【作者】 邢东兴；

【导师】 常庆瑞；

【作者基本信息】 西北农林科技大学 ， 土地资源与空间信息技术， 2009， 博士

【摘要】 以渭北旱塬上的几种代表果树为研究对象,分别基于其冠层或单鲜叶片尺度上的反射光谱数据,对果树的树种、病虫害情、冻害害情、挂果数量、三大营养元素（TN、TP、TK）含量、四种微量元素（Fe、Mn、Cu、Zn）含量等生物理化性状信息进行了探索性提取研究。此研究试图探寻利用地面光谱测试数据对果树理化性状信息进行有效监测、识别的方法与途径,并为我国今后利用机载、星载高（超）光谱传感器来对果树相关理化性状信息进行快速监测做一理论准备。主要研究内容及结论如下:1.基于处于开花期的7种果树（砂红桃、酥梨、金太阳杏以及秦冠、红嘎那、黄元帅、红富士苹果树）的冠层光谱反射率,探寻了利用地面光谱测试数据对处于开花期的多种果树树种进行有效识别的方法与途径。首先对比分析了各种果树的反射光谱特征;其后通过光谱数据重采样、植被指数求算等相关数据处理,对比分析了6种卫星传感器、4种植被指数分别对7种果树树种的识别效能,并对其识别效能进行了排序;最后在光谱反射率及其22种变式数据中,优选了最佳的数据形式,并在优化模型参数的基础上,建立了识别花期多种果树树种的3层BP神经网络模型。主要结论为:（1）6种卫星传感器识别果树树种的效能由强到弱的排列顺序为:MODIS、ETM+、QUICKBIRD、IKONOS、HRG、ASTER;（2）在4种植被指数中,RVI对果树树种的识别效能最强,其次是NDVI,SAVI与DVI的识别效能相对较弱;（3）用MODIS或ETM+传感器的近红外通道与蓝光通道上的反射率数据,求算的RVI与NDVI对果树树种的识别效能相对较强;（4）在R_λ及其22种变换数据中,波长间隔设为9 nm的d¹[㏒（1/R_λ）],是建立BP神经网络模型的首选数据形式;（5）利用波长间隔设为9 nm的d¹[㏒（1/R_λ）]这一数据形式,建立了识别果树树种的3层BP神经网络模型。2.基于处于果实成熟期的7种挂果果树（酥梨、满天红柿子、大红袍石榴以及秦冠、黄元帅、套袋红富士、未套袋红富士苹果）的冠层光谱反射率,探寻了利用地面光谱测试数据对处于果实成熟期的多种果树树种进行有效识别的方法与途径。首先对比分析了各种挂果果树的反射光谱特征;其后通过光谱数据重采样、植被指数求算等相关数据处理,对比分析了6种卫星传感器、4种植被指数分别对处于果实成熟期的多种果树树种的识别效能,并分别对其识别效能进行了排序;最后在光谱反射率及其22种变式数据中,优选了最佳的数据形式,并在优化模型参数的基础上,建立了识别处于果实成熟期的多种果树树种的3层BP神经网络模型。主要结论为:（1）6种卫星传感器识别果树树种的效能由强到弱的排列顺序为:MODIS、ASTER、ETM+、HRG、QUICKBIRD、IKONOS;（2）在4种植被指数中,RVI对果树树种的识别效能最强,其次是NDVI,SAVI与DVI的识别效能相对较弱;（3）用MODIS或ETM+传感器的近红外通道与红光通道上的反射率数据,求算的RVI与NDVI对果树树种的识别效能相对较强;（4）在R_λ及其22种变换数据中,波长间隔设为9 nm的d1[㏒（1/R_λ）],是建立BP神经网络模型的首选数据形式;（5）利用波长间隔设为9 nm的d1[㏒（1/R_λ）]这一数据形式,建立了识别果树树种的3层BP神经网络模型。3.基于3种果树（即酥梨树与秦冠、红富士苹果树,它们是研究区内三大主要果树树种）分别在5个不同时相（开花期、6月中旬、8月中旬、果实成熟期、果实采摘后）中的冠层光谱反射率,分析研究了利用高光谱数据对研究区内三大主要果树树种进行有效识别的最佳时相选取问题。首先对光谱数据进行了预处理,并对3种果树分别在各个时相中的反射光谱特征及其差异进行了对比分析;其后对处于各时相中的3种果树的光谱反射率,求算出其在每一波长处的标准差,并依据与各时相对应的反射率的标准差的大小及其波动变化情形,分析得出了利用高光谱数据对果树树种进行有效识别的最佳时相。主要结论为:（1）在5个时相中,3种果树反射率的标准差在各波长处相对较大且波动变化较大者要属开花期和果实成熟期;（2）在果实成熟期中,3种果树反射率的标准差的波动变化相对最大;（3）依据利用多、高光谱遥感影像数据对地物进行分类的先验知识可知,果实成熟期是利用高光谱数据对果树树种进行识别的最佳时相。4.基于3种果树（即酥梨、砂红桃、红富士苹果树）4个受冻级别的花朵的光谱反射率数据,对比分析了每种果树各级受冻花朵的光谱特征;利用一阶微分波谱在特征波段区间内的积分求值,分别对3种果树花朵的受冻级别建立了定量化评估模型。主要结论为:（1）每种果树各级受冻花朵的反射波谱均在360 nm附近出现了最低谷区,而在360～440 nm的波段区间均出现了坡度最大的陡坎,陡坎坡度的大小顺序为:未受冻级别>轻度受冻级别>中度受冻级别>重度受冻级别,并且陡坎的最大坡度均出现在400 nm附近;（2）当波长间隔同时设为9 nm时,酥梨、砂红桃、红富士苹果树的各级受冻花朵的一阶微分波谱分别在396±20,400±20,410±20 nm波段区间内的积分求值差异最大;（3）利用上述的3组积分值,分别对3种果树花朵的受冻级别建立了定量化评估模型。5.基于患有不同程度的红蜘蛛虫害或黄叶病害的红富士苹果树的冠层反射光谱,探寻了红富士苹果树在各级红蜘蛛虫害、黄叶病害胁迫下的反射光谱特征以及利用高光谱数据对果树病、虫害情进行有效测评的方法与途径。首先,分别对各级红蜘蛛虫害、黄叶病害胁迫下的果树的反射光谱特征进行了分析;其后,建立了定量化测评相应病、虫害程度的数学模型。主要结论为:（1）各级红蜘蛛虫害对应的反射波谱在630～695 nm波段,R_重度>R中度>R_轻度>R_正常,四者在684 nm处变异系数最大;在730～950 nm波段,R_重度<R_中度<R轻度<R_正常,四者在762 nm处变异系数最大。各级黄叶病害对应的反射波谱在515～716 nm波段,R_重度>R_中度>R_轻度>R_正常,四者在603 nm处变异系数最大;在740～950 nm波段,R_重度<R_中度<R_轻度<R_正常,四者在764 nm处变异系数最大;（2）果树反射波谱的“红边”位置随病、虫害情的加重而依次发生“蓝移”;（3）建立的害情测评模型对病、虫害情的测评准确率分别为98%、96%。6.以红富士苹果树为例,基于不同挂果量的果树的冠层反射光谱,探寻了不同挂果量的果树的反射波谱特征以及利用光谱数据对果树挂果量进行有效测评的方法与途径。在此子研究中,对比分析了不同挂果量的果树的反射光谱特征,并在优选光谱指数的基础上,建立了测评果实面积比（或果实总重量）的数学模型。主要结论为:（1）在560～673,760～950 nm两个波段内,5个级别挂果量分别对应的反射率的大小差异较大,且其反射率分别在634 ,760 nm处变异系数较大。在592,655,696 nm附近的波段内,5条反射波谱的一阶微分值的差异均较大;（2）反射波谱的“红边”位置随果树挂果量（或果实面积比）的增加而依次发生“蓝移”;（3）建立了优选的光谱指数与果实面积比（或果实总重量）之间的数量关系模型。7.以红富士苹果树为例,利用其单鲜叶片的光谱反射率以及叶片中的TN、TP、TK含量数据,探试了光谱分析手段估测果树N、P、K营养元素含量的精度与应用潜力。首先对果树叶片中的TN、TP、TK含量与叶片的光谱反射率（R_λ）及其多种变式数据（1/R_λ、lg（1/R_λ）、d1R_λ、d2R_λ、d1[lg（1/R_λ）]、d2[lg（1/R_λ）]、lg（1/BNC）、?