利用淨最小平方迴歸、主成分迴歸及類神經網路模式於近紅外光譜資料以預測糙米中粗蛋白質含量之比較

Comparison of the Partial Least Squares Regression, Principal Component Regression, and Neural Networks Models for Predicting the Crude Protein Content of Brown Rice on the Near-infrared Spectral Data

徐永衡(Yung-Heng Hsu) ； 林汶鑫(Wen-Shin Lin) ； 何兆銓(Siu-Chuen Ho) ； 郭寶錚(Bo-Jein Kuo)

《作物、環境與生物資訊》 4卷1期 (2007/03) Pp. 41-48

https://doi.org/10.30061/CEB.200703.0005

摘要

近年來利用近紅外光分析儀來測定物質中的各種化學成分含量，有愈來愈普遍的趨勢，而分析這類的資料，需要藉助多變數檢量的技術。本研究利用糙米中粗蛋白質含量的近紅外線光譜資料，以主成分迴歸(principal component regression, PCR)、淨最小平方迴歸(partial least squares regression, PLSR)以及類神經網路(neural networks, NN)來建立檢量模式，並比較這些方法在預測糙米中粗蛋白質含量上之優劣性。本研究中NN模式的建立，除了以原始351個全波段當作輸入變數外，將經主成分分析(principal component analysis, PCA)後所得之主成分(principal component, PC)及經淨最小平方法(partial least squares, PLS)後所得的得點t，當作輸入變數分別來建立PC-NN與PLS-NN兩種模式。最後得到的結果以PLSR所建立的模式預測能力表現最佳，PCR與NN模式次之，PLS-NN模式最差；其中NN的建模之預測能力以及效率，不如雙線性(bilinear)檢量模式中的PLSR外，另需耗費較長的時間來進行學習訓練。其可能的原因在於給予NN初始的加權值(weight)時，擁有重要訊息的變數得到較小的加權值，較不重要的變數反而得到較大的加權值。因此，如能發展出一個能事先給予重要變數較大初始加權值的準則，應可提高NN網路學習的能力及效率。

關鍵字

類神經網路 ； 主成分迴歸 ； 淨最小平方迴歸 ； 雙線性

並列摘要

Near-infrared reflectance spectroscopy (NIRS) has been widely used in quantitative applications of chemometrics in recent years. However, the analysis of NIR spectral data needs to use the technique of multivariate calibration. In this study, calibration methods including principal component regression (PCR), partial least squares regression (PLSR), and neural networks (NN) were employed in conjunction with the near-infrared reflectance spectroscopy technique to determine the raw protein content of brown rice. Furthermore, the performance of these calibration methods was compared. Except the reflectance values of the 351 wavebands of full spectral range, scores of the principal component analysis and partial least squares were also used as the input variables for neural networks (i.e., PC-NN and PLS-NN). It was found that performance of the PLSR model in the predictive ability was better than that of the PCR and NN models. The PLS-NN model did not perform well. It was also shown that performance of the NN model was worse than that of the PLSR model in the predictive ability and efficiency. Additionally, the NN model needed more learning time. It was quite likely that the initial weights of the important variables between input and hidden layer in NN model were given smaller values, while the initial weights of the unimportant variables were given larger values. If the criteria of offering a larger initial weight to the important variable could be exploited in advance, the learning ability and efficiency might be improved.

並列關鍵字

Neural networks ； Principal component regression ； Partial least squares regression ； Bilinear