【作者】 杨燕；

【导师】 何勇；

【作者基本信息】 浙江大学 ， 农业电气化与自动化， 2012， 博士

【摘要】 精细农业技术体系包括数据获取与数据采集、数据分析与可视化表达、决策分析与制定和精细农田作业的控制实施等主要组成部分,其中植物生长信息的快速获取与生长状况快速诊断,对提高农作物的产量和质量具有重要意义。传统的水稻稻瘟病病害检测主要基于分子生物学检测方法,不利于在线实时水稻稻瘟病病害识别。本研究论文主要围绕植物生长信息检测与解析中的病害胁迫信息快速获取和识别关键技术展开,利用光学成像传感装置快速无损获取水稻稻瘟病早期病害信息,通过研究不同的病害检测预报模型,及时监控水稻稻瘟病发生发展状况,构建稻瘟病预测预报系统,为变量施药决策系统的实施提供决策支持。本论文的主要研究内容和结论如下：(1)提出了针对稻瘟病病害检测的高光谱特征提取方法,实现了高维光谱数据压缩和特征提取,建立了基于光谱特征的稻瘟病病害识别模型,实现了稻瘟病的精确、无损检测。系统地研究了高斯函数拟合光谱特征提取、植被指数光谱特征提取和小波近似系数光谱特征提取,建立了基于高斯拟合参数、植被指数和小波近似系数的水稻稻瘟病病害分类判别模型。研究表明,3种光谱特征提取方法均可有效提取水稻稻瘟病病害特征光谱信息,其中分析得到的水稻稻瘟病优化识别模型为基于1阶导数光谱高斯拟合参数(峰高、峰宽、峰面积)LDA病害判别模型,该模型在校正集和预测集分类准确率分别为100%和96%。(2)应用数字图像处理技术,实现了水稻稻瘟病高光谱图像特征信息获取,建立了基于高光谱图像统计信息的水稻稻瘟病分类判别模型。系统研究了主成分图像特征提取,概率统计滤波图像特征提取和二阶概率统计滤波图像特征提取方法,得到了基于图像特征提取的优化模型——基于主成分图像统计信息的逐步线性判别模型,该模型在校正集判别准确率为98.3%,预测集判别准确率为97.5%。(3)应用高光谱和高光谱图像技术,提取了对水稻稻瘟病识别敏感的特征波长,建立了基于特征波长下光谱和图像信息的水稻稻瘟病害识别模型,实现基于图像信息稻瘟病病害识别模型的优化。研究应用主成分分析-载荷系数法得到了基于图像特征波长的病害识别优化模型为基于特征波段长(419nm,502nm,569nm,659nm,675nm,699nm,742nm)图像信息构建PCA-LDA分类模型,校正集判别准确率为92.7%,预测集判别准确率为92.5%。(4)首次建立了水稻冠层光谱信息与抗氧化酶活性关系预测模型,实现稻瘟病可见症状显症之前的早期病害识别。应用光谱技术,利用偏最小二乘回归法建立基于冠层光谱信息的抗氧化物酶(POD、SOD、CAT)酶值活性预测模型。基于可见-近红波段(400-1100nm)冠层光谱漫反射信息对POD预测相关系数,校正集为97.57%,预测集为90.79%；SOD预测相关系数,校正集为96.82%,预测集为86.65%；CAT预测相关系数,校正集为85.98%,预测集为66.63%。研究首次建立了基于特征波长光谱信息抗氧化物酶(POD、SOD、CAT)活性预测模型,实现了抗氧化酶活性预测模型的简化。基于特征波长(491nm,545nm,676nm,707nm,741nm)的漫反射值对POD预测相关系数,校正集为83.35%,预测集为75.19%；基于特征波长(526nm,550nm,672nm,697nm,738nm,747nm)的漫反射值对SOD预测相关系数,校正集为69.45%,预测集为54.88%；基于特征波长(491nm,503nm,544nm,673nm,709nm,744nm)漫反射值对CAT的预测相关系数,校正集为66.09%,预测集为46.91%。上述研究成果表明基于高光谱图像技术可以实现水稻稻瘟病病害的快速、无损检测,研究为水稻稻瘟病病害的快速检测仪器和传感器开发奠定了理论基础,具有广阔的应用前景。更多还原

