×
【作者】 杨杭
【导师】 童庆禧；王晋年；张立福
【 学位年度 】 2011
【论文级别】 博士
【关键词】 TASI，温度，比辐射率，热红外高光谱，尺度
【Key words】 TASI; Temperature; Emissivity; Hyperspectral Thermal Infrared; Scale
【中文摘要】
航空热红外成像光谱数据具有极其广泛的应用需求。目前国际上，美国、芬兰、加拿大等发达国家已开展了航空热红外成像光谱的应用研究，其最核心的问题是温度与比辐射率的精确分离。本文以航空热红外成像光谱仪（TASI）为主要数据源，配以地面同步测量数据，围绕热红外温度与比辐射率的分离及尺度影响展开，研究取得了以下成果和结论：
1. 针对TASI数据的特点，建立了基于TASI数据的基本处理流程：辐射定标、大气纠正、几何纠正、图像镶嵌、温度与比辐射率分离，温度和比辐射率产品的应用。
2. 在TASI数据的温度与比辐射率分离算法研究方面：
在研究Aster_TES算法、alpha剩余法和ISSTES算法的基础上结合TASI热红外高光谱数据对alpha 剩余法和ISSTES 算法进行了改进，从而较好地反演出地表温度和比辐射率。
以Aster_TES算法为基础分别构建min与MMD、MMR 和VAR的经验关系，进行温度与比辐射率分离。对于TASI 数据来讲，MMR 经验关系反演的温度精度最高，VAR 经验关系反演的比辐射率精度最高。
对alpha 剩余法引进了大气纠正项和维恩近似纠正项，分析了改进算法的敏感性和反演精度。虽然修正后的算法在理论上得到完善，但是和Aster_TES 算法相比，温度和比辐射率精度较低。
ISSTES 算法中，采用二阶差分作为代价函数反演的样本点的温度精度最高，但是温度图像的空间噪声大；ISSTES 算法只能获取比辐射率的相对谱形，要精确获取比辐射率的值还需要增加其他约束条件。
借鉴Aster_TES 和ISSTES 算法的优秀思想：保持比辐射率波谱形状和真实地表比辐射率光谱曲线平滑假设，提出了一种基于噪声分离的温度与比辐射率分离（NSTES）算法，该算法考虑了噪声去除以及其他经验约束使得反演的比辐射率精度有了明显提升，反演的温度图像的信噪比明显提高。
光谱分辨率是TES 算法精度的重要影响因素：随着光谱分辨率的提高，温度反演精度逐渐提高，其中当光谱分辨率小于0.172μm 时，精度随光谱分辨率提高而迅速增加。当光谱分辨率大于0.172μm 时，精度随光谱分辨率提高而增加缓慢。因此0.172μm 可作为未来热红外高光谱传感器光谱分辨率指标的重要理论参考。
应用示范中成功将温度图像用于日较差模型，下一步将深入研究把温度产品用于热惯量、蒸散模型以及城市能量平衡的研究中。
3. 尺度问题方面：
不同尺度转换方法对比分析表明，中心像元法不适合于尺度转换；点扩散函数法在保持图像的空间自相关性方面优于简单平均法和小波变换法，但点扩散函数法获取的图像的标准差明显低于简单平均法和小波变换法；小波变换将空间域变换到频域后再进一步运算，在尺度转换中每个移动窗口都存在边缘像元的处理，因此会增加一些误差。从运算效率上看，平均法的运算效率最高，最容易实现，因此实际应用中采用此法者居多；其次是点扩散函数法，效率最低的是小波变换法。
通过分析P1T 和P2T的差异可知，相同条件下城市温度的尺度效应大于农村，Aster_TES 算法的尺度效应小于NSTES 算法的尺度效应。随着尺度转换窗口的变大，温差的直方图分布也越来越离散，峰值越来越低，其中Aster_TES 算法的尺度效应随尺度变化最慢，其次是NSTES 算法。
P1 T 和P2T、P1和P2之间具有显著的相关性。根据回归方程对温度与比辐射率进行尺度校正结果表明，基于统计方法建立的回归方程能有效校正温度与比辐射率的尺度效应，但没有改变不同TES 算法之间的尺度效应相对大小。
选用局部方差和半方差进行最优尺度选择的结果是：城市热红外高光谱研究的最优空间分辨率为6.25m~8.75m。
【Abstract】
Airborne hyperspectral thermal infrared data spans a very broad requirement of application. Presently, serveral developed countries carried out the application study of airborne hyperspectral thermal infrared imaging. It is of core theoretical and practical problem to study the algorithm of temperature and emissivity separation algorithms. The main data resources include two aspects: TASI data and field measurement results.The research discusses serveral algorithms of temperature and emissivity separation, and their scaling effect. Main results and conclusions are summarized as follows:
1. Based on the characters of TASI data, the basic process flowing for TASI data is built, whose main steps are radiance calibration, atmospheric correction, geometric correction, image mosaic, temperature and emissivity separation, the application of temperature and emissivity products.