′（R_λ）、Dn）之间的相关性强弱分别进行了分析对比,找出了与每种元素含量相关性最强的光谱数据形式;其后采用逐步回归法对每种元素含量和与其相关性最强的光谱数据形式进行了回归分析,得到了入选波长;最后利用入选的波长,进行了基于最小误差平方和的偏最小二乘回归建模。主要结论为:（1）叶片TN、TK含量与波长间隔设为5 nm的d1R_λ的相关性最强,TP含量与波长间隔设为9 nm的d1[lg（1/R_λ）]的相关性最强;（2）利用入选的特征波长,建立的估测模型均具有较好的线性趋势,R2均在0.8以上;（3）光谱分析手段对果树N、P、K营养元素含量的估测精度较高,具有一定的应用潜力。8.以红富士苹果树为例,利用其单鲜叶片的光谱反射率和叶片中的Fe、Mn、Cu、Zn 4种元素含量数据,探试了光谱分析手段估测果树微量元素含量的精度与应用潜力。首先对果树叶片中的Fe、Mn、Cu、Zn含量与叶片的光谱反射率（R_λ）及其多种变式数据（1/R_λ、lg（1/R_λ）、d1R_λ、d2R_λ、d1[lg（1/R_λ）]、d2[lg（1/R_λ）]、lg（1/BNC）、?′（R_λ）、Dn）之间的相关性强弱分别进行了分析对比,找出了与每种元素含量相关性最强的光谱数据形式;其后采用逐步回归法对每种元素含量和与其相关性最强的光谱数据形式进行了回归分析,得到了入选波长;最后利用入选的波长,进行了基于最小误差平方和的偏最小二乘回归建模。主要结论为: （1）叶片Fe、Mn、Cu、Zn含量都与R_λ的相关性较弱,而分别与四点差分的一阶微分光谱?′（R_λ）、波长间隔设为17 nm的d1R_λ、波长间隔设为25 nm的d1R_λ、波长间隔设为15 nm的d1R_λ的相关性最强;（2）用入选的波长建立的估测模型均具有较好的线性趋势,R2均在0.8以上;（3）光谱分析手段估测果树叶片Fe、Mn、Cu、Zn元素含量的精度较高,具有一定的应用潜力。更多还原

【Abstract】 The paper takes fruit trees as the research object, bases on reflectance spectroscopy of canopy or single leaf of fruit trees, exploratory research on Extraction of the information of Physical and chemical properties of fruit trees（the species of Fruit trees; the characteristics of the spectral reflectance of fuji apple trees which are intimidated by disease or insect pest; the characteristics of the spectral reflectance of fruit trees blossoms suffered frost with different degrees; the relation between the fruit quantity and the spectral reflectance of fuji apple tree; the three major nutrients〈TN, TP, TK〉content, 4 kinds of trace elements〈Fe, Mn, Cu, Zn〉content）,the study attempts to explore the effective ways and means to monitor or identify the information of physical and chemical properties of fruit trees.The main research content and conclusions are as follows:1. Using the spectral reflectance data （R_λ） of canopies, the paper identifis seven species of fruit trees during flowering period. Firstly, it compares the identification capacity of six kinds of satellite sensors and four kinds of vegetation index on the basis of resampling the spectral data with Six kinds of pre-defined filter function and calculating vegetation index. Then, it structures a BP neural network model for identifying seven species of fruit trees on the basis of choosing the best transformation of R_λand optimizing the model parameters.The main conclusions are:（1） The order of the identification capacity of Six kinds of satellite sensors from power to weak are : MODIS、ETM+、QUICKBIRD、IKONOS、HRG、ASTER; （2） Among four kinds of vegetation index, the identification capacity of RVI is the most powerful, next is NDVI, the identification capacity of SAVI or DVI is relatively weak; （3） The identification capacity of RVI and NDVI that are calculated with the reflectances of near-infrared and blue channels of ETM + or MODIS sensor are relatively powerful;（4） Among R_λand it`s 22 kinds of transformation data, d1[㏒（1/R_λ）]（ derivative gap is set 9 nm） is the best transformation for structuring BP neural network model.2. Using the spectral reflectance data （R_λ） of canopies, the paper identifies seven species of fruit trees bearing fruit in the fruit mature period. Firstly, it compares the fruit tree species identification capability of six kinds of satellite sensors and four kinds of vegetation index through re-sampling the spectral