【作者】 王琦；

【导师】 何勇；

【作者基本信息】 浙江大学 ， 生物系统工程， 2019， 硕士

【摘要】 秋葵是我国常见的食用作物,具有高营养价值和观赏价值,其品种繁多,产地遍布亚洲、欧洲、北美洲、非洲和澳洲。不同品种秋葵的环境抗性存在较大差异,因此在本地化育种栽培过程中,往往需要借助表型信息筛选高产量秋葵品种。目前在生产和育种筛选上,对于作物表型的无损检测技术有着迫切的需求。传统作物表型参数的获取依靠人工测量,人力成本高、时效性差且测量结果易受主观判断的影响。高光谱成像技术是一种非破坏性且快速高效的分析手段,近几年在植物表型分析领域发挥着越来越重要的作用。在利用光谱技术分析植物表型过程中,根据不同作物开发鲁棒性强的算法是作物表型测量和分析的关键所在。此外,结合光谱成像技术解析的表型信息,能有效地减少环境因素的干扰,也是实现精细化生产和育种筛选的关键技术。本研究针对秋葵育种栽培中的植株定位、表型信息获取和耐盐性检测等重要问题,基于光谱成像、图像处理和数学建模技术,采用图像分割、高光谱特征波长提取、光谱图像信息融合、层次聚类等理论方法,实现了对秋葵表型参数的快速检测,应用表型信息和光谱信息实现多品种秋葵耐盐性水平快速分类,主要研究成果如下:（1）盐胁迫下秋葵生理生化机理研究为探究盐胁迫对秋葵表型的影响机理,试验测量了秋葵植株在盐胁迫下的生物量变化、Na、K等元素含量、光合参数等信息,并与未胁迫的对照组做了比较。试验结果表明,盐胁迫会导致秋葵植株根系大量吸收Na离子,K/Na比例遭到破坏。同时,盐胁迫对秋葵叶片光合作用有较大的抑制作用,在胁迫初期,光合速率受影响的主要原因是叶片气孔关闭,导致CO₂供应受阻,在长时间的胁迫后,叶肉光合活性下降导致叶片光合强度下降。（2）基于深度学习的秋葵自动识别分割和植株生物量预测本研究提出一个基于深度学习技术的图像分割算法,融合了语义分割模块和实例分割模块,可减少模型参数并降低计算量。语义分割模块用于分割植株的整体区域,可实现植株冠层图像面积获取、生物量估测、光谱信息可视化;实例分割模块将每片叶片分割为单独个体,为叶片计数、叶面积估测、叶位估计和叶片平均光谱计算提供了有力的技术支持。针对叶片高重叠度的问题,在算法中引入注意力机制循环解码,实现了对叶片的分割。为了计算各个叶片在解码过程中的注意力分配,提出候选概率图结构,通过策略搜索方式循环选取待分割叶片,并将计算得到的注意力点转换为注意力编码特征图（Attention Coding Channel,AAC）嵌入到解码网络中,从而得到更精细的分割结果。同时,采用级联式掩码预测以提高解码精度,在各个分辨率上预测叶片的分割掩码并与人工标注的掩码进行对比以计算损失函数。为训练秋葵分割模型,将高光谱成像获得的732张秋葵植株图像进行人工标注,共包含了2000余株秋葵植株。模型在标注好的秋葵数据集上训练,待收敛后,语义分割模块在预测集上的分割精度（IoU）达到0.94,能将植株完整的从背景中分离,与传统阈值分割方法相比有明显提升;实例分割模块在预测集上的对称最优Dice系数（SBD）为81.4,实现了对每片叶片的独立分割。冠层图像面积与植株的生物量间有较高的相关性。利用图像分割得到的冠层图像面积建立一元线性回归模型以预测植株的生物量。鲜重与冠层面积的相关系数达0.808,干重与冠层面积的相关系数达0.774。结果表明利用光谱成像技术能够有效预测植株的生物量。（3）基于可见光-近红外波段光谱信息的秋葵SPAD值预测对全波段建立偏最小二乘回归模型,使用K-折叠交叉验证,当主成分数为12时效果最好,此时训练集决定系数为0.7368,均方根误差3.3796,交叉验证集决定系数0.7148,均方根误差3.4838。由于高光谱各波长信息有高度相关性。因此采用特征波长提取算法提取少量特征波长进行建模,有效地缓解模型过拟合的问题。同时,少量的特征波长更有利于成分分析,并降低光谱检测仪器开发成本。本研究提出一种新的特征波长搜索算法——自适应快速蒙特卡洛树搜索（AFMCTS）并与经典特征波长提取算法进行比较。该算法实现了特征波长组解空间的高效搜索,能够在少量波长数下迅速逼近最优解。采用AFMCTS算法提取特征波长并通过多元线性回归建立秋葵SPAD值预测模型,当选择的特征波长数为6时,训练集决定系数0.7336,均方根误差3.4001,交叉验证集决定系数0.7217,均方根误差3.4352,此时结果已基本收敛。采用CARS算法提取得到18个特征波长,对提取的特征波长采用偏最小二乘回归建模,训练集决定系数0.7225,均方根误差3.4703,交叉验证集决定系数0.7086,均方根误差3.5205,该方法提取的特征波长数较多,且互信息较大。采用SPA算法在提取14个特征变量时得到最优结果,训练集决定系数0.7351,均方根误差3.3905,交叉验证集决定系数0.7182,均方根误差3.4656。对比AFMCTS算法和经典的特征波长提取算法,本研究提出的方法鲁棒性更强,达到同样精度需要的特征波长数更少,能得到更丰富的搜索结果,因此该方法提取的特征波长组更适合于实际应用。（4）基于表型信息和光谱信息的多品种秋葵耐盐性水平聚类通过表型信息评价作物的生长状况在育种筛选中具有重要意义。为评价盐胁迫对于秋葵生长的影响,引入双样本检测度量方法最大均值差异（MMD）。使用MMD计算各个品种秋葵植株在盐胁迫前后表型信息和理化信息的分布差异,前后分布差异越大表明该品种在盐胁迫下受影响越严重,在此基础上统计了14类基因型秋葵样本的耐盐性水平,其中耐盐性最强的品种是‘丹指’,而‘纤指’、‘东京五角’、‘五福’三个品种耐盐性相对较差。试验结果表明,通过表型信息能有效地评估秋葵的耐盐性水平。但作物的表型信息和理化信息存在检测繁琐的缺点,难以在实际生产中普及。而光谱具有快速无损检测的优点,且光谱信息与植物表型存在内在联系,通过光谱信息对各品种秋葵进行聚类分析,聚类结果与基于表型信息的聚类结果相近,耐盐性最强的‘丹指’被单独分为一类。因此,利用光谱信息评估秋葵耐盐性水平是可行的,可以作为替代表型信息的有效手段。更多还原

【Abstract】 Okra is a common edible crop in China.It has high nutritional value and ornamental value.It has a wide range of varieties and origins in Asia,Europe,North America,Africa and Australia.The environmental resistance of different varieties of okra varies greatly.Therefore,in the process of local breeding and cultivation,it is necessary to use phenotypic information to screen high-yield okra varieties.Nowadays,non-destructive testing technology of crop phenotype has become an urgent need in production and breeding screening.Traditional crop phenotypic parameters acquisition relies on manual measurement,which has high human cost,poor timeliness and the measurement results are vulnerable to subjective judgment.Hyperspectral imaging is a non-destructive,fast and efficient method for plant phenotypic analysis,which has played an increasingly important role in recent years.In the process of plant phenotypic analysis by spectral technology,developing robust algorithms based on different crops has become the key to crop phenotypic measurement and analysis.In addition,the combination of spectral imaging technology to analyze the phenotypic information can effectively reduce the interference of environmental factors,and is the key technology to achieve fine production and breeding screening.In view of the important problems of plant location,phenotypic information acquisition and salt tolerance detection in okra breeding and cultivation,based on spectral imaging,image processing and mathematical modeling technology,this study adopts image