摘要
烤烟烟叶厚度是烤烟烟叶分级和品质评价的重要指标之一,采用近红外光谱分析技术实现对烟叶厚度的快速测定具有一定的可行性。对比分析了采用竞争自适应重加权法(CARS)算法优选特征变量和采用全波长变量(1 000~2 500 nm)结合偏最小二乘法(PLS)建立初烤烟烟叶厚度近红外校正模型的效果。结果表明:模型的输入变量数由1 543个降低到180个,决定系数由0.846提高到0.941;适宜主成分数由10降低到6,校正标准误差和交互验证均方根误差分别降低了0.003 4和0.010 3。采用30个外部样品对模型进一步进行验证,模型的验证标准误差和验证标准误差的偏差由0.018 2降低到0.001 1,在α=0.05显著水平,两个模型预测值与实测值间均不存在显著差异,采用CARS筛选特征变量近红外模型预测值与实测值间的差异性更小。CARS筛选特征变量提高了烟叶厚度近红外校正模型的稳定性和预测准确性。
The thickness of tobacco is an important factor to measure the quality and grading of flue- cured tobacco. The present study shows that it is feasibility to prediction thickness of tobacco by fourier transform near-infrared （FT-NIR） spectrometry. Two methods of screening spectrum range that calcu- lated by competitive adaptive reweighted sampling method （CARS） and directly use the spectrum range （ 1 000 - 2 500 nm） had been combined with partial least squares （PLS） to establish the FT- NIR model. The results show that the number of variables of the thickness of tobacco FT-NIR model also had reduced to 180 from 1 543. The number of principal components also had reduced 6 from 10. The lower of standard error of calibration （SEC） and root mean square error of cross validation （ RM- SECV） was 0. 003 4 and 0. 010 3, respectively. Thirty samples had been used as external validation. The standard error of validation （SEV） and standard deviation of validation error （SDV） reduced to 0. 001 1 from 0. 018 2. According NIR model of paired-t tests of measured and prediction value, the results have no significant variance at level of 0. 05. The CARS-NIR model have smaller variance. The stability and forecasting accuracy of Fr-NIR model had been developed improved by the CARS method to select the sensitive wavelengths.
出处
《云南农业大学学报（自然科学版）》
CSCD
北大核心
2017年第2期288-293,共6页
Journal of Yunnan Agricultural University:Natural Science
基金
云南中烟工业有限责任公司科技项目(HYHH2013YL06)
关键词
初烤烟
厚度
近红外光谱法
模型集群分析
竞争自适应重加权法
flue-cured tobacco
thickness
fourier transform near-infrared （FT-NIR） spectrometry
model population analysis （MPA）
competitive adaptive reweighted sampling （CARS）