【Abstract】 The system of precision agriculture technology consists of 4 major components, which includes:data acquisition & data collection, data analysis and visualization, decision analysis and decision making, the implementation of agricultural job control. In these key technology areas, rapid acquiring and diagnoses of plant growth information is of great importance to the improvement of crop yield and quality.Traditional detection of rice blast disease is mainly based on molecular biology methods, which militate against on-line and real time detection. The main focus of this paper is on the key technology of acquiring and identifying the diseases stress in the plant growth stage. The early rice blast disease information was detected by optical imaging sensing device rapidly and non-destructively, and by the study of different prediction models, the incidence of rice blast development was monitored timely. The blast forecasting system constructed will provide the decision support for the variable spraying system.The main research contents of this paper and conclusions are as follows:(1) The high spectral feature extraction method was proposed for rice blast detection, the proposed method realized the high-dimensional spectral data compression and feature extraction. The Blast disease recognition model was constucted based on spectral feature and realized the identification of rice blast nondestructively and pricisely. The gaussian function fitting spectrum feature extraction, vegetation index spectrum feature extraction and wavelet approximate coefficient spectrum feature extraction are evaluated systematically and the classification discrimination model of rice blast based on gaussian fitting parameters, vegetation index and Wavelet approximate coefficient was constructed respectively.Research shows that three spectrum feature extraction method can effectively extract characteristics spectrum information for rice blast detection. The Optimization identification model based on spectral feature obtained by analysis,The classification accuracy of LDA identification model based on gaussian fitting parameters (peak high, peak width and peak area)extracted from 1 st derivative spectra is 100% on correction set and 96% on pridiction set, respectively.(2) The rice blast hyperspectral image feature informationwas acquired based on digital image processing techniques.The Blast disease recognition model was constucted based on image feature. Principal component image feature extraction, probability statistics filter image feature extraction and second order probability statistics filter image feature extraction are evaluated systematically.The optimization model based on feature extraction was obtained. The classification accuracy of stepwised LDA identification model based on principal component image statistical information (image mean and variance) is 98.3% on correction set and 97.5% on pridiction set, respectively.(3) The feature wavelengths to rice blast recognition sensitive was extracted based on spectrum and high hyperspectral imaging technology. The Blast disease recognition model was constucted based on spectrum and image information of feature wavelengths. The blast disease identification model based on image information was optimized. The optimimation identification of disease based on image feature wavelengths was acquired using principal component analytical-load coefficient method.The classification performance of PCA-LDA discriminant analysis model based on feature wavelengths (419nm,502nm,569nm,659nm,675nm,699nm, 742nm) image information, the classification accuracy is 92.7% on correction set and 92.5% on pridiction set, respectively.(4) The canopy spectral information and antioxidant enzymes activity relationship forecast model was set up for the first time.The early disease recognition was realized before the visible symptoms of the disease. The antioxidant enzyme (POD SOD CAT) enzyme activity value prediction model was constructed based on canopy spectral information using PLS regression method. Based on visible-infared band (400-1100 nm) canopy spectral diffuse information, the POD prediction correlation coefficient is 97.5% on calibration set and 90.79% on pridiction set; the SOD prediction correlation coefficient is 96.82% on calibration set and 86.65% on pridiction set. the CAT prediction correlation coefficient is 85.98% on calibration set and 66.63% on pridiction set. The antioxidants enzyme (POD SOD CAT) activity forecasting model based on was set up based feature wavelengths spectrum information for the first time,and simplified the forecasting model. Based on diffuse value of feature wavelengths (491 nm,545 nm,676nm,707 nm,741nm), the POD prediction correlation coefficient is 83.35% on calibration set and 75.19% on pridiction set. Based on diffuse value of feature wavelengths (526nm,550nm,672nm, 697nm,738nm,747nm), the SOD prediction correlation coefficient is 69.45% on calibration set and 54.88% on pridiction set. Based on diffuse value of feature wavelengths (491 nm,503nm,544nm,673nm,709nm,744nm), the SOD prediction correlation coefficient is 66.09% on calibration set and 46.91% on pridiction set.The research results indicated that the rice blast disease can identified quickly and non-destructively based on hyper-spectral image technology. For developing the rice blast disease fast detection instruments and sensors, the research laid a theoretical foundation and has the broad application prospect.更多还原