【作者】 张俊；

【导师】 成芳；

【作者基本信息】 浙江大学 ， 农业机械化工程， 2022， 博士

【摘要】 玉米(Zea mays L.)是重要的粮食和饲料作物,在我国农业生产中占有举足轻重的地位。种子是农业生产最基本的生产资料,其质量的高低对播种后种子的生长发育成苗等有很大的影响。河西走廊地区拥有我国最大的玉米制种基地,该地区常有寒潮和冻害发生,对玉米种子质量产生严重影响。为实现种子质量的精细管控,冻害玉米种子的无损、准确识别尤为重要,对提高种子利用率、保障作物生产具有重要的积极作用。本文首先针对目前冻害玉米种子评价指标单一、缺少相对全面的冻害程度判定标准等问题,从种子本身出发,测试不同冻害玉米种子生理生化指标并研究其变化情况,综合各项指标的聚类结果设置冻害程度判定标准,为冻害种子的显微特征分析和检测识别时冻害样本的制备提供依据;然后针对玉米种子冻害特征分析不深入的问题,对受不同程度冻害影响的种子主要部位(种皮和种胚细胞)的微观结构进行表征,明确玉米种子受冻害影响的微观变化规律;进而针对目前冻害玉米种子缺少无损检测方法的问题,基于近红外光谱技术和高光谱成像技术,研究群体和单粒玉米种子的冻害识别方法,利用二维相关分析方法提取光谱特征波长,并与其他特征提取算法提取的波长进行融合,为建立基于特征波长的冻害玉米种子快速无损多光谱检测系统提供依据;最后基于深度学习在数据处理方面的优势,对比卷积神经网络方法与传统建模算法对冻害玉米种子的识别效果,建立了一种端到端的冻害玉米种子无损识别模型。主要研究内容和结论如下:(1)根据多个理化指标的聚类结果设置了冻害程度判定标准,为制备具有代表性和可重复性的冻害种子样本提供依据。测定了高、低含水率(16-20%,24-28%)下3个玉米品种在不同冻害条件下的理化指标。结果表明:在相同冻害条件下,低含水种子的发芽率、主芽长度和主根长度变化情况小于高含水玉米种子。随着冻害程度的加深,电导率数值增加、过氧化物酶和超氧化物歧化酶活性先增加后减小、过氧化氢酶活性增加、可溶性糖含量降低。最终3个玉米品种种子的冻害程度判定标准为:低含水情况下,未受冻或-5℃(5h,10 h)或-10℃(5 h,10 h)冻害条件的玉米种子为正常样本;-15℃(5 h,10 h)或-20℃(5 h,10 h)冻害条件的玉米种子为轻/中度冻害样本。高含水情况下,未受冻与或-5℃,5 h冻害条件的玉米种子为正常样本;-5℃,10 h或-10℃(5 h,10 h)或-15℃,5 h冻害条件的玉米种子为轻/中度冻害样本;-15℃,10 h或-20℃(5 h,10 h)冻害条件的玉米种子为重度冻害样本。(2)对受不同程度冻害影响的种皮纹理和种胚细胞的微观结构进行表征,明确了微观变化规律。利用扫描电子显微镜观察种皮表面纹理的变化情况,结果表明冻害对种皮表面纹理有一定影响。随着冻害程度的加深,种皮的纹理不清晰、平行度降低并且出现交叉纹理的情况,冻害程度高的平行纹理之间交叉严重、褶皱明显。利用透射电子显微镜观察种胚细胞结构,结果表明冻害对种胚细胞结构有明显影响。冻害温度越低,对细胞结构完整性的影响越显著,细胞质壁分离情况越严重,细胞和细胞核的完整性越差;冻害程度更高时,会出现核膜消失、磷脂分子严重缺失等情况。(3)基于近红外光谱技术实现了群体玉米种子冻害的无损分类识别。结果表明:基于标准正态变换预处理和多元散射校正预处理算法的结果优于无预处理和5点3次平滑预处理的分类结果;基于主成分分析方法获得的分类结果优于连续投影算法、载荷系数法和二维相关分析方法;k最邻近算法的分类结果优于线性判别分析和支持向量机算法,标准正态变换预处理-k最邻近模型获得最优的分类结果,校正集最高准确率达到99.4%,验证集最高准确率达到100%。(4)基于高光谱成像技术分别针对胚乳面和胚面实现了单粒冻害玉米种子的分类识别。利用图像处理算法分别对700 nm波长和500 nm波长的灰度图像进行阈值分割获得胚乳面和胚部平均光谱。胚乳面光谱分析结果表明:基于全波段的标准正态变换预处理-支持向量机模型获得最优的分类结果,校正集最高准确率达到92.9%、验证集最高准确率达到90.0%;基于特征波长的标准正态变换预处理-连续投影算法融合二维相关分析方法-支持向量机模型获得最优的分类结果,校正集最高准确率达到92.9%、验证集最高准确率达到91.2%。胚面光谱分析结果表明:基于全波段的5点3次平滑预处理-线性判别分析模型获得最优的分类结果,校正集准确率最高达到97.7%、验证集准确率最高达到95.9%;基于特征波长的无预处理-连续投影算法融合二维相关分析方法-线性判别分析模型获得最优的分类结果,校正集准确率最高达到97.5%、验证集准确率最高达到94.9%。融合了二维相关分析方法提取的波长后的建模效果比融合前有了一定的提高。胚乳面种子仅用19个特征波长就获得比全波段更好的建模效果,而胚面种子仅用15个特征波长就获得良好的建模效果,这为后续建立基于特征波长的冻害玉米种子快速无损多光谱检测系统提供依据。与胚乳面冻害玉米种子识别效果相比,胚面的识别效果更优。(5)建立了基于高光谱成像技术和深度卷积神经网络算法的单粒冻害玉米种子识别模型。对比了传统建模算法和深度卷积神经网络算法的单粒玉米种子冻害识别的建模效果,并对玉米种子胚乳面和胚面的结果进行对比。结果表明:相较于支持向量机模型,深度卷积神经网络模型在准确率、灵敏度、特异度和精确度等指标上结果更优,其中验证集准确率至少有3.5%的提高。玉米种子胚面的分类结果优于胚乳面,胚面深度卷积神经网络模型验证集最低准确率为96.7%,而胚乳面深度卷积神经网络模型验证集最高准确率为93.8%。更多还原

【Abstract】 Maize(Zea mays L.)is a major food and feed crop,which occupies a crucial position in the agricultural production of China.As the most basic means of production in agricultural production,the seed quality has a great impact on the subsequent growth and development of seeds after sowing.The Hexi Corridor area has the largest corn seed production base in China,due to frequent cold weather and freezing injuries occur in this area,which have serious impact the corn seed quality.In order to realize the fine management and control of seed quality,the non-destructive and accurate identification of freeze-damaged corn seeds will play a positive role in promoting corn production and improving seed utilization.Firstly,in view of the current situation that the evaluation indicator of freezing-damaged is single and there is no relatively general freezing damage judgment standard,this paper explores the changes of physiological and biochemical indicators of freezing damaged corn seeds,and set the freezing damage judgment standard based on the clustering results of each indicator,which provides a basis for the preparation of freeze-damaged samples of the microscopic characteristic analysis and identification of freezing-damaged seeds;Then,in view of the incomplete analysis of freezing damage characteristics of corn seeds,the microstructure of the main parts(seed coat and seed embryo cells)affected by different degrees of freezing damage