【作者】 樊景超；

【导师】 周国民；

【作者基本信息】 中国农业科学院 ， 作物信息科学， 2011， 博士

【摘要】 苹果果实采后病害可分为生理性病害和细菌性病理两大类:生理性病害已有很好的前人研究基础,通过气调、温度调整及其他方法可以有效控制;细菌性病害的无损检测目前除了有少量基于电学参数特性等方法外,相关研究很少,利用近红外光谱进行无损检测还鲜有报道。因此利用近红外光谱的快速、绿色、无损等特性建立苹果病害的检测模型,具有重要理论意义和实用价值。本文以苹果果实为研究对象,采后病害识别(轮纹病和炭疽病)为目标、近红外光谱检测为研究手段,在对前人对苹果病害无损检测研究的基础上,分别从近红外光谱信息采集的影响因素、样品选择与模型参数和病害分类识别模型等方面开展研究,并在此基础上研制了基于近红外光谱的苹果病害识别软件系统。本文主要研究成果如下:(1)苹果光谱信息采集影响因素研究。对同一样品、同一位置连续9个小时共53次光谱采集证明,ASD公司的Field Spec 3型光谱仪性能稳定;杂散光仅对光谱的可见光部分影响明显,在近红外波段几乎没有影响;果实的不同部位对近红外光谱的影响不同,苹果赤道线上的采样点光谱重复度好,可信度高,而果实顶部和底部的光谱差异较大;在0mm距离下由于光纤探头对照射光源产生了一定的遮挡,光谱差异很大;在2.5mm-12.5mm的距离上,由于裸光纤的视角是25°,对应的采光界面没有超过苹果的最大高度,因此变化不大;超过12.5mm之后由于入射能量的损失,信号能量降低,误差变大;不同货架期的苹果失重率变化明显,由于内部水分的流失,所采集的近红外光谱差异显著;果实的不同色差在赤道面上的近红外光谱的近红外波段差异不明显而对可见光波段差异明显,因此在实际的近红外检测中可以忽略果实的表面颜色差异。(2)样品选择和建模参数研究。以相关度最好的糖度为研究对象,通过实验证明通过剔除异常样品可以大幅提高模型的精度;通过对比分析不同光谱预处理方法、不同平滑点数、不同交互验证因子数和不同建模波段对苹果糖度PLS定量预测模型预测精度的影响,得到最佳预处理方法是SNV和MSC,最佳平滑点数为3,交互验证时最佳因子数为3,建模波段的选择为整个近红外区。(3)苹果病害识别模型研究。以苹果果实轮纹病和炭疽病为实验对象,使用PCA定性分析模型可以将健康苹果和病害苹果分开,但是两种病害之间互相识别率低。利用主成分分析法将得分矩阵的前3个主成分作为输入参数,使用KNN最近邻域法、BP神经网路分类算法,将这两种病害之间的识别率提高到85%和90%。(4)实现了基于近红外光谱的苹果病害识别软件系统。更多还原

【Abstract】 Post harvest fruit diseases can be divided into a living bacterial diseases and physiological reason two categories: physiological diseases have a very good basis of previous studies, through the atmosphere, temperature and other methods can be effectively transferred control; bacterial disease of non-destructive evaluation With the exception of a small amount of electrical parameters based on characteristics of methods, very few studies, using near infrared spectroscopy for nondestructive evaluation has rarely been reported. Therefore, using fast near infrared spectroscopy, green, and other characteristics of non-destructive test to establish disease models Apple has important theoretical and practical value.Taking Apple fruit as the research object to post harvest diseases (ring rot and anthracnose) and near infrared spectroscopy to study the means, in the previous study on apple diseases on the basis of non-destructive evaluation were collected from the near-infrared spectral information Factors, sample selection and classification model parameters and model of disease areas such as research and development on this basis, near infrared spectroscopy-based disease identification software system Apple. In this paper, results are as follows:(1) Apple spectral factors of information acquisition. The same sample, the same location a total of 9 hours straight 53 proof of spectral acquisition, ASD’s Field Spec 3 Spectrometer performance and stability; stray light only visible part of the spectrum showed that the effect in the near infrared band almost no effect; fruits of different Parts of the near infrared spectra of different samples Apple equator line spectral repeatability is good, reliability, and the pedicel and fruit quite different spectrum E; 0mm distance in the fiber optic probe of the radiation sources as generated A certain block, spectrum very different; in a distance of 2.5mm-12.5mm, the perspective of the bare fiber is 25°, corresponding to the lighting interface, Apple did not exceed the maximum height, so little change; more than 12.5mm After the incident Energy loss, the signal energy is reduced, the error becomes larger; different weight loss shelf life of apple changed significantly, due to internal water loss, the acquisition significantly different near-infrared spectroscopy; fruit of different color in the equatorial plane near infrared spectroscopy No significant difference between the near infrared and visible bands of significant differences, the actual detection of near-infrared surface color of the fruit can be ignored.(2) Sample selection and modeling parameters: with the best sugar content of the relevant degree of proof for the study sample by removing the anomalies can significantly improve the accuracy of the model; comparative analysis of the different spectral pretreatment methods, different smooth points, the number of different factors and different cross validation PLS modeling band Sugar Apple predicted quantitative prediction accuracy. Concluded that the best pretreatment method for the SNV and MSC; optimal smoothing of 3 points; interaction factor authentication is the best number is 3; modeling the choice of band for the entire near infrared region.(3) Building the identify model of Apple’s disease. Taking the apple ring rot and anthracnose fruit as experimental subjects, the qualitative analysis PCA model can be controlled separately from health apples and infected apple, but low recognition rate among the two diseases. Using principal component analysis, the PCA principal component as an input parameter, use the nearest neighbor KNN, BP neural network classification algorithm, and the identification between these two diseases was increased to 85% and 90%.(4) To achieve near-infrared spectra based on Apple’s disease identification, a software system was developed.更多还原