最小二乘支持向量回归与偏最小二乘回归建立烟草总糖NIR预测模型比较

Comparison of Tobacco Total Sugar NIR Prediction Models Developed with Least Square Support Vector Regression and Partial Least Square Regression

采用最小二乘支持向量回归（LSSVR）法和偏最小二乘回归（PISR）法及192个烟叶样品的近红外（NIR）光谱与总糖含量的测定数据，分别建立了烟叶总糖含量的NIR预测模型，并利用这两种模型对95个烟叶样品进行了预测。结果表明：LSSVR法模型的预测误差范围为-3．08％～3．71％，预测回收率范围为90．0％～112．2％。LSSVR法模型的预测准确度比PLSR法的高。

The two prediction models of total sugar in tobacco were developed with least square support vector regression （LSSVR）, partial least square regression （PLSR） based on the total sugar contents in and near infrared spectra （NIRs） of 192 flue-cured tobacco samples. The contents of total sugar in other 95 samples were predicted with the models. It was found that the prediction range of recoveries of LSSVR model were from 90.0% to 112.2% and the prediction errors were in the range of - 3.08% to 3.71 %, while those of PLSR model were from 89.8 % to 113.6 % and from - 4.75 % to 3.85 %, respectively. LSSVR model was superior to PLSR model in prediction accuracy.