利用特征波段提取及结合机器学习对小米淀粉的高光谱检测研究  

Hyperspectral Detection Methods of Starch of Millet byFeature Bands Extraction Combined with Machine Learning Algorithm

作　　者：王国梁 赵媛 刘敏 郭二虎 王瑞[1] 范惠萍 李瑜辉 张艾英[1] Wang Guoliang;Zhao Yuan;Liu Min;Guo Erhu;Wang Rui;Fan Huiping;Li Yuhui;Zhang Aiying(Millet Research Institute,Shanxi Agricultural University,Changzhi 046000;College of Agriculture,Shanxi Agricultural University,Taigu 030801)

机构地区：[1]山西农业大学谷子研究所,长治046000 [2]山西农业大学农学院,太谷030801

出　　处：《中国粮油学报》2024年第4期149-157,共9页Journal of the Chinese Cereals and Oils Association

基　　金：杂粮种质资源创新与分子育种国家实验室项目(202204010910001-13);国家现代农业产业技术体系建设专项(CARS-06-14.5-A21);山西省现代农业产业技术体系谷子体系项目(2023CYJSTX04-04)。

摘　　要：运用高光谱检测技术实现小米淀粉的快速检测在小米定级、定价及降低加工成本中具有重要意义。本研究基于高光谱检测技术,采用化学计量学及机器学习相关知识对小米直链、支链淀粉含量进行检测,并提出特征波段提取联用预处理方法及Logistic结合COOT(coot optimization algorithm)优化算法。结果表明采用特征波段提取联用算法建立的PLSR(partial least squares regression)模型能够在减少波段冗余情况下不影响模型预测精度,其中直链淀粉较好模型为MSC(multiplicative scatter correction)-RF(random frog)-IRIV(iteratively retains informative variables)-PLSR,支链淀粉较好模型为MSC-CARS(competitive adaptive reweighted sampling)-IRIV-PLSR。为了进一步提高模型预测精度,基于最佳预处理算法结合Logistic-COOT建立BP(back propagation)预测模型能够较好地预测小米直链、支链淀粉的含量,模型评价直链、支链淀粉相关系数(correlation coefficient,R)、均方根误差(root mean squared error,RMSE)、相对分析误差(relative percent deviation,RPD)分别为0.74、1.19、1.51和0.72、5.25、1.40,研究可为小米其他营养成分的高光谱检测及产品分类、定级等提供理论参考。The rapid detection of millet starch by hyperspectral technology is of great significance in millet grading,pricing and reducing processing costs.In this paper,based on hyperspectral detection technology,the content of amylose and amylopectin in millet was detected by using chemometrics and machine learning algorithm,and the pretreatment methods feature bands extraction sequential combination and Logistic combined with coot optimization algorithm optimization algorithm were proposed.The results indicated that partial least squares regression model established by feature bands extraction sequential combination can reduce the bands redundancy without affecting the prediction accuracy of the model,the better prediction model for amylose was MSC-RF-IRIV-PLSR,and the better prediction model for amylopectin was MSC-CARS-IRIV-PLSR.In order to further improve the accuracy of the model prediction,BP model based on the best pretreatment method combined with Logistic-COOT could predict the content of amylose and amylopectin in millet,R(correlation coefficient),RMSE(root mean squared error)and RPD(relative percent deviation)of amylose&amylopectin were 0.74,1.19,1.51;0.72,5.25,1.40,respectively.This study can provide a reference for hyperspectral in other nutritional components of millet and product classification or grading.

关 键 词：小米淀粉 高光谱检测 特征波段提取联用 机器学习 

分 类 号：O657.3[理学—分析化学]