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

【作者】             罗文斐                                                                                                                                             

【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，光谱解混，零空间端元提取，独立成分分析，空间分辨率 

【Key words】   Hyperspectral Remote Sensing, Spectral Unmixing, Null Space, Endmember Extraction, Independent Component Analysis, Spatial Resolution 

【中文摘要】

高光谱遥感图像所提供的丰富光谱信息可以使人们更加精确地区分地物类型和提取更为详尽的地物理化参量，但由于遥感器技术条件的限制，相对于其它光学数据，高光谱图像往往具有相对较低的空间分辨率，导致了图像中的混合像元问题比较严重。另一方面，高光谱数据所具有的丰富光谱信息也为人们开展像元分解与光谱解混提供很好的客观条件，这就使得光谱解混成为了高光谱图像分析中的一个关键问题。混合像元机理和针对不同尺度图像的光谱解混技术研究，将为解决高光谱图像的混合像元问题，进而提高高光谱数据分类和目标识别的精度起到关键作用，对高光谱遥感技术和应用的发展具有重要意义。

        本论文在系统归纳了高光谱图像线性光谱解混的整体技术流程基础上，对其中端元提取以及独立成分分析等关键技术作了深入分析和探讨，解决了目前在线性光谱解混技术中存在的问题，包括：基于几何学端元提取算法的有效性问题、独立成分分析的成分独立性条件和不确定性问题，并发展了具有一定灵活性的、能够支持多种单次端元提取策略的新算法。同时，结合图像数据的多尺度效应，对线性光谱解混技术进行了不同空间分辨率图像的适应性研究。论文的主要研究内容和成果如下：

1. 分析了高光谱遥感图像光谱解混技术的发展现状，系统总结并提炼出了高光谱图像线性光谱解混的整体技术流程，实现了对传统的高光谱图像线性光谱解混技术路线的扩展，涵盖了线性光谱解混过程中所涉及到的各主要方面问题，并融入了更多先进的技术和理念。

2. 从一个新的角度——零空间——探讨了基于几何学的线性光谱解混算法，利用零空间在投影距离计算方面的优势，提出了最大零空间投影距离算法，并结合零空间与端元之间的关系对算法进行了严格的数学证明，为基于子空间距离的端元提取算法提供了重要的理论依据。讨论了零空间的凸不变性，提出了零空间光谱投影分析的端元提取思路，改变了传统的以当前端元作为新端元衡量基准的逐端元提取思想，通过制定和选择不同的单次端元提取策略，设计了多种零空间光谱投影分析算法，实验表明这些算法在不同的条件下具有不同的应用特点。

3. 针对独立成分分析在高光谱图像光谱解混中存在的成分独立性问题，提出了成分相关性最小化的解决思路，并探讨了相关性最小化的最佳夹角问题。在此基础上发展了基于最佳夹角角度修正的斜交独立成分分析算法；针对独立成分分析的不确定性问题，提出了端元定量化以及丰度定量化算法，对幅值不确定性进行修正，并利用初始估计端元，来消除次序的不确定性；最后，根据上述的改进方法，提出了高光谱图像定量化独立成分分析的总体技术方案，在一定程度上改善了独立成分分析在高光谱图像光谱解混中存在的问题。

      针对地物的多尺度和遥感图像的多分辨率问题，给出了高光谱图像光谱解混技术的空间分辨率适用性评价方法。开展了不同空间分辨率下的不同像元解混技术的适应性研究，研究结果可为高光谱图像光谱解混技术的进一步改进提供参考，并为基于地面成像光谱仪的像元光谱解混技术研究提供了关键技术支持。

【Abstract】

Hyperspectral remote sensing has been developed in hundreds of narrow contiguous bands and may provide a wealth of spectral information for data exploitation. However, due to the limitation on sensor cost, the resolution cell corresponding to a single pixel in hyperspectral imagery often contains several kinds of distinctsubstances. This phenomenon is a little more serious than other optical data. In this situation, a great challenge in information extraction from hyperspectral remote sensing data is decomposing a mixed pixel into a collection of endmembers and their corresponding abundance fractions, namely spectral unmixing. However, the wealth of spectral information in hyperspectral data enhances the ability for spectral unmixing. The issues of subpixels and mixed pixels analysis is one of the most focused applications of hyperspectral remote sensing data analysis. The research on the mechanism of mixed pixel and the adaptability of the spectral unmixing techniques in different scale data is the key point to solve the mixed pixel problem and further to improve the accuracy of classification and target detection.

      Based on an integrated workflow of end-to-end stages of spectral unmixing, this dissertation makes a study of Endmember Extraction (EE) and Independent Component Analysis (ICA) for linear spectral unmixing techniques, addresses the problems arising in the following: the validity of geometric based EE algorithms, the violation of independent restriction in ICA and the ambiguities of ICA. And it proposes a group of novel flexible algorithms by making different strategies for one step EE. Moreover, considering the multi-scale effect of hyperspectral remote sensing data, this dissertation engages in the adaptability research of spectral unmixing techniques based on different spatial scale.

The topics addressed in this dissertation and its main contributions are summarized as follows:

1. Gives a survey for most recent linear spectral unmixing techniques and extends the traditional workflow of end-to-end stages to one that includes state of the art techniques of linear spectral unmixing which makes it more distinct from ordinary ones.

2. Discusses the geometry-based EE techniques from the viewpoint of null space. Two advances based on the null space are presented. First, the subspace projection distance can be easily obtained by null space and a maximal null space projection distance EE algorithm is developed. It is mathematically proven right depended on the relationship between null space and endmember and provides a mathematical basic for the maximal subspace projection distance EE algorithms. Second, null space spectral projection has the invariability of convex. In the light of this characteristic of null space, faster and more flexible method is proposed, called null space spectral projection method. Through making strategies for one step EE, a group of algorithms are designed which present their inimitable virtue in this dissertation.

3. On the issue of ICA for spectral unmixing, component independent restriction is violated due to the abundance constrains. This dissertation presents a solution, called correlation minimizing, to this embarrassment. An optimum angle is provided under the condition of correlation minimizing. Two kinds of modified ICA algorithms, called Sequence oblique-ICA Algorithm (Sob-ICA) and Parallel oblique-ICA Algorithm (Pob-ICA), are presented to adjust the searching direction to the optimum angle for the components. Futhermore, to against the uncertainty of ICA, this dissertation builds up two quantificators to correct the endmember matrix and the abundance respectively and uses an estimated endmembers obtained by EE techniques, such as null space spectral projection algorithm, for the initialized endmember matrix. A workflow of quantified ICA analysis in hyperspectral imagery is presented and it improves the accuracy of spectral unmixing.

      Finally, this dissertation presents a technique for evaluating the linear spectral unmixing algorithms, considering the multi-scale effect of hyperspectral remote sensing data. A research on the adaptability of linear spectral unmixing algorithms to different levels of spatial resolution is carried out. The result can provide technical support for the ground spectroradiometer based spectral unmixing research and produce profound thoughts for imporving the spectral unmixing techniques.