~植物叶片是植物体的重要组成部分，其中含有如叶绿素、类胡萝卜素、水分、木质素及纤维素等诸多的生化组分，蕴含了多种信息。对植被生化参数含量的合理估计，不管是对于农村的发展，还是对于当地生态系统的平衡，当地生态安全等方面均有着极其重要的意义。此外，机器学习算法可以很好的解释植物生物化学参数与光谱反射率之间隐含的、潜在的非线性函数关系，这使得机器学习算法更适用于与辐射传输模型相结合反演植被的叶片和冠层生化组分含量。
本文的研究内容与结论如下：
（1）在叶片尺度，为提高植被叶片生化参数的反演精度，本文利用扩展傅立叶幅度敏感性检验法对阔叶辐射传输模型(PROSPECT)进行敏感性分析，提取出了叶绿素、类胡萝卜素、水分和干物质四种生化组分对应的敏感波段。其中，叶绿素的提取波段为531nm-737nm、类胡萝卜素的提取波段为400nm-530nm、水分的提取波段为1312-2500nm、干物质的提取波段为400nm-2500nm。其次，本文在标准鲸鱼算法(Whale Optimization Algorithm, WOA)的基础上,通过引入改进的混沌序列，对种群进行初始化，产生分布均匀的个体。随后引进非线性收敛因子与权重对鲸鱼优化算法的参数进行改进，利用改进后的参数对个体位置进行更新，使算法更易遍历全局寻找最优解。然后，在一次迭代完成时对适应度值进行排序，选择出适应度高的精英个体，通过扰动概率对精英个体进行高斯扰动，增加算法跃出局部最优的概率，形成了改进的鲸鱼优化算法(Improved Whale Optimization Algorithm, IWOA)。最后将改进鲸鱼算法与常见的粒子群优化算法的反演精度进行比较。结果表明，改进鲸鱼算法反演的叶绿素、类胡萝卜素和等效水厚度的精度均高于粒子群优化算法反演的精度。具体来说，改进鲸鱼算法（粒子群优化算法）反演的叶绿素、类胡萝卜素和等效水厚度的R2分别为0.969（0.175）、0.917（0.013）、0.925（0.738）。改进鲸鱼算法反演（粒子群优化算法）的叶绿素、类胡萝卜素和等效水厚度的RMSE分别为6.899（18.022）、0.384（27.065）、0.013（0.004）。
（2）在冠层尺度，本文以冠层叶绿素含量为研究对象。为提高作物冠层叶绿素含量反演的精度，以廊坊市2017年冬小麦试验小区为基础，测量了研究区冬小麦冠层太阳辐射亮度和冠层辐射亮度数据及其对应的叶片叶绿素含量和叶面积指数LAI。通过对测量得到的辐射亮度数据进行处理，得到了研究区冠层的反射率数据，此外，利用3FLD荧光反演算法得到了冠层尺度日光诱导叶绿素荧光数据作为模型的数据源。运用分数阶微分法计算了0-2阶步长为0.1的分数阶光谱，通过相关性分析与连续投影算法相结合的高光谱特征提取方法，首先利用相关性分析这一方法对冠层叶绿素含量与高光谱特征参数之间进行分析，筛选出与冠层叶绿素含量极显著相关的前50个特征波段。其次，为降低各波段间互相关性，并考虑到神经网络模型的结构复杂度对反演精度的影响，对筛选出的50个特征波段进行连续投影算法处理，消除光谱矩阵中冗余的信息，最终提取出1.1阶739nm;1.6阶740nm;1.9阶791nm和766nm共计4个与冠层叶绿素含量关系紧密的波段作为模型的输入特征。此外，由于BP神经网络(Back Propagation Network，BP)存在收敛速度慢、易陷入局部极小值等问题。因此，本文使用差分进化灰狼优化算法(Differential evolution of Grey wolf optimization algorithm，DE-GWO)对BP神经网络的权值和阈值进行迭代优化。利用优化后的神经网络模型进行冠层叶绿素含量的预测。并比较了引入荧光数据对反演结果的影响。结果表明，运用差分进化灰狼算法优化的BP神经网络模型对加入荧光参数的数据集反演精度相较于未加入荧光参数的数据集有所提高。两个数据集反演的R2为0.925，RMSE分别为10.134和16.833。
 

    Plant leaves are an important part of the plant body, which contains many biochemical components such as chlorophyll, carotenoids, water, lignin and cellulose, which contain a variety of information. Reasonable estimation of the content of vegetation biochemical parameters is of great significance to the development of rural areas, to the balance of local ecosystems, and to local ecological security. In addition, machine learning algorithms can well explain the implicit and potentially nonlinear functional relationship between plant biochemical parameters and spectral reflectance, which makes machine learning algorithms more suitable for inversion of vegetation leaves in combination with radiative transfer models. and canopy biochemical components.