data with six kinds of pre-defined filter function and the related data processing of calculating vegetation indexes. Then, it structures a BP neural network model for identifying seven species of fruit trees on the basis of choosing the best transformation of R_λand optimizing the model parameters.The main conclusions are: （1） the order of the identification capability of six kinds of satellite sensors from strong to weak is : MODIS, ASTER, ETM +, HRG, QUICKBIRD, IKONOS; （2） among four kinds of vegetation indexes, the identification capability of RVI is the most powerful, the next is NDVI, the identification capability of SAVI or DVI is relatively weak; （3） The identification capability of RVI and NDVI those are calculated with the reflectance of near-infrared and red channels of ETM + or MODIS sensor are relatively powerful;（4） Among R_λand it’s 22 kinds of transformation data, d1[㏒（1/R_λ）]（ derivative gap is set 9 nm） is the best transformation for structuring BP neural network model; （5） The paper structures a 3-layer BP neural network model for identifying seven species of fruit trees using the best transformation of R_λwhich is d1[㏒（1/R_λ）]（ derivative gap is set 9 nm）.3. Basing on the spectral reflectance of three species of fruit trees which are in five different period of time, the study Analyses the problem of the best period of time to identify the species of fruit trees, the main conclusions are: the period of autumn is the best period of time to identify the species of fruit trees.4. The study attempts to explore the characteristics of the spectral reflectance of fruit trees blossoms suffered frost with different degrees, and try to evaluate quantitatively the levels of fruit trees blossoms suffered frost by spectra data . Firstly, we pretreat the spectra reflectance data of the blossoms of three species of fruit trees suffered frost with four levels , and analyze the spectra characteristics of the frosted fruit trees blossoms .Subsequently, we transform the spectra data by the first derivative with nine kinds of different wavelength intervals, and find out the special wavelength and Special wavelength range, and Select three groups of specific derivative Spectra in the transformation results. Finally, we build the quantitative assessment models for the frosted blossoms of the three Species of fruit trees using the integral values which are calculated by the selected three groups of specific differential Spectra within the corresponding Special wavelength range respectively.The main conclusions are:（1） the reflectance spectroscopy of the blossoms suffered frost at each level of each species of the fruit trees emerges the lowest Valley area near 360 nm, and emerges the Scarp with the largest slope within the Wavelength range from 360 nm to 440 nm, the order of the slopes is: no suffered frost >suffered frost lightly >suffered frost moderate > suffered frost severe, the largest slopes of the four Scarps are all near 400 nm;（2） when the wavelength interval is set at 9 nm, the integral values that are calculated by the derivative spectra of the frosted blossoms at all levels of Crisp Pear, Shahong Peach, Fuji apple trees within the Wavelength range 396±20 nm, 400±20 nm, 410±20 nm