基于Si-cPLS的小麦种子发芽率近红外模型优化研究  

Research on Optimization of Wheat Seed Germination Rate NIR Model Based on Si-cPLS

作　　者：吴静珠[1,2] 董文菲 董晶晶[1] 陈岩[1] 毛文华[3] 刘翠玲[1] 

机构地区：[1]北京工商大学计算机与信息工程学院,北京100048 [2]土壤植物机器系统技术国家重点实验室,北京100083 [3]中国农业机械化科学研究院,北京100083

出　　处：《光谱学与光谱分析》2017年第4期1114-1117,共4页Spectroscopy and Spectral Analysis

基　　金：国家国际科技合作专项项目(2014DFA31660);土壤植物机器系统技术国家重点实验室开放课题(2014-SKL-05);北京工商大学两科基金培育项目(19008001110)资助

摘　　要：为了提高近红外光谱技术快速测定小麦种子发芽率的准确度和稳健性,比较分析了基于全光谱的单一偏最小二乘(PLS)模型和多模型共识PLS模型(cPLS)性能,并提出了基于特征光谱的多模型共识PLS模型(Si-cPLS)。实验收集84份小麦种子,通过SPXY法将样本集划分为训练集样本66个,预测集样本18个。从训练集中随机抽取50个样本作为校正集建立一系列PLS子模型,选取其中性能较好100个子模型作为成员模型建立cPLS模型,取成员模型预测结果的均值来分析未知样本。在此基础上,采用组合间隔偏最小二乘法(SiPLS)筛选特征谱区建立多模型共识PLS模型(Si-cPLS)。各模型均采用均值中心化预处理方法,PLS模型、cPLS模型以及Si-cPLS模型对预测集的小麦种子发芽率进行50次重复预测的平均相关系数r分别为0.935,0.949和0.967,平均预测均方根误差RMSEP分别为13.735%,12.533%和10.273%,RMSEP的标准偏差分别为1.144%,0.096%和0.08%。实验结果表明cPLS模型较单一PLS模型更加稳定可靠,而基于特征光谱的Si-cPLS模型则进一步提高了cPLS的稳定性与预测精度,为建立性能优秀的小麦种子发芽率近红外模型提供了新思路。To improve the detecting accuracy and robustness of wheat seed germination rate with near infrared spectroscopy tech- nique, single PLS model and consensus PLS model（cPLS） developed on the full-spectral were compared and analyzed, thus, the Si-cPLS model which developed on the characteristic spectral regions was put forward. There were 84 samples partitioned into 66 training samples and 18 prediction samples using SPXY method. By randomly selecting 50 samples from training set as calibra- tion set, a series of sub PLS models were build. 100 PLS sub models satisfying predefined criterion were selected and combined one cPLS model by averaging all predicted results. With this basis Si-cPLS model were developed on characteristic spectral re- gions selected with synergy interval method. Statistics on 50 repeat prediction of wheat seed germination by full-spectral PLS model, full-spectral cPLS model, and Si-cPLS model showed that, the mean correlation coefficient（R） were 0. 901, 0. 922 and 0. 936 respectively, the mean RMSEP were 13. 735%, 12. 533% and 10. 273% respectively with standard deviation of RMSEP of 1. 144%, 0. 096% and 0. 080% respectively. Results showed that cPLS model was more stable and reliable than single PLS model, While Si-cPLS could further increase the stability and prediction accuracy of cPLS model.

关 键 词：近红外 特征光谱 多模型共识 发芽率 

分 类 号：O657.3[理学—分析化学]