評估以不同統計方法建立白米粉末直鏈澱粉含量的近紅外線光譜模型

Evaluation of Building the Models of Amylose Content in Rice Powder from Near-Infrared Spectrum by Different Statistical Methods

蘇堃綺(Kun-Chi Su) ； 鄭心婷(Xin-Ting Zheng) ； 洪梅珠(Mei-Chu Hong) ； 郭寶錚(Bo-Jein Kuo)

《農林學報》 50卷2期 (2001/06) Pp. 33-52

https://doi.org/10.30089/JAF.200106.0004

摘要

本研究係分別以MLR，PCR與PLS三種統計法，建立自米粉末直鏈澱粉含量的NIRS模型。由研究結果發現，MLR法並非一個十分穩定的建模方法，無法解決共線性問題所帶來的困擾；PCR法雖可解決此一問題，但PLS又可以較PCR法為少的成分數，提升模型的建模與預測效能。MLR法的分析中，是選擇以8個波長所建立的模型為最適模型，但此模型預測能力表現不佳，且IRV值高達15509.2，顯然逢機變異過大，以致於模型穩定度不足。PCR法乃採用前19個最重要的主成分來建立最適模型，此模型的逢機變異大幅下降，預測的能力也獲得提升。至於以PLS法分析的結果，則是選擇前13個最具解釋能力的成分來建立最適模型，其有關建模效能的指標R(上標 cal)、SEC、ISV及IRV分別為：0.963、1.705%、19.06與6489.4，而有關預測效能的指標R(下標 val)、SEP、Bias、Skew及RPD比值則分別為0.953、1.947%、-0.333%、0.950與3.225，除了各項指標均表現十分良好外，IRV值也顯示模型已將共線性問題排除，因此模型顯得相當穩定，同時PLS法亦更進步提升模型的建模能力與預測能力，明顯於三種統計模型中有最好的表現。

關鍵字

近紅外線光譜 ； 多元共線性 ； 逐步回歸 ； 主成分回歸 ； 淨最小平方法 ； 直鏈澱粉含量

並列摘要

In our study, we used three different statistical methods, MLR, PCR, and PLS. to build mi1led-rice amylase model of NIRS, respectively. But it has been found that MLR is not a stable method to build a suitable model. It cannot solve the problems coming from multicollinearity. Although PCR resolved the problem, PLS raised the model-building and prediction ability further with lesser components. By the results, 8 wavelengths were used to build the best MLR model. It seems that its prediction ability was not sufficient. Furthermore, the IRV was as high as 15509.2. Obviously, the random variation was 50 extremely large that the model came into a very unstable state. We also used the 19 foremost principal components to build the best PCR model. The random variation declined a lot, and the prediction ability also promoted. Final1y, we selected the model with 13 components of best explanatory ability as the most suitable PLS model. For this model, the indexes of model-building ability, such as R(subscript cal), SEC, ISV and IRV, were 0.963, 1.705%, 19.06 and 6489.4, respectively. The indexes of prediction ability, such as R(subscript val), SEP, Bias, Skew and RPD ratio, were 0.953, 1.947%， -0.333%, 0.950 and 3.225, respectively. All indexes showed that the selected PLS model performed quite well. As the value of IRV, it implied that the multicollinearity problem had been resolved. Hence the model had become very stable. Besides, the model-building and prediction ability of this model was progressed further at the same time. Evidently, the PLS method performed as the best in the three statistical models.

並列關鍵字

near-infrared reflectance spectroscopy ； multicollinearity ； stepwise regression ； principal component regression ； partial least-squares regression ； amylose content