【作者】 唐月明；

【导师】 何勇；

【作者基本信息】 浙江大学 ， 农业工程， 2008， 硕士

【摘要】 水稻是世界上最重要的粮食作物之一,也是我国最重要的粮食作物之一。水稻的品种共有五万多种。稻谷品种、年份鉴别和内部品质检测一直是农业生产、作物育种和种子检验上的重要问题。随着社会经济的发展,人们越来越青睐于质优价高的优质稻谷,随之而来的是制售假冒伪劣种子等现象的日益增多,每年造成的经济损失是巨大的。因此,稻谷品种、年份的鉴别和内部品质的检测问题日益受到种子质检部门、水稻育种研究以及粮食企业等单位的重视。长期以来水稻品种的鉴别和稻谷内部品质的检测主要由人工结合化学方法来完成,操作过程繁琐,工作量大,所需时间长,检测效率和检测结果的一致性都比较差。在稻谷品种的鉴别方面,目前常用的方法有形态法、化学法、幼苗法、田间小区种植法、电泳法等,近年来国内外学者采用计算机图像技术进行稻谷品种鉴别也取得了很大进展。在稻谷内部品质成分的检测方面,生产实际中主要还是采用碘比色法、凯氏定氮法等一些标准化学测量方法。建立一套简便、快捷、高效、经济、准确的稻谷检测技术体系成为当前的一个迫切需要,而红外光谱技术的发展为此提供了可能。本文针对上述目标,以稻谷为研究对象,采用红外光谱技术结合化学计量学方法和数据挖掘技术,对稻谷进行了品种和不同年份的鉴别,以及辐照剂量、直链淀粉和蛋白质含量的检测,主要研究内容和结果如下:(1)使用Field Spec Handheld光谱仪对五个水稻品种的150个样本进行可见/近红外光谱测定,采用小波变换结合主成分分析和人工神经网络的组合化学计量学方法对对获得的光谱特征进行分析,建立了稻谷品种的识别模型,对五个水稻品种进行了定性的聚类和定量的鉴别,正确率达96%。提出的组合算法为复杂体系的光谱非线形建模提供了一种有力的工具。结果表明,应用可见/近红外光谱技术结合化学计量学方法可简单、快速、无损地鉴别稻谷品种,为稻谷的品种的快速无损鉴别提供了一种新的方法。(2)试验得到了2003-2005年三年的晚粳谷的可见/近红外光谱,经处理后采用独立组分分析提取稻谷样本的敏感波段作为神经网络的输入,建立稻谷年份的BP神经网络鉴别模型,对不同年份的稻谷的识别率达到100%。通过独立组分分析,找到了晚粳谷主要成分对应的敏感波段,其中770nm,970nm对应水分含量,880nm对应脂肪含量,922nm,972nm,996nm对应稻谷中蛋白质的含量。该结果对不同种类的稻谷具有普适性。研究结果表明利用可见/近红外光谱技术结合化学计量学的方法对同年份的稻谷进行快速鉴别是可行的,它为稻谷年份的快速检测提供了一种新方法。(3)采用偏最小二乘法(PLS)和最小二乘支持向量机(LS-SVM)分别建立辐照谷物剂量及内部品质(直链淀粉和蛋白质)的线性数学模型。用偏最小二乘法(PLS)建立的较优模型所得的辐照谷物剂量测定,直链淀粉和蛋白质含量的预测相关系数分别为0.978,0.911和0.943,预测均方根误差为114.902,0.250和0.102,偏差为-6.032E-03,-6.012E-04和2.150E-05,其中对于辐照剂量的预测和蛋白质含量测定,中红外波段好于近红外波段;对于直链淀粉含量的测定,近红外波段优于中红外波段。用最小二乘支持向量机(LS-SVM)建立的最优模型所得的辐照谷物剂量测定,直链淀粉和蛋白质含量的预测相关系数分别为0.989,0.951和0.982,预测均方根误差为95.763,0.201和0.052,偏差为-3.621E-03,-1.302E-06和-2.105E-07,其中对于辐照剂量的预测和蛋白质含量测定,中红外波段好于近红外波段,对于直链淀粉含量的测定,近红外波段优于中红外。研究结果表明最小二乘支持向量机具有较好的预测准确性和抗干扰性,可以得到较好的预测结果。更多还原

