【作者】 王晓丽；

【导师】 周国民；

【作者基本信息】 中国农业科学院 ， 管理科学与工程， 2011， 硕士

【摘要】 本文主要研究了利用光谱数据诊断水稻胡麻斑病和水稻纹枯病叶片的理论和方法,通过对多种算法的比较分析研究,找出了最优的光谱预处理和分析算法,并建立了最优的病害严重度诊断模型和识别模型。为今后通过航空航天遥感平台大面积监测水稻胡麻斑病和水稻纹枯病提供了依据,也可以为水稻其他病害的遥感监测提供借鉴和参考。本文的研究对象是自然条件下发病的水稻叶片,主要流程如下:光谱数据采集阶段,比较了运用内置光源的反射探头和裸光纤测量水稻叶片光谱信息的优劣;通过比较不同宽度水稻叶片对光谱反射率的影响,发现宽度的变化对近红外区域水稻叶片光谱反射率影响较大;同时提出了水稻病害叶片光谱数据采集的注意事项,对进入预处理阶段的光谱进行了筛选,最后获取101个水稻病害叶片样本用于论文研究。光谱数据预处理阶段,针对光谱数据存在噪声和散射的问题,主要研究了S-G平滑、kernel平滑、导数算法、多元散射校正等预处理算法。结果表明,平滑点数和多项式阶数需根据实际进行调整;kernel平滑比S-G平滑算法更优,但是处理速度较慢;光谱信息的一阶导数较二阶导数对噪声敏感度低,且能得出重要的光谱参数;多元散射校正能很好的消除基线的平移和偏移。水稻病害光谱特征分析阶段,对水稻病害叶片光谱信息对比发现:在400～700nm范围内,随着胡麻斑病和纹枯病病害等级的增加反射率逐渐增高,纹枯病较胡麻斑病光谱反射率增高速度迅速;在700~1300 nm近红外区域,水稻胡麻斑病和纹枯病叶片反射率随病害等级的增加而逐渐降低;在1900nm~2000nm范围内,水稻纹枯病叶反射率随等级增加逐渐增高,而水稻胡麻斑病叶片反射率随等级增加而降低;而其他波段无明显规律。特征提取阶段,针对近红外光谱的数据量大、波段众多的问题,先根据严重度与反射率的相关性数据,选取了水稻胡麻斑和纹枯病的敏感波段,再运用主成分分析的算法,选取累积贡献率达到85%以上的前2个分量,最终选取990nm、1850nm、660nm、1921nm、1933nm这5个主要的特征波段用于建模;原始光谱反射率经一阶导数处理后,选取与严重度相关性最好的红边面积作为区分健康与生病叶片的重要参数。模型建立阶段,通过对得到的5个特征波段和红边面积参数分别建立水稻胡麻斑和纹枯病的严重度诊断模型,通过模型验证表明以下模型精度最高:(1)胡麻斑入选的组合反射率R1933nm- R990nm建立的模型y=11.4971x+5.0313, r=0.8912;(2)纹枯病入选的红边面积和组合反射率R660nm- R990nm建立的模型y=-11.1037x+3.5195,r=0.89502;y=6.2834x+2.8464,r=0.8920;运用逐步回归法和BP神经网络方法对水稻胡麻斑和纹枯病进行识别,其中60个样本用于建模,41个样本用于模型检验,结果表明,与逐步回归分析相比较,当对660nm、990nm、1933nm三个特征波段进行组合识别时,BP神经网络识别准确率比逐步回归分析的准确率高。更多还原

【Abstract】 This paper studies the spectral data of rice flax leaf spot and sheath blight disease; find the optimal spectral pre-processing and analysis algorithms; and established the optimal diagnostic model and severity identify model; to provide theoretic foundation for further monitoring rice flax leaf spot and sheath blight disease at large scale using airborne and airspace remote sensing as flats and reference for monitoring rice other disease. In this study the works were completed as follows:Spectral data acquisition stage, the influence of external environment, machine and sample to study fully considers for the spectral reflectance; use tools of blade grippers and naked fiber optic to measure rice leaf spectra data, and find which tool have the better result; find the change on the width of near infrared spectral reflectance have a big affect on rice leaf; also puts forward the matters needing attention when collect rice disease spectrum data, choose spectrum data for pre-processing stage.Spectral data pretreatment stage, for the problem of spectra data have noise and scattering, this pepar studied the S-G smoothing, kernel smoothing, derivative algorithm, multiplicative scatter correction of the data pre-preprocessing algorithm, and compared the results of preprocessing algorithms, and ultimately to find suitable preprocessing algorithm of rice diseases.Rice disease spectrum characteristic analysis stage, compared and analyzed rice leaves of normal and have difference disease which have the same width and variety, we found that: In the range of 400 ~ 700nm, with the level of flax leaf spot disease and the sheath blight disease was gradually increased to increase reflectivity, the increased speed are more rapid than flax spot; In the range of 700 ~ 1300 nm near-infrared region, with the level of flax spot disease and the sheath blight disease increase leaf reflectance gradually decreased; In the range of 1900nm ~ 2000nm, with the level of the sheath blight disease increase the leaf reflectance gradually increase, while the other band no law.Feature extraction stage, for the problems of near infrared spectrum have large amount of data and band numerous, according to the correlation of severity and reflectivity data, select rice sensitive band of the spot and sheath blight, then using principal component analysis algorithm select two component which accumulated more than 85 percent contribution; finally select five main features bands used for modeling: 990nm, 1850nm, 660nm,1921nm, 1933nm; after the original spectral reflectance process with the first derivative, select the best correlation with the severity of red edge area as a health and sick leaves important parameter.Model establishment stage, rice spot and sheath blight disease severity diagnosis model were established through 5 obtained the characteristic bands and the red edge parameters. Through the model validation showed that the following model has the highest: (1) rice spot, R1933nm- R990nm, y=11.4971x+5.0313, r=0.8912;(2) sheath blight, red edge area parameters and R660nm-R990nm, y=-11.1037x+3.5195,r=0.89502;y=6.2834x+2.8464, r=0.8920;use stepwise regression method and the BP neural network method to Identify the rice spot and sheath blight disease, including 60 samples used for modeling, 41 samples used to model test, the results show that compared to stepwise regression analysis, when 660nm, 990nm, 1933nm three features bands combined, the BP neural network identification accuracy than stepwise regression analysis.更多还原