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】             张兵                                                                                                                                            

【导师】             童庆禧；郑兰芬

【 学位年度 】    2002

【论文级别】      博士    

【关键词】          高光谱遥感，时间数据，空间数据，信息提取  

【Key words】   Hyperspectral remote sensing, Temporal data, Spatial data, information extraction

【中文摘要】  

遥感技术的发展一直伴随着空间分辨率与光谱分辨率的进步。传统的多光谱扫描仪对光谱曲线的采样是零散的，从可见到短波红外它通常只记录10 个左右的光谱波段，其光谱分辨率也在0.10μm 数量级。而高光谱成像或成像光谱却能够得到上百通道、连续波段的图像，从而可以从每个图像像元中提取一条完整的光谱曲线。

        高光谱遥感将确定地物性质的光谱与确定地物空间与几何特性的图像有机地结合在一起。从空间对地观测的角度来说，高光谱遥感信息无论对地物理化特性的深层探索，还是对地物间微小差异的精细识别，以及对自然界的知识发现，都为人类提供了前所未有的丰富信息。随之而来的问题就是，面对如此多的波段、如此海量的光谱图像信息如何去处理、如何从中提取我们需要的信息，黑白图像以及多波段彩色图像的一些处理算法甚至主要处理和分析手段已经无法满足高光谱遥感信息的处理要求。高光谱图像立方体图谱合一的特点，要求人们从光谱维去理解地物在空间维的变化，人们对二维空间图像的处理与分析需要转化成对每个像元所提取出的光谱曲线的处理与分析。而本论文的核心正是围绕着高光谱数据处理与分析这个核心，从系统性与创新性的角度展开相关论述。

        本篇论文在第一章高光谱遥感综述的基础上，利用第二章和第三章，就高光谱数据的定量化和参量化、高光谱图像分类与地物识别这两个关键的高光谱信息处理技术环节进行论述。它们在总结国内外研究成果的基础上，提出了作者自己的研究思路和研究成果，这也为后两章研究内容的展开提供了数据处理和分析的技术铺垫。第四章和第五章就时间和空间信息辅助下的高光谱遥感数据挖掘问题进行系统化地研究和论述，这两章内容拟在探讨以高光谱图像数据为核心的前提下，面对不同的问题，如何从不同角度发挥空间和时间辅助信息的作用，以更加有利于高光谱遥感应用研究的发展。归纳起来，本研究在高光谱信息处理与提取领域取得了如下几点进展：

（1）在光谱特征选择方面，提出了高光谱图像波段选择和目标在图像中快速查找技术。在光谱特征提取和光谱减维方面给出了多种光谱参量化方案。结合高光谱地质遥感和地层分析的应用实际，首次提出了光谱柱状图的概念。它通过变差分析技术放大了相似沉积地层的光谱差异，将光谱曲线变换成彩色光谱条码，从而建立起新疆吐鲁番背斜14 套地层的光谱柱状图。

（2）本论文在实践和分析前人高光谱图像分类方法的基础上，提出了特征优化的专家决策分类算法。这种算法主要体现了两大原则，一是基于待分类别的光谱特征优化与参量化原则，二是类别判定中的模糊定义与专家决策原则。论文给出了这种算法的普遍性运算流程，并通过与其它几种方法的对比显示了它的优越性。（3）本论文在高光谱伪装探测方面比较了几种不同伪装材料的光谱曲线数据，说明了在短波红外鲜活植被所特有的液态水吸收光谱特征是很难模拟的，这将成为植被伪装识别的重点。同时，本文在北京亚运村地区基于凸面几何体投影变换技术成功地完成了建材市场屋顶板材的高光谱探测。

（4）本论文在植被光谱特征分析与高光谱植被指数构建基础上，将多时相的高光谱图像立方体变换成多时相指数图像立方体（MIIC），并利用MIIC 模型分析了日本生菜、中国大白菜、氮与水胁迫下的小麦生化参量时空变化规律。

（5）本论文在空间信息辅助下的高光谱数据挖掘方面，提出了基于图像光谱复原的空间域遥感数据融合模型、像元空间关联分析模型、图斑级光谱分解与分类模型、DGM 辅助高光谱图像分析模型。它们分别从四个方面论述了空间辅助信息在高光谱遥感数据处理与应用中的作用。 

【Abstract】

From the beginning of remote sensing, imaging technology has advanced in two major ways: one is the improvement in the spatial resolution of images; another is the improvement in the spectral resolution of images. Conventional multispectral scanners record up to 10 or so, spectral bands with bandwidths on the order of 0.10μm in visible to short wave infrared bands. Furthermore, hyperspectral imaging, or called imaging spectrometry, can acquire images in hundreds of registered, contiguous spectral bands such that for each picture element it is possible to derive a complete reflectance spectrum.

      Hyperspectral remote sensing effectively make the spectral feature and geometric characters of objects together. From the view of earth observation from space, hyperspectral data provide human being more abundant information, not only in the deep explorations of object’s physical and chemical characters, but also in the precise classification of different objects and knowledge innovation. In case of so much spectral bands and such huge quantities of data, some conventional data processing methods cannot play good roles. Aiming at the hyperspectral image cube, the understanding and data processing in image spatial dimension must be changed to that completed in the spectral dimension.

      This dissertation is just concentrated on above aspects and evolved in the systematic and innovative views. This dissertation begins from the introduction on hyperspectral remote sensing technology. In the second and third chapters, two key points in hyperapectral data processing and analysis area, hyperspectral data calibration and parameterizationand, and hyperspectral image classification and identification, were dissertated. The fourth and fifth chapter pays more attentions to the hyperspectral data mining supported by the temporal and spatial information. In general, this study has some advantages as follows:

(1) As for spectral feature selection, spectral bands selection and objects quickly finding in image cube were provided. On the other hand of spectral featureextraction, several selections of spectral parameterization were also provided. Considering the hyperspectral geological remote sensing, stratum spectral histogram was established specially for 14 strata in Tulufan anticline.

(2) After discussion on the traditional image classification, a new method, Expert Decision Classification Based on Feature Optimization, was provided here. It is designed out in accord with two principles: one is the spectral feature optimization and parameterization, another is fuzzy and expert decision in pixel identification. Comparing with other method, this method can acquire more accurate classification results.

(3) Several spectra of man-made camouflage materials were provided here. In the SWIR, the position and relative intensities of the major absorption features associated with water are difficult to duplicate due to the complex architecture of vegetation. In addition, convex geometry projection was successfully used in the different metal material detection.

(4) On the bases of vegetation spectral analysis and hyperspectral vegetation indices, Multi-temporal Indices Image Cube was put forward and used in the dynamic growing analysis of Japanese lettuce, Chinese cabbage, and wheat stressed by nitrogen or water contents.

(5) In the area of hyperspectral data analysis supported by spatial information, Four application aspects were provided: spatial fusion based spectral reversion, hyperspectral data analysis associated with pixel position analysis, spectral unmixing and classification in the field patch units, and image classification supported by digital geomorphology model.