【Abstract】 Rice is one of the most important crops in the world,also in China,and the total species of rice are more than 50,000.Rice varieties,stored years identification and internal quality testing are the important issues in agricultural production,crop and seed breeding.With the development of social economy,more and more people favor the high-quality rice.But there also exist phenomena like manufacturing and selling fake and shoddy seeds,which bring the enormous economic losses every year. Therefore,seed qu.ality inspection departments,rice breeding and grain enterprises and other units give more attention on rice varieties,year identification and rice internal quality prediction.Presently,the identification of rice varieties and rice internal quality are mainly by the combination of chemical methods,which are with heavy workload,long time,poor efficiency and bad test results.In the identification of rice varieties,the current methods are morphology,chemical method,seedlings, planting field plot,electrophoresis and so on.In recent years,domestic and foreign scholars have also made great progress in computing imaging technology to identify rice varieties.In rice quality prediction,the main methods like iodine colorimetry, Kjeldahl method,and some other standard methods.So,establish a simple,fast, efficient,economic and accurate detection of rice become a pressing need,and the infrared spectra technology provides the possibility for development.In this paper,according to the above objectives,we used infrared spectroscopy, chemometrics and data mining technology to identify rice variety and different years, and internal quality prediction（starch and protein） after radiation,the main research contents and findings are as follows:（1） A simple,fast and non-destructive method was put forward for discriminating varieties of paddy,and the method was based on visible/near infrared reflectance （Vis/NIR） spectroscopy and chemometrics.Firstly,thousands of spectral data of different varieties were denoised by wavelet transform（WT）.Secondly,principal component analysis（PCA） compressed the above data into several variables.The analysis suggested that the first four PCs（principle components） could account for 99.89%of the original spectral information.In order to set up the model for discriminating varieties of paddy,the four diagnostic PCs were applied as inputs of back propagation artificial neural network（BP-ANN）,and the values of different paddy varieties were applied as the outputs of BP-ANN.The whole 150 samples were randomly divided into two parts,one of them that consisted of 100 samples was used to model,and the other one contained 50 samples were used to predict.This model had been used to predict the varieties of 50 unknown samples,and the discrimination rate 96%had been obtained.It was proved that the model was very reliable and practicable.In short,this paper could offer a new approach to discriminate varieties of paddy.（2） A new method for discrimination stored years of rough rice from 2003-2005 based on independent component analysis was developed through using visible/near infrared spectroscopy（Vis/NIRS）.First,the vis/NIR loading weight of rough rice with different years were got through using independent component analysis（ICA）,set the wavelengths corresponding to the maximal correlation as inputs of artificial neural network（ANN）,then build the discrimination model.120 samples（40 with each year） from three years were selected randomly as calibration set;the left 60 samples（20 with each year） were as perdition set.The discrimination rate of 100%was achieved. Synchronously,the sensitive wavelengths corresponding to the main components in rough rice were obtained through ICA.The wavelengths 770nm and 970nm are corresponding to the moisture content,880nm corresponding to the starch,and 922nm, 972nm,996nm to the protein content in rice.It indicated that the result for discrimination years of rough rice was very good based on ICA method,and it offers a new approach to the fast discrimination years of rough rice.（3） Partial least squares（PLS） analysis and least squares-support vector machine （LS-SVM） were implemented to predict the radiation doses and components of starch and protein in rice treated with different radiation doses（0Gy,250Gy,500Gy,750Gy, 1000Gy,1500Gy,2000Gy,2500Gy,3000Gy） based on sensitive wavelengths（SWs） and chemometrics.In the PLS models,the correlation coefficient（r）,root mean square error of prediction（RMSEP） and bias in prediction set were 0.978,114.902, -6.032E-03 for radiation doses detection,0.911,0.250 and-6.012E-04 for starch, 0.943,0.102 and 2.150E-05 for protein,respectively.To radiation doses detection and protein prediction,mid infrared（MIR）(400-4000 cm^-1) region was better than near infrared（NIR）（1100-2500nm）,to starch prediction,the NIR region was better.In the LS-SVM models,the correlation coefficient（r）,root mean square error of prediction （RMSEP） and bias in prediction set were 0.989,95.763,-3.621E-03 for radiation doses detection,0.951,0.201 and-1.302E-06 for starch,0.982,0.052 and -2.105E-07 for protein,respectively.To radiation doses detection and protein prediction,mid infrared（MIR）(400-4000cm^-1) region was better than near infrared（NIR） （1100-2500nm）,to starch prediction,the NIR region was better.The overall results indicated that IR spectroscopy combined with LS-SVM could be applied as a high precision and fast way for the radiation doses detection and starch and protein prediction in rice after radiation.更多还原