are respectively the largest; （3） we establish respectively quantitative evaluation models based on the three groups of integral values which mentioned in （2）.5. Yellow Leaves Disease and Red Mite Insect Pest of FuJi Apple Tree were used as Samples, the research attempts to explore the characteristics of the spectral reflectance of fuji apple trees which are intimidated by disease or insect pest, and try to evaluate quantitatively the degrees of disease or insect pest stress by spectra data.The main conclusions are: （1） The Spectral reflectance of each level of red mite insect pest within the wavelength range from 630 to 695 nm, R_severe>R_moderate>R_lightly>R_normal,the largest Coefficients of variation of the four Reflectances is at 684 nm, within the wavelength range from 730 to 950 nm, Rsevere < R_moderate < R_lightly <R_normal,the largest Coefficients of variation of the four Reflectances is at 762 nm.the Spectral Reflectance of each level of yellow leaves disease within the wavelength range from 515 to 716 nm, R_severeR_moderate>R_lightly>R_normal,the largest Coefficients of variation of the four Reflectances is at 603 nm,within the wavelength range from 740 to 950 nm, Rsevere < R_moderate < R_lightly <R_normal,the largest Coefficients of variation of the four Reflectances is at 764 nm. （2） The red border position of the Spectral Reflectance of fruit trees moves to shortwave with the level of disease or insect pest increasing.（3）The models that are used to evaluate quantitatively the degrees of disease or insect pest stress have higher precision .6. Red Fuji Apple Tree was used as object of the study, based on the spectra reflectance of apple tree which hanging on the quantity of fruit were not equal, The research attempts to explore the relation between the fruit quantity and the spectral reflectance of fuji apple tree, and try to evaluate quantitatively the fruit quantity with spectra data.The main conclusions are: （1） within the wavelength range from 560 to 673 nm, or from 760 to 950 nm, the difference of R_λof five levels of fruit quantity are large, the coefficient of variation of R_λof five levels of fruit quantity is larger when wavelength is 634 or 760 nm. the wavelength range nearby 592 or 655 or 696 nm, the difference of the first derivative spectral reflectance of five levels of fruit quantity is larger. （2） The red border position of the Spectral Reflectance of fruit trees moves to shortwave with the fruit quantity increasing.（3） the quantitive relation model between the optimal spectra index and area ratio of fruit is constructed.7.Red Fuji Apple Tree was used as object of study, the spectral reflectance （R_λ） of fresh leaves of fruit trees and the leaves total nitrogen （TN）, total phosphorus （TP）, total potassium （TK） contents were measured, and analyze the statistical correlation between the each element content and R_λas well as its several transformations （1/R_λ、lg（1/R_λ）、d¹R_λ、d²R_λ、d¹[lg（1/R_λ）]、 d²[lg（1/R_λ）]、lg（1/BNC）、f′（R_λ）、Dn） within the wavelength range from 400 nm to 900 nm by factor analysis method, and discover the spectral reflectance variant of the highest correlation coefficient . Subsequently, we carry on the regression analysis of each element content