2. The study of temperature and emissivity separation After the study of the algorithm of Aster_TES, alpha derived emissivity method, and ISSTES, this paper improves alpha derived emissivity, and ISSTES method combination with the TASI data, and then retrieves the accuracy temperature and emissivity.
During the Aster_TES method, the new empirical relationships between the εmin and the statistical parameters of emissivity spectrum, such as MMD, MMR and VAR are created. For TASI data, the accuracy of temperature is best using empirical relationships between the εmin and MMR, and the accuracy of emissivity is best using empirical relationships between the εmin and VAR.
The improved derived emissivity method introduced to the corrections of atmospheric effects, and Wien approximation is perfected theoretically. After analysis of the sensitivity and retrieve accuracy of improved method, the result shows that accuracy of temperature and emissivity is worse than Aster_TES method.
For ISSTES method, the accuracy of temperature retrieved by the cost function of second order difference is best, but the noise of temperature image is more than the other cost functions; the ISSTES method can only get the relatively emissivity shape, so other constraint conditions must be used for the accuracy value of emissivity.
Using for reference of good idea of Aster_TES and ISSTES methods, sunch as keeping the emissivity spectral shape and the smoothness of emissivity spectrum, the author puts forward the new algorithm named as emissivity and temperature separation based on noise separation (NSTES). This method takes the noise removing and other empirical relationship into consideration, and makes a difference for the precision of emissivity and the SNR of temperature images.
Spectral resolution is main affect factor for the pericision of TES algorithms. With the improvement of spectral resolution, the pericision of retrieved temperature becomes more and more high. When spectral resolution is less than 0.172μm, precision of temperature increases fast with the improvement of spectral resolution; when spectral resolution is higher than 0.172μm, precision of temperature increases slow with the improvement of spectral resolution. So 0.172μm can be taken as theoretical reference for the design of hyperspectral thermal infrared sensors.
Lastly, the author makes the temperature images application for diurnal Temperature range model, next work will apply the temperature to the thermal inertia, Evapotranspiration and the urban energy banlance.
3. Scale Effect This dissertation discusses four scale transformation methods: center pixels method, simple average method, point spread function method (PSF), wavelet transform method (WTM).
Center pixels method is not fit for scale transformation; the spatial autocorrelation of images using PSF method is higher than that of images using simple average method and wavelet transform method; but the standard deviation of images using PSF method is lower than that of images using simple average method and wavelet transform method. Wavelet transformation method transforms the image from spatial domain to frequent domain, and causes to error because of processing the edge pixels for each moving windows.
The implementation efficiency of average method is best, next best is PSF method, and the wavelet transformation method is worse. After analysis of the difference between T P1 and T P2 , the scale effect of temperature for city is greater than that for country; the scale effect of Aster_TES algorithm is less than that of NSTES algorithm. With increase in scale transformation windows, the histogram of difference in temperature becomes more and more scattered, and the peak value becomes lower and lower. There are significant relationship between T P1 and T P2, εP1 and εP2 . After regression analysis and scale correction, the result shows that the regression equation on the base of statistical method can decrease the scale effect of temperature and emissivity, but cant change the relative size of TES algorithm. Local variance and semi-variance are fit for scale choice. The optimal spatial resolution for hyperspectral thermal infrared is 6.25m~8.75m.
