【作者】 邵咏妮；

【导师】 何勇； 赵春江；

【作者基本信息】 浙江大学 ， 生物系统工程， 2010， 博士

【摘要】 精细农业是21世纪全球农业发展的必然趋势,是实现农业低耗、高效、优质与安全的重要途径。它的技术核心是农田信息的获取、信息的管理与决策及变量作业三个部分。其中如何快速实时地获取土壤和作物的状态信息,是开展精细农业最为基本和关键的问题,也是精细农业研究的一个热点和难点。基于国内外在农作物方面的研究成果,本论文以水稻为对象进行了详细深入的研究。通过光谱技术与多光谱成像技术的有机结合实现了对水稻生长、生理信息的采集,并运用化学计量学方法和数据挖掘技术对采集数据进行分析,实现了水稻品质信息、养分需求信息、病虫害信息的全方位检测和诊断,为水稻等田间作物的生长、生理信息无损检测仪器的开发奠定了较为扎实的理论基础。本论文通过二次正交回归设计和设置不同氮肥梯度的方法进行了水稻不同施肥状态的田间试验,采用可见-近红外光谱技术研究了水稻冠层及叶片SPAD值和氮素含量与水稻冠层和叶片光谱反射特性间的关系；应用光谱技术和数据挖掘技术建立了水稻叶片叶绿素含量及微量元素含量(铁、锌)的数学模型；通过分析稻瘟病病变叶片的光谱特征信息,进行了水稻稻瘟病等级判别等的研究；研究了水稻植株及叶片多光谱图像与SPAD值和氮素含量间的关系；通过大量试验分析证实了采用水稻冠层光谱信息反演土壤养分(氮、磷、钾)信息的可行性。此外,还探讨了辐照处理对稻谷的光谱反射特性的影响,并结合中红外光谱技术对辐照谷物的内部成分含量(直链淀粉和蛋白质)作了深入研究,对水稻品质的无损检测提供了依据。本论文的主要研究成果和结论如下：1)首次采用化学计量学方法结合特征波段选取方法,提取能够反演水稻冠层及叶片SPAD值的敏感波段,为仪器的开发奠定基础。对于水稻冠层SPAD值的预测模型,非线性的偏最小二乘支持向量机(PLS-LS-SVM)模型具有较高的预测精度。对水稻叶片SPAD值的预测模型,基于全波段的预测效果最好。文章将预处理方法与特征波段提取及数据压缩技术结合,建立水稻叶片叶绿素含量的预测模型,并提取敏感波段反演叶绿素含量信息。其中直接信号校正算法(DOSC)结合连续投影算法(SPA)的最优波长选择方法要优于多元散射校正(MSC)结合SPA的波长选择方法。2)将基于独立组分分析的特征波段提取方法应用于水稻冠层及叶片氮素含量与光谱反射特性关系的研究中,建立了不同波段水稻冠层氮素含量的偏最小二乘(PLS)模型,并通过试验证实了基于全波段的水稻冠层氮素预测模型的效果最好。基于ICA-LS-SVM模型的水稻叶片氮含量研究,获取了叶片氮含量的敏感波段,为仪器开发提供了理论依据。采用PLS建立了全波段、多波段和多波长的水稻稻瘟病病变叶片的鉴别模型,结果显示：采用全波段建模,模型的鉴别率最高,高于采用特征波段和特征波长所建的模型。试验还通过对病变叶片建立ICA-LS-SVM模型,达到了86.7%的鉴别率。3)首次将光谱技术结合数据挖掘技术应用于水稻叶片微量元素含量(铁、锌)的研究。采用偏最小二乘(PLS)建立了全波段、多波段和多波长的水稻叶片微量元素铁和锌的预测模型。对于微量元素铁,基于全波段的模型预测精度要高于采用多波段的模型,高于采用多波长的模型。对于微量元素锌,基于多波长的模型预测精度高于全波段的模型,高于采用多波段的模型。建立了基于ICA-LS-SVM模型的水稻微量元素铁和锌的预测模型,结果显示,采用独立组分分析(ICA)结合非线性LS-SVM回归模型的预测精度高于PLS-LS-SVM模型,高于线性PLS模型。4)研究了利用冠层光谱特性评价土壤养分信息,包括氮、磷、钾信息的可行性。基于PLS的模型对土壤养分(氮、磷、钾)的预测精度拔节期高于分蘖期,其中氮含量的预测精度高于磷含量,钾的预测效果相对较差。非线性PLS-LS-SVM模型建立的分蘖期及拔节期土壤氮、磷、钾的预测结果要优于ICA-LS-SVM.PLS-BPNN及PLS的模型结果。基于植被指数对土壤氮素含量的研究,得到模型的反演效果好于ICA-LS-SVM模型。5)采用多光谱成像技术建立植被指数与水稻植株及叶片SPAD值和氮素含量间的关系模型。采用的植被指数,包括有归一化植被指数、绿波归一化植被指数和比值植被指数。其中对叶片SPAD值的预测,模型的相关系数为0.8756；植株拔节期SPAD值的预测结果好于分蘖期；植株分蘖期氮素含量的预测结果好于拔节期的预测精度。6)利用光谱技术建立稻谷年份的判别模型,并对经辐照处理后谷物的辐照剂量进行预测,同时建立辐照谷物内部成分(直链淀粉和蛋白质)的预测模型。基于独立组分分析(ICA)结合BP神经网络模型(ICA-BPNN)对不同年份谷物的鉴别率达到100%。采用LS-SVM模型建立谷物不同辐照剂量的预测模型,预测结果优于PLS模型。对辐照后谷物直链淀粉含量的预测,基于近红外光谱模型的预测结果优于中红外区域的模型,LS-SVM模型优于PLS模型的预测结果。对辐照后谷物蛋白质含量的预测,基于中红外光谱模型的预测结果优于近红外区域的模型,LS-SVM模型优于PLS模型的预测结果。更多还原