    The research contents and conclusions of this paper are as follows:
(1) At the leaf scale, in order to improve the inversion accuracy of leaf biochemical parameters of vegetation, the sensitivity analysis of broad-leaf radiative transfer model (PROSPECT) was carried out by using the extended Fourier amplitude sensitivity test, and the corresponding sensitive bands of chlorophyll, carotenoid, water and dry matter were extracted. Among them, the extraction bands of chlorophyll are 531nm-737nm, carotenoid is 400nm-530nm, water is 1312-2500nm, and dry matter is 400nm-2500nm. Secondly, on the basis of the standard Whale Optimization Algorithm (WOA), the improved chaotic sequence is introduced to initialize the population and produce uniformly distributed individuals. Then, nonlinear convergence factor and weight are introduced to improve the parameters of whale optimization algorithm, and the improved parameters are used to update individual positions, making the algorithm easier to traverse the whole world to find the optimal solution. Then, the fitness values are sorted at the completion of an iteration, and elite individuals with high fitness are selected. Gaussian perturbation is performed on elite individuals through perturbation probability to increase the probability of the algorithm jumping out of the local optimal. Improved Whale Optimization Algorithm (IWOA) is formed. Finally, the inversion accuracy of the improved whale algorithm is compared with that of the common particle swarm optimization algorithm. The results show that the accuracy of chlorophyll, carotenoid and equivalent water thickness inversion by improved whale algorithm is higher than that by particle swarm optimization algorithm. Specifically, the R2 of chlorophyll, carotenoid and equivalent water thickness retrieved by the improved whale algorithm (particle swarm optimization) were 0.969 (0.175), 0.917 (0.013) and 0.925 (0.738), respectively. The RMSE of chlorophyll, carotenoid and equivalent water thickness were 6.899 (18.022), 0.384 (27.065) and 0.013 (0.004), respectively. In general, compared with PSO standard particle swarm optimization algorithm, the improved whale optimization algorithm and band inversion model have higher prediction accuracy for biochemical parameters.
(2) At canopy scale, chlorophyll content in the canopy was taken as the research object. In order to improve the accuracy of crop canopy chlorophyll content inversion, the canopy solar radiation brightness and the corresponding leaf chlorophyll content and leaf area index (LAI) of winter wheat in the study area were measured based on the 2017 winter wheat experimental plot in Langfang city. By processing the measured radiation brightness data, the canopy reflectance data of the study area was obtained. In addition, the 3FLD fluorescence inversion algorithm was used to obtain the canopy scale sun-induced chlorophyll fluorescence data as the data source of the model. 2-0 is calculated by using the fractional order differentiation order fractional order spectrum of step length is 0.1, through correlation analysis and projection algorithm combined with continuous hyperspectral feature extraction method, first using the method of correlation analysis of the characteristics of canopy chlorophyll content and high spectral analysis between parameters, select canopy chlorophyll content significantly related top 50 characteristic bands. Secondly, in order to reduce the mutual correlation between the bands, and considering the impact of the structural complexity of the neural network model on the inversion accuracy, the 50 selected characteristic bands were processed by continuous projection algorithm to eliminate the redundant information in the spectral matrix, and finally the 1.1-order 739nm was extracted. 1.6 order 740 - nm; Four bands of 791nm and 766nm in order 1.9, which are closely related to canopy chlorophyll content, were used as input features of the model. In addition, the Back Propagation Network (BP) has some problems, such as slow convergence speed and easy to fall into local minimum. Therefore, Differential evolution of Grey Wolf optimization algorithm (DE-GWO) is used to optimize the weights and thresholds of BP neural network iteratively. The optimized neural network model was used to predict canopy chlorophyll content. The influence of introduced fluorescence data on inversion results is compared. The results show that the BP neural network model optimized by the differential evolution gray Wolf algorithm improves the inversion accuracy of the data set with fluorescence parameters compared with the data set without fluorescence parameters. The R2 and RMSE of the two datasets were 0.925 and 10.134 and 16.833, respectively.