灰霉病早期胁迫下花椰菜多酚氧化酶活性的高光谱研究  

Hyperspectral Study on Polyphenol Oxidase Content of Cauliflower at the Early Stages of Gray Mold Infection

作　　者：王凯[1] 薛建新[1] 李尧迪 张明月 WANG Kai;XUE Jian-xin;LI Yao-di;ZHANG Ming-yue(College of Agricultural Engineering,Shanxi Agricultural University,Jinzhong 030801,China)

机构地区：[1]山西农业大学农业工程学院,山西晋中030801

出　　处：《光谱学与光谱分析》2024年第2期534-541,共8页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金青年科学基金项目(31801632)资助。

摘　　要：利用高光谱技术实现灰霉病早期胁迫下花椰菜多酚氧化酶(PPO)活性的快速无损检测。为了使预测效果更好,在900~1700 nm光谱范围内采集253个健康花椰菜样本及257个染病花椰菜样本的光谱信息,并使用分光光度计法对样本中多酚氧化酶活性进行测定。对健康及染病花椰菜样本PPO活性均值进行分析,发现健康花椰菜PPO活性均值(10.257 U·g^(-1))小于染病花椰菜PPO活性均值(12.324 U·g^(-1))。利用光谱-理化值共生距离(SPXY)算法对样本进行校正集(健康样本193个,染病样本197个)和预测集(健康样本60个,染病样本60个)的划分,对划分后的样本集进行六种单一预处理(卷积平滑算法SG、去趋势算法DT、中值滤波MF、归一化处理NOR、标准正态变量变换SNV、基线校正Baseline)。利用相关系数(R)和均方根误差(RMSE)作为模型评价指标,发现预处理能够有效提高模型的精度和稳定性,其中健康样本经NOR预处理后的预测集建模效果最好;染病样本经DT预处理后的预测集建模效果最好。采用连续投影算法(SPA)与回归系数法(RC)提取特征波长,建立PPO活性值的偏最小二乘回归(PLSR)、最小二乘支持向量机(LS-SVM)及BP神经网络三种预测模型,探究不同特征波长提取方法对模型精度的影响,比较不同建模方式对花椰菜PPO活性预测的准确度。结果说明提取特征波长可以优化光谱信息,SPA和RC对两种样本提取的波长个数分别为9,12,7和11,其中SPA优化后的光谱数据对健康样本PPO活性预测效果较好,RC优化后的光谱数据对染病样本PPO活性预测效果较好。对比分析模型效果,发现LS-SVM模型对两种样本和其对应的酶活性产生了很好的拟合效果。最终发现,SPA-LS-SVM模型对于健康花椰菜PPO活性预测效果较好,其预测相关系数(Rp)为0.832,预测均方根误差(RMSEP)为1.676;RC-LS-SVM模型对于染病花椰菜PPO活性预测效果较好,其Rp为0.848,RMSEP为1.15Hyperspectral technique was applied to detect polyphenol oxidase(PPO)content in cauliflower with early botrytis stress.A total of 253 healthy cauliflower samples and 257 infected cauliflower samples were used to acquire hyperspectral within the range of 900~1700 nm,and the corresponding PPO content in the cauliflowers were measured with the spectrophotometry method in order to make the prediction effect better.The mean value was applied to the analysis of the PPO with cauliflower samples,and results showed that the mean PPO content of healthy cauliflower(10.257 U·g^(-1))was less than that of infected cauliflower(12.324 U·g^(-1)).The SPXY method divides the cauliflower sample set into a calibration set(193 healthy,and 197 infected samples)and a validation set(60 healthy and 60 infected samples).Six kinds of single pretreatment were performed on the divided sample set.The R(correlation coefficient)and RMSE(root mean square error)were used as the model evaluation index,and results showed that pretreatment can effectively improve the accuracy and stability of the mode.It was found that the predictive set modeling effect of healthy samples after NOR pretreatment is the best and that of infected samples after DT pretreatment is the best.Successive projection algorithm(SPA)and regression coefficient(RC)were used to select the characteristic wavelengths.Partial least squares regression,least squares support vector machines,and BP neural networks were built to explore the impact of different feature wavelength extraction methods on the accuracy of the model and compare the accuracy of different modeling methods on the prediction of PPO content of cauliflower.The results showed that extracting the characteristic wavelength can optimize the spectral information,and the number of wavelengths extracted by SPA and RC for the two samples were 9,12,7 and 11 respectively.It was found that the LS-SVM model has a good fitting effect on the two samples and their corresponding enzyme activities by comparing and analyzing the effect

关 键 词：花椰菜 灰霉病 多酚氧化酶 高光谱 无损检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]