segmentation,Hyperspectral Feature wavelength extraction,spectral image information fusion,hierarchical clustering and other theoretical methods to realize the rapid detection of okra phenotypic parameters,and applies phenotypic information and spectral information.The main research results are as follows:（1）Study on physiological and biochemical mechanism of okra under Salt StressIn order to explore the mechanism of salt stress on the phenotype of okra,the changes of biomass,element content and photosynthetic parameters of okra under salt stress were measured and compared with those of the control group without salt stress.The results showed that salt stress could cause root system to absorb large amounts of Na ions,and the K/Na ratio was destroyed.At the same time,salt stress inhibited the photosynthesis of okra leaves.In the early stage of stress,the main reason of photosynthesis rate affected was stomatal closure of leaves,which led to the obstruction of CO₂ supply.After long-term stress,the decline of mesophyll photosynthetic activity resulted in the loss of photosynthetic capacity of okra leaves.（2）Automatic recognition and segmentation of okra based on in-depth learning and prediction of plant biomassIn this study,an image segmentation algorithm model based on deeplearning technology is proposed,which integrates the semantic segmentation module and the instance segmentation module,reduces the model parameters and reduces the computational load.Semantic segmentation module is used to segment the whole plant area,so as to achieve the acquisition of plant canopy area,biomass estimation and visualization of spectral information.Example segmentation module can divide each leaf into individual,providing technical support for the realization of leaf count,leaf area estimation,leaf position estimation and leaf average spectral calculation.Aiming at the problem of high overlap degree of blades,attention mechanism cyclic decoding is introduced to realize blade segmentation.In order to calculate the attention allocation of each blade in decoding process,a candidate probability graph structure is proposed.The blade to be segmented is selected by strategy search cycle,and the calculated attention points are converted into attention coding channel（AAC）embedded in decoding network,so as to obtain more fine segmentation results.At the same time,in order to improve the decoding accuracy,cascade mask prediction is used to predict the blade segmentation mask at each resolution and to calculate the loss function by comparing it with the manual labeling mask.In order to train the segmentation model of okra,732 images of okra plants obtained by hyperspectral imaging were labeled artificially.The data included more than 2000 okra plants.The model is trained on the labeled okra data set.After convergence,the semantics segmentation module achieves 0.94segmentation accuracy（IoU）on the prediction set,which can separate the plant completely from the background,and obviously improves compared with the traditional threshold segmentation method.The symmetrical best dice coefficient（SBD）of the example segmentation module on the prediction set is 81.4,which achieves the independent segmentation of each leaf.There was a high correlation between canopy image area and plant biomass.The canopy image area obtained by image segmentation was used to establish a linear regression model to predict plant biomass.For the prediction model of fresh weight of okra plant,the determination coefficient was 0.6526,and for the prediction model of dry weight of okra plant,the determination coefficient was 0.5997.The results showed that spectral imaging technology could effectively predict plant biomass.