target in the image on the one hand. On the one hand it can replace the Expert Decision to finish the target detection tasks by the method of Artificial Intelligence. It is a flexible remote sensing procedure to detect the specified targets through hyperspectral images. This dissertation first introduces each detail of this procedure to validate the feasibility of target detection by hyperspectral remote sensing. Many kinds of decorrelation transforms and subspace projection transforms both in Spectral Dimensions are studied for the purpose to understand the hyperspectral analysis in Spectral Dimensions well. And then based on the research of scholars in many fields, this dissertation introduces the contents about mathematical models, design procedures, category and performance estimations of different detectors. Then pointing at two different mathematical models, the probability statistics model and the subspace model, the discussions, designs and experiments to target detection algorithms are opened out. The main fruits of this dissertation are as follows: 1. To affirm the equivalence of the MNF and NAPC transform, they were discussed as two different mathematical transform methods and finally this conclusion is proved. Then it is proved that how the sample classes are scattered have great influences on classification accuracy when using MNF transform. An improved MNF method, in which the Noise Covariance Matrix estimation method is optimized, is introduced for this problem. 2. For the purpose that to estimate the performance of different hyperspectral target detectors, this article sum up the detector performance parameters of different fields, such as the Probability of False Alarm and the Probability of Detection, the character of Constant False-Alarm-Rate, Signal to Interference-plus-Noise Ratio ROC curve and so on. And many detectors are estimated in accordance with these performance parameters through experiments. 3. The detector algorithms based on the probability statistics model are uniformly described in the whitened space. Pointing at how to select a proper decision sufface between targets and backgrounds in the whitened space, the hypothesis that the background submits to normal distribution is denied in this dissertation, and two algorithms based on Elliptically Contoured Distribution (ECD) function are introduced. They are named ECD Detector with Hyperbola Threshold and ECD Detector with Parabola Threshold. They both perform better than ACE detector on Dection Probability when testified by expermiments 4. The detector algorithms based on the subspace model are uniformly described in the Euclidean space in which some target detector are proved equivalent. And a new Generalized Likelihood-Ratio algorithm based on Oblique Subspace Projection is presented. Testified by expermiments, this detector performs well than GLRT detector when the vectors of the background signature are not enough to describe background subspace. We try to use the MNF transform in the CEM detector and OSP detector. And find that the Eigenvector of MNF can describe the background subspace accurately, but not the image subspace. So MNF can be used in the detectors which are based on background subspace projection. A new Unsupervised OSP algorithm based on MNF transform is presented.
(责任编辑：admin) |
