【作者】 刘强；

【导师】 田国良； 李小文；

【作者基本信息】 中国科学院研究生院（遥感应用研究所） ， 地图学与地理信息系统， 2002， 博士

【摘要】 表面温度是地表能量平衡中的重要输入参数，自然状态下地表温度的分布极不均匀，像元尺度的平均温度缺乏对真实温度分布的代表性，很多关于地表能量平衡过程的模型都需要输入组分温度以提高精度。利用多角度或多光谱热红外遥感数据，在有地面实验数据支撑的重点实验区域实现组分温度反演的条件已经具备，本文从实践上证明这一套算法的可行性。 在前向模型上，本文指出热红外波段分辨率的遥感像元通常是多种覆被类型的混合，混合像元的分解应采用“像元-端元-组分”这样的层次结构。现有的辐射方向性模型通常针对单一的覆被类型，是端元模型。根据像元的尺度结构，端元模型复合成为像元模型可以采用线性组合的形式。另外，组分温度反演需要以最简洁的形式刻画地表辐射方向性，本文通过直接改写辐射传输方程的方式把一个典型的BRDF模型——SAIL模型扩展到热红外波段，可以用于计算水平均匀冠层的热辐射方向性或组分有效发射率。 反演算法上，本文首先以像元结构参数和材料参数为媒介把可见光／近红外波段的信息引入组分温度反演，并且利用了可见光／近红外波段与热红外波段分辨率的差异实现“像元-端元”这一层次的分解，利用BRDF模型反演实现“端元-组分”这一层次的分解。接下来，在组分有效发射率模型及其矩阵表示的基础上，结合贝叶斯反演理论，实现了组分温度反演，反演结果以后验概率均值和标准差的形式给出，并且以后验标准差相对于先验标准差的减小作为参数在反演中获得信息量的度量。对一个典型案例做的误差和敏感度分析表明：(A)多角度数据比单一角度的多光谱数据更适于组分温度反演；(B)在数据或模型参数有误差的情况下，增强先验知识约束能明显改善反演结果；(C)如果有先验知识的支持则单一角度的多光谱数据也可以用于反演组分温度。因此本文提出利用参数的空间相关性改进先验知识的迭代反演方法，新先验知识的生成过程需要用到上一次反演的后验均值和标准差。 为了将组分温度反演用于实测遥感数据，本文在第四章中先用一定篇幅介绍数据的基本处理，尤其重点介绍了多角度遥感图像的自动配准算法。该算法针对机载多角度遥感图像的特点而设计，结合了金字塔式配准框架和小波变换、多元变量的相关系数、B-样条等数学工具，可在图像间有复杂局部形变以及光谱差异的条件下实现配准。接下来，文中重点介绍了用多角度和多光谱数据反演典型农业区组分温 中国科学院遥感应用研究所博士学位论文 度的实践。AMTIS是我国自行研制的机载多角度成像系统，其数据采集和处理全部 是课题组完成的；ASTER是Terra上唯一的高分辨率多光谱传感器，其标准数据产 品己包含基本的几何和大气校正。这两种数据分别用于组分温度反演，组分温度的 分离都基本成功。我们还给出地面同步测量数据对反演结果的验证，结果虽然不很 完整，但还是可以借此分析存在的问题。 最后，本文在可见光／近红外波段分辨率足够高的假设下，通过人为降低热红外 波段分辨率的方式，分析了热红外波段分辨率对组分温度反演结果以及像元平均温一 度的影响，指出数据的空间分辨率对反演结果的影响是渐变的，而组分温度相对于 像元平均温度则有质的区别。本文结果与MODIS－LST、ASTER－TES标准产品的做 了对比，存在的差别一部分来源于大气校正或地表发射率的处理方法的差异，但最 主要还是因为数据分辨率的差异以及组分温度与像元平均温度的不同。更多还原

【Abstract】 Land surface temperature (LST) is an important parameter in the study of energy balance/exchange over land surface. Because natural land surface is usually heterogeneous and not isothermal, the pixel-mean temperature cannot adequately represent the actual thermal state of the surface. Many applications will appreciate component temperature input for better results. With the support of ground-based experiments and other auxiliary information, it is possible to retrieve component temperature from multi-angular or multi-spectral thermal infrared (TIR) remote sensing data. Presented in this paper is the theory, as well as practice, of component temperature inversion.Because a remote sensing pixel in TIR band is usually a mixture of several land cover types, the "pixel-subpixel-component" structure was chosen to be the framework of forward model and inversion strategy in this paper. Here, a "subpixel" was defined as the fraction inside one pixel that is of pure land cover type instead of its usual meaning that is ambiguous between "component" and "endmember". It was pointed out that the "subpixel-component" structure is usually complex and should be modeled with 3-D model, while "pixel-subpixel" structure can be approximated with 2-D model. So, most complex models that describe the directionality of land surface are subpixel models, and they can be integerated into pixel models with linear combination. In the component temperature inversion algorithm, the "pixel-subpixel" decomposition was done by exploiting resolution difference between TIR and VNIR (visible and near infrared) data; and the decomposition of "subpixel-component" was done by inverting BRDF model with VNIR data.It is important that the forward models used in inversion should be effective as well as simple. So, a well-known BRDF model, the SAIL model, was adapted and extended to thermal bands by directly modifying its RT differential equations. The extended SAIL model can be used to predict directional thermal radiance or to derive component equivalent emissivity for horizontally homogeneous canopy.Inversion algorithm was built on the concept of component equivalent emissivity and its matrix representation. Bayes inference was also incorporated into the frame so that a priori knowledge could be introduced to be constraint of result. It was also shown in the error analysis that a priori information was very critical for inversion of those unknown variables that are not sensitive to observation. A priori information wasexpressed as mean and std. deviation and the result of inversion was expressed as posterior mean and posterior std. deviation. The gained information for a variable could be measured with the decrement of its std. deviation before and after inversion. Based on this measurement, it was shown that (1) component temperature got more information with multi-angular data set than with single-angle and multi-spectral data set; (2) single-angle and multi-spectral data set could be used to retrieve component temperature only if a priori information is enhanced. Because component temperature varies much throughout a scene of remote sensing image, globally defined a priori information can never be very precise. So it is necessary to give separate a priori mean and std. deviation for each pixel. Exploiting spatial correlation of the unknown variables, an iterative upgrade method is proposed to extract new a priori mean and std. deviation from a neighborhood of previous inverted posterior mean and std. deviation.As part of work in component inversion practice, chapter 4 gave some details for basic data processing and correction. Mainly introduced here was an automatic image registration algorithm, which was specially designed for airborne multi-angular remote sensing images. It cooperated with wavelet decomposition, multi-variant correlation coefficient and B-spline warping function in the frame of pyramidal resolution matching; and it performed well with images that contain localized distortion and spectral change. After registration, su更多还原