基于高光谱信息的柑橘叶绿素含量预测模型研究

发布时间：2018-11-13 16:34

【摘要】：柑橘是世界第一大果树作物,我国的柑橘产量和面积均已位居世界首位。柑橘产业已成为我国南方果农的主要经济来源。通过对柑橘叶绿素含量的分析能准确掌握果树的光合能力、营养状况和生长态势,为果园管理提供科学指导。测定叶绿素含量的传统方法是分光光度法,利用化学试剂萃取叶片中的叶绿素,依据不同波长下叶绿素吸光度不同的原理计算得到叶绿素含量。这种检测手段耗时长,具有破坏性,同时还依赖于检测者的操作技术,无法在数字化农业中推广。随着高光谱技术和遥感技术的发展,基于光谱信息建立预测模型已成为作物估产和营养检测的新手段。这种方法依据的是物质固有的吸收、发射或散射光谱特性。与传统化学分析手段相比,具有无损、快捷、准确的优点,符合现代农业的发展要求。目前基于遥感技术的营养诊断主要应用在玉米、水稻等大田作物上,对柑橘等单株植物的研究相对较少。建立预测模型的常用方法有支持向量回归、BP神经网络、多元线性回归等,但涉及BP神经网络优化技术的研究却很少。本文以12年生枳橙[C sinensis(L.)Osbeck×P.trifoliate(L.)Raf.'Carrizo citrage']砧纽荷尔脐橙(C sinensis(L.)Osbeck'Newhall navel orange')为研究对象,重点研究粒子群优化神经网络算法,并利用高光谱信息在叶片级别建立叶绿素含量预测模型。旨在提高叶绿素含量预测精度,同时为高光谱遥感在柑橘长势监测中的应用提供理论依据和技术支持。本文的主要研究内容可归纳为以下两个方面:(1)提出了一种改进的粒子群优化算法,并用来优化BP神经网络。传统的BP算法在训练网络时容易陷入局部最优,为解决这一问题,研究者们提出了多种优化算法。研究表明,这些优化技术能有效提高BP神经网络的性能,已在多个领域得到了应用。本文分析了以往叶绿素含量检测研究中常用的建模方法,发现较少涉及BP神经网络的优化算法。为了进一步提高预测模型的准确性,本文将粒子群优化BP神经网络算法作为研究重点。针对粒子群算法中适应度信息未被充分利用的问题,在原有算法的基础上提出了一种改进的粒子群优化算法(FDPSOs),并用它替代原有粒子群算法对BP神经网络权值进行优化。为了验证算法的性能,本文选择四组分类数据集进行实验,并与BP及其他多种改进的PSOs-BP网络进行比较。实验结果表明,算法能有效优化BP神经网络,提高模型的学习和泛化性能。(2)建立基于光谱信息的柑橘叶绿素含量预测模型。本研究将采集的柑橘叶片室内光谱转换为一阶导数形式、二阶导数形式和log(1/r)形式。通过主成分分析法和连续投影法处理多种形式的光谱数据,分别得到降维后的特征向量和特征波长。选择多元线性回归、偏最小二乘法、支持向量回归、BP神经网络以及FDPSOs优化后的BP神经网络建立柑橘叶绿素含量预测模型。比较不同模型的预测结果得到如下结论:经FDPSOs优化后的BP神经网络在预测叶绿素含量时比单纯的BP神经网络准确性更高(R=0.8786,RMSE=0.1683);使用原始光谱数据和log(1/r)形式的光谱数据进行柑橘叶绿素含量预测比导数形式的数据更有效。
[Abstract]:Citrus is the first fruit tree in the world, and the yield and area of citrus in China are the first in the world. The citrus industry has become the main economic source of the fruit farmers in the south of China. Through the analysis of the content of the chlorophyll of the citrus, the photosynthetic capacity, the nutritional status and the growth of the fruit trees can be accurately controlled, and the scientific guidance is provided for the management of the orchards. The traditional method for measuring the content of chlorophyll is to determine the content of chlorophyll from the principle of the different chlorophyll absorbance at different wavelengths by the method of spectrophotometry and the extraction of the chlorophyll in the leaves by the chemical reagent. The detection method is time-consuming and destructive, and also depends on the operator's operation technology and cannot be popularized in the digital agriculture. With the development of high-spectrum technology and remote sensing technology, the establishment of the prediction model based on the spectral information has become a new means of crop estimation and nutrition detection. This method is based on the intrinsic absorption, emission or scattering spectrum characteristics of the substance. Compared with the traditional chemical analysis method, the method has the advantages of non-destructive, rapid and accurate, and accords with the development requirement of modern agriculture. At present, the nutrition diagnosis based on remote sensing technology is mainly applied to large field crops such as corn, rice and the like, and the research on the plant of the single plant such as the citrus is relatively small. The common methods to set up the prediction model are support vector regression, BP neural network, multiple linear regression, etc., but the research of BP neural network optimization technology is very low. This paper takes 12 years of raw orange[C sinensis (L.) Osbeck and P. trifoliate (L.) Raf. 'Carrizo citrage.'] Anvil's navel orange (C sinensis (L.) Osbeck 'Newhall namel) In order to study the object, the particle swarm optimization neural network (PSO) algorithm is focused on, and the chlorophyll content prediction model is established at the blade level by using the high spectral information. The aim of this paper is to improve the accuracy of the prediction of the content of chlorophyll, and to provide the theoretical basis and technical support for the application of high-spectral remote sensing in the monitoring of the long-term citrus. The main contents of this paper can be summarized as follows: (1) An improved particle swarm optimization algorithm is proposed and used to optimize the BP neural network. The traditional BP algorithm is easy to get into the local optimal when training the network, and to solve the problem, the researchers put forward a variety of optimization algorithms. The results show that these optimization techniques can effectively improve the performance of BP neural network and have been applied in many fields. In this paper, the commonly used modeling method in the research of chlorophyll content detection is analyzed, and the optimization algorithm of BP neural network is found to be less. In order to further improve the accuracy of the prediction model, the particle swarm optimization BP neural network algorithm is used as the research focus. In order to solve the problem that the fitness information is not fully utilized in the particle swarm optimization algorithm, an improved particle swarm optimization algorithm (FDPSOs) is proposed on the basis of the original algorithm, and the BP neural network weight value is optimized by using it instead of the original particle swarm optimization algorithm. In order to verify the performance of the algorithm, four groups of classified data sets are selected and compared with BP and other modified PSOs-BP networks. The experimental results show that the algorithm can effectively optimize the BP neural network and improve the learning and generalization performance of the model. and (2) establishing a citrus chlorophyll content prediction model based on the spectral information. In this study, the indoor spectrum of citrus leaves was converted into a first derivative form, a second derivative form and a log (1/ r) form. and processing the spectral data in various forms by the principal component analysis method and the continuous projection method, and respectively obtaining the characteristic vector and the characteristic wavelength after the dimension reduction. The prediction model of the chlorophyll content of citrus was established by using the BP neural network after the multiple linear regression, the partial least square method, the support vector regression, the BP neural network and the FDPSOs optimization. The results of the prediction of different models are as follows: the BP neural network optimized by FDPSOs is more accurate than the simple BP neural network in predicting the content of chlorophyll (R = 0.8786, RMSE = 0.1683); The use of the spectral data in the form of the original spectral data and the log (1/ r) is more effective than the data in the derivative form.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S666;TP18

【参考文献】