【Abstract】 Precision agriculture is the inevitable trend for the development of agricultural in the 21st century, and it is an important way for achieving low energy consumption, high efficiency, high quality and security. The key technology includes field information acquisition, information management and decision-making, and the feature of variable operating, which means precision agriculture can be processed relying on the existence of in-field variability. So far, how to quickly catch real-time status information of soil and crops growth information is one of the most critical issues.Based on the research on the application of spectroscopy in biosystem engineering field, this dissertion focuses on the spectral investigation on rice concerning the quality, nutrient and disease, etc. And validate the feasibility of designing non-destructive equipment for the field application for plant monitoring and diagnosis.This thesis designed an experimental plan regarding the methodology of quadratic orthogonal regression and the set-up of differential level of fertilizers, especially the nitrogen. The visible-near infrared spectroscopy was adopted to build the relationship between the reflectance characteristics of rice canopy or leaf and the SPAD values or nitrogen content of rice canopy or leaf, and the relationship between the reflectance characteristics of rice leaf and the chlorophyll content or trace element content of rice leaf. The spectral characteristic of infected rice leaf was also evaluated concerning the relation between the disease and shortage of nutrient. The thesis later conducted field experiments to prove the feasibility of using rice canopy spectral information to predict soil nutrients (nitrogen, phosphorus, kalium). The correlation between the SPAD value or nitrogen content with the multi-spectral images of rice plant and leaf was also studied. In addition, a preliminary study of the irradiation treatment on the rice reflectance characteristics was studied, and the internal components for irradiated grain (amylase and proteins) was predicted combined with the mid-infrared spectroscopy.The conclusions for this thesis are as follows:1) With the methodology of Chemometrics combined with the sensitive waveband acquisition, the thesis built the model for SPAD value prediction for canopy and leaf of rice. The prediction results showed that, with respect to SPAD values of rice canopy, the prediction accuracy for nonlinear PLS-LS-SVM model is higher, while for the prediction model for SPAD value of rice leaf, the model with the full wavelengths is better. Taking the advantage of preprocessing method and sensitive waveband acquisition together with the data compression, the thesis developed the model for predicting chlorophyll content. The optimal wavelength selection method combined DOSC with SPA is more accurate than MSC combined with SPA for chlorophyll content prediction.2) Based on the ICA, the relationship between the reflectance characteristics and the nitrogen content of rice canopy or leaf was investigated. The prediction model for nitrogen content of rice canopy was established, and the model with the full wavelengths has the highest prediction accuracy. The prediction model for the nitrogen content of rice leaf, the prediction accuracy based on ICA-LS-SVM model is higher and can be taked as preliminary reference for machine development. PLS models were established based on the full wavelengths, characteristic wavebands, and characteristic wavelengths to distinguish the infected rice leaves. The results showed that, the model with full wavelengths had the highest identification rate. The ICA-LS-SVM model built for identification of infected leaf can be as high as 86.7%.3) The thesis firstly took research on the trace elements investigation with technology of spectroscopy and data mining. The prediction model based on full wavelengths, characteristic wavebands and characteristic wavelengths were established by the PLS model. For the trace elements Fe, the prediction accuracy based on full wavelengths was higher than using characteristic wavebands, higher than model with characteristic wavelengths. For the trace elements Zn, the model based on characteristic wavelengths was higher than full wavelengths model, and higher than using characteristic wavebands. The prediction model for trace elements Fe and Zn based on the ICA-LS-SVM models were established. It indicated that, the prediction accuracy using independent component analysis (ICA) combined with nonlinear LS-SVM regression model higher than PLS-LS-SVM model, higher than linear PLS model.4) The feasibility of using rice canopy spectral information to evaluate soil nutrients (nitrogen, phosphorus, kalium) was investigated. The prediction accuracy for soil nutrients (nitrogen, phosphorus, kalium) based on PLS model in rice booting stage was higher than in tillering stage. The prediction accuracy for nitrogen content is higher than phosphorus content and the results for kalium is relatively poor. The prediction results for soil nitrogen, phosphorus and kalium in rice tillering and booting stage showed, models based on nonlinear PLS-LS-SVM is better than ICA-LS-SVM, PLS-BPNN and PLS models. The prediction model for soil nitrogen content based on vegetation index is better than ICA-LS-SVM model.5) The relationship between the vegetation index and the SPAD values or nitrogen content of rice canopy or leaf was studied based on the multi-spectral imaging technique. The vegetation index used is the normalized difference vegetation index, green normalized vegetation index and ratio vegetation index. For the SPAD value of rice leaf, the correlation coefficient is reached 0.8756 for the prediction model. It showed that for SPAD value of the rice plant, the prediction results in rice booting stage was better than the tillering stage. For the nitrogen content of rice plant, the prediction accuracy in rice tillering stage was higher than the booting stage.6) The age discrimination model for rice was built and evaluated, and the radiation dose prediction for grain was also investigated. The prediction model for internal content (amylase and protein) of irradiated grain was studied. The differential rate of 100% was reached for the age prediction of grain based on the independent component analysis (ICA) combine with BP neural network model. The results for different irradiation doses prediction of grain showed that LS-SVM model was better than PLS model. For the amylase content prediction model of irradiated grain, models based on near infrared spectroscopy was better than the mid-infrared spectroscopy, and the prediction model with LS-SVM was superior to PLS model. For the protein content prediction model of irradiated grain, models based on mid-infrared spectroscopy was better than the near infrared spectroscopy, and the prediction model with LS-SVM was also superior to PLS model.更多还原