（3）SPAD Value Prediction of Okra Based on Visible-Near Infrared Spectral InformationSPAD value is a commonly used indirect evaluation index of chlorophyll content.In this study,the prediction model of SPAD value of okra was established by using visible-near infrared spectroscopy information.Partial least squares regression model is established for the whole band.The K-folding interactive validation shows that when the number of principal components is 12,the effect is the best.At this time,the decision coefficient of training set is 0.7368,the root mean square error is 3.3796,the cross-validation set is 0.7148 and the root mean square error is 3.4838.Because the information of each wavelength in hyperspectral spectrum is highly correlated.Therefore,the feature wavelength extraction algorithm is used to extract a small number of feature wavelengths for modeling,which can effectively alleviate the over-fitting of the model.At the same time,a small number of characteristic wavelengths are more conducive to component analysis and reduce the development cost of spectral detection instruments.In this study,a new feature wavelength search algorithm,adaptive fast Monte Carlo tree search（AFMCTS）,is proposed and compared with the classical feature wavelength extraction algorithm.The algorithm achieves efficient searching in the solution space of characteristic wavelength group,and can quickly approach the optimal solution with a small number of wavelengths.The characteristic wavelengths are extracted by using AFMCTS algorithm and the SPAD value prediction model is established by multiple linear regression.When the characteristic wavelengths are 6,the decision coefficient of training set is 0.7336,the root mean square error is 3.4001,the cross validation set is 0.7217,and the root mean square error is 3.4352.At this time,the results are basically convergent.The eighteen characteristic wavelengths are extracted by CARS algorithm.The partial least squares regression model is used to model the extracted characteristic wavelengths.The determination coefficient of training set is 0.7225,the root mean square error is 3.4703,the determination coefficient of cross validation set is 0.7086,and the root mean square error is 3.5205.The SPA algorithm is used to extract 14 characteristic variables,and the optimal results are obtained.The decision coefficient of training set is 0.7351,the root mean square error is 3.3905,the cross validation set is 0.7182,and the root mean square error is 3.4656.Compared with the classical feature wavelength extraction algorithm and the AFMCTS algorithm,the proposed method has stronger robustness and requires fewer feature wavelengths with the same accuracy.It can get more abundant search results.Therefore,the feature wavelength group extracted by this method is more suitable for practical application.（4）Clustering of salt tolerance level of multi-cultivars of okra based on phenotypic and spectral information.Evaluating crop growth through phenotypic information is of great significance in breeding screening.In order to evaluate the effect of salt stress on the growth of okra,the maximum mean difference（MMD）was introduced by using two-sample detection method.MMD was used to calculate the distribution differences of phenotypic information and physical and chemical information before and after salt stress.The greater the difference was,the more seriously the variety was affected under salt stress.Based on this,the salt tolerance levels of 14 genotypes of okra were calculated.The results showed that the most salt tolerant varieties were`Danzhi’,`Xianzhi’,`Tokyo Wujiao’,`Wufu’.The salt tolerance of the three varieties was relatively poor.The results showed that the salt tolerance level of okra could be effectively evaluated by phenotypic information.However,crop phenotypic information and physical and chemical information have the disadvantage of tedious detection,which is difficult to popularize in practical production.Spectrum has the advantages of fast non-destructive testing,and there is an inherent relationship between spectral information and plant phenotype.Clustering analysis of various varieties of okra by spectral information shows that the clustering results are similar to those based on phenotypic information.Salt tolerance of the strongest’Danzhi’is divided into a separate category.Therefore,it is feasible to evaluate the salt tolerance level of okra using spectral information,which can be used as an effective means to replace phenotypic information.更多还原