are characterized,and the microscopic changes of the freezing-damaged corn seeds is clarified;Furthermore,in view of the lack of non-destructive testing methods for freezing damage of corn seeds,the freezingdamage identification methods for batch corn seeds and single corn seeds based on near-infrared spectroscopy and hyperspectral imaging technology are proposed.Two-dimensional correlation analysis method is used to extract spectral feature wavelengths,and the wavelengths are fused with the wavelengths obtained by other feature extraction algorithms to provide a basis for establishing a fast non-destructive multispectral detection system for freezing-damaged corn seeds based on feature wavelengths.At last,based on the advantages of deep learning in data processing,the convolutional neural network method and the traditional modeling algorithm are used to compare the classification effects of freeze-damaged corn seeds,in order to establish an end-to-end non-destructive and good identification model of freeze-damaged corn seeds.The main conclusions of the paper are as follows:(1)A freezing damage judgment standard was comprehensively set using the clustering results of multiple physical and chemical indicators,which provided the basis for preparing representative and repeatable freezing-damaged seed samples.Physiological and biochemical indicators of three corn varieties under different freezing-damaged conditions with the condition of high and low moisture content(16-20%,24-28%)were measured.The results showed that under the same freezing-damage conditions,the germination condition of low-moisture corn seeds was better than that of high-moisture corn seeds.With the deepening of freezing damage,the conductivity increased,the peroxidase and superoxide dismutase activities increased first and then decreased,the catalase activity increased,and the soluble sugar content decreased.The freezing damage judgment standard of the three corn varieties were: Under low moisture content condition,the corn seeds that are not frozen or frozen at-5°C(5 h,10 h)or-10°C(5 h,10 h)conditions are normal samples;the corn seeds frozen at-15°C(5 h,10 h)or-20°C(5 h,10 h)are light/moderate freezing-damaged samples.Under the high moisture content condition,the corn seeds that are not frozen or frozen-5°C,5 h conditions are normal samples;corn seeds frozen at-5°C,10 h or-10°C(5 h,10 h)or-15°C,5 h are light/moderate freezing-damaged samples;the corn seeds frozen at-15℃,10 h or-20℃(5 h,10 h)were the severe freezing-damaged samples.(2)The microscopic changes of the seed coat texture and seed embryo cell structure affected by different degrees of freezing damage were characterized,and the microscopic changes rule was clarified.The scanning electron microscope was used to observe the uniformity and texture change of the seed coat.The results showed that freezing damage had a certain influence on the texture of the seed coat.With the deepening of freezing damage,the texture of the seed coat was unclear,the parallelism was reduced,and the cross-texture appears.The transmission electron microscope was used to observe the cell structure of the seed embryo,the results showed that the freezing damage had obvious influence on the seed embryo cell structure.The lower the freezing temperature,the more serious the impact on the integrity of the cell structure,the more severe the separation of the cytoplasmic wall,and the worse the integrity of the cell and nucleus.The nuclear membrane disappears and the phospholipid molecule is seriously missing sometimes.