and the corresponding spectral reflectance variant of the highest correlation coefficient by stepwise regression method, and choose the eigenvalue wavelengths with which to carry on partial least squares regression modeling based on the least square error. The research is expected to evaluate the possibility and application potential of the method of spectral analysis on predicting the nutritional element of fruit trees.The results show : （1） The correlation coefficientis the highest between the leaves TN or TK content and the first derivative of spectral reflectance with derivative gap = 5 nm; the correlation coefficient is the highest between the leaves TP content and d¹[lg（1/R_λ）] with derivative gap = 9 nm; （2） The regression models that are established using the wavelengths selected by stepwise regression method when Sig（Tt′s a Probability numerical Higher than F detection value） is set to 0.02 and Ramoval is set to 0.03 have the better linear trend , and R2 values are higher than 0.8; （3） The method of spectral analysis have some applications potential to predict TN、TP、TK element of fruit trees .8. Red Fuji Apple Tree was used as object of study, the spectral reflectance （R_λ） of fresh leaves of fruit trees and the Fe,Mn,Cu,Zn element contents in the leaves were measured, and analyze the statistical correlation between the each element content and R_λas well as its several transformations （1/R_λ、lg（1/R_λ）、d¹R_λ、d²R_λ、d¹[lg（1/R_λ）]、d²[lg（1/R_λ）]、lg（1/BNC）、f′（R_λ）、Dn）within the wavelength range from 400 nm to 900 nm by factor analysis method, and respectively find out the spectral reflectance Variant, which has the highest absolute value of the correlation coefficient with each of the element’s content. Subsequently, we carry on the regression analysis to each element content and the corresponding spectral reflectance variant of the highest Correlation coefficient through the stepwise regression method, and choose the eigenvalue wavelengths with which to carry on partial least squares regression modeling based on the least error square sum. The research is expected to evaluate the possibility and application potential of the method of spectral analysis on predicting trace elements of fruit trees.The results show: （1） The correlation between the Fe, Mn, Cu, Zn elements content with R_λis rather weak and the correlation coefficient is the highest between the leaves Fe content and the first derivative of 4 points Difference of spectral reflectance f′（R_λ）, or between the leaves Mn content and the first derivative of spectral reflectance with derivative gap = 17 nm d¹R_λ, or between the leaves Cu content and the first derivative of spectral reflectance with derivative gap = 25 nm d¹R_λ, or between the leaves Zn content and the first derivative of spectral reflectance with derivative gap = 15 nm d¹R_λ; （2） The result also show that the regression models that are established using the wavelengths selected when Sig（it′s a Probability numerical Higher than F detection value） is set to 0.01 and Ramoval is set to 0.02 have the better linear trend, and R2 values are higher than 0.8; （3） The method of spectral analysis have some applications potential to predict trace elements of fruit trees .更多还原