(3)The non-destructive classification and identification of freezing damage of batch corn seeds was realized based on near infrared spectroscopy.The results showed that the classification results under standard normal transformation pretreatment and multivariate scattering correction pretreatment were better than the results under no pretreatment and 5-point3-time smoothing pretreatment;Comparing with the modeling results of other feature extraction algorithms,the classification results obtained by the principal component analysis method were better than other algorithms;The classification results of the k nearest neighbor algorithm were better than the results of linear discriminant analysis and support vector machine.The optimal result was the standard normal transformation pretreatment-k nearest neighbor model,with the99.4% accuracy rate of the calibration set,and the 100% accuracy rate of the validation set.(4)The classification and identification of freezing damage on the endosperm side and embryo side of single-grain corn seeds were realized respectively based on hyperspectral imaging technology.The average spectra of endosperm and embryo were obtained by threshold segmentation of gray image in 700 nm and 500 nm by the image processing algorithm.The spectral analysis results of the endosperm side showed that: the standard normal transformation pretreatment-support vector machine model based on the the full wavelength achieved the optimal result,with the highest accuracy rate of 92.9% for the calibration set and 90.0% for the validation set;the standard normal transformation pretreatment-successive projection algorithm combined with two-dimensional correlation analysis feature extraction-support vector machine model based on the feature wavelengths achieved the optimal result,with 92.9% accuracy rate of the calibration set and 91.2% accuracy rate of the validation set.The spectral analysis results of the embryo side showed that: the 5-point 3-time smoothing pretreatment-linear discriminant analysis model based on the the full wavelength achieved the optimal result,with the highest accuracy rate of 97.7%for the calibration set and 95.9% for the validation set;the none pretreatment-successive projection algorithm combined with two-dimensional correlation analysis feature extractionlinear discriminant analysis model based on the feature wavelengths achieved the optimal result with 97.5% accuracy rate of the calibration set and 94.9% accuracy rate of the validation set.The model effects after fusing the wavelengths extracted by the two-dimensional correlation analysis method had a certain improvement compared with those before fusion.Better modeling results than full wavelengths were obtained with only 13 feature wavelengths in endosperm side,while wonderful modeling results obtained in embryo side with only 15 feature wavelengths,which provides feature wavelengths basis for the rapid non-destructive multispectral detection system of freezing-damaged corn seeds.Better identification effects were obtained in embryo side than the rusults of endosperm side.(5)A single-grain freeze-damaged corn seed identification model based on hyperspectral imaging technology combined with deep convolutional neural network algorithm was established.The modeling effects of the traditional modeling algorithm and the deep convolutional neural network algorithm for the identification of freezing damage of single corn seeds were compared,and the results of the endosperm and embryo side were compared.The results showed that compared with the support vector machine model,the deep convolutional neural network model achieved better results in terms of accuracy,sensitivity,specificity and precision,and the accuracy on the validation set was improved by at least 3.5%.The classification results of the embryo side of corn seeds were better than that of the endosperm side.The lowest96.7% accuracy of the validation set of the embryo deep convolutional neural network model was higher than the highest 93.8% accuracy of the endosperm deep convolutional neural network model.更多还原