Search

Elastic net is an improvement of the least-squares method by introducing in L1 and L2 penalties, and it has the advantages of the variable selection. The quantitative analysis model build by Elastic net can improve the prediction accuracy. Using 89 wheat samples as the experiment material, the spectrum principal components of the samples were selected by Elastic net. The analysis model was established for the near-infrared spectrum and the wheat's protein content, and the feasibility of using Elastic net to establish the quantitative analysis model was confirmed. In experiment, the 89 wheat samples were randomly divided into two groups, with 60 samples being the model set and 29 samples being the prediction set. The 60 samples were used to build analysis model to predict the protein contents of the 29 samples, and correlation coefficient (R) of the predicted value and chemistry observed value was 0. 984 9, with the mean relative error being 2.48%. To further investigate the feasibility and stability of the model, the 89 samples were randomly selected five times, with 60 samples to be model set and 29 samples to be prediction set. The five groups of principal components which were selected by Elastic net for building model were basically consistent, and compared with the PCR and PLS method, the model prediction accuracies were all better than PCR and similar with PLS. In view of the fact that Elastic net can realize the variable selection and the model has good prediction, it was shown that Elastic net is suitable method for building chemometrics quantitative analysis model.

In the present paper, taking 66 wheat samples for testing materials, ridge regression technology in near-infrared (NIR) spectroscopy quantitative analysis was researched. The NIR-ridge regression model for determination of protein content was established by NIR spectral data of 44 wheat samples to predict the protein content of the other 22 samples. The average relative error was 0.015 18 between the predictive results and Kjeldahl's values (chemical analysis values). And the predictive results were compared with those values derived through partial least squares (PLS) method, showing that ridge regression method was deserved to be chosen for NIR spectroscopy quantitative analysis. Furthermore, in order to reduce the disturbance to predictive capacity of the quantitative analysis model resulting from irrelevant information, one effective way is to screen the wavelength information. In order to select the spectral information with more content information and stronger relativity with the composition or the nature of the samples to improve the model's predictive accuracy, ridge regression was used to select wavelength information in this paper. The NIR-ridge regression model was established with the spectral information at 4 wavelength points, which were selected from 1 297 wavelength points, to predict the protein content of the 22 samples. The average relative error was 0.013 7 and the correlation coefficient reached 0.981 7 between the predictive results and Kjeldahl's values. The results showed that ridge regression was able to screen the essential wavelength information from a large amount of spectral information. It not only can simplify the model and effectively reduce the disturbance resulting from collinearity information, but also has practical significance for designing special NIR analysis instrument for analyzing specific component in some special samples.

The present paper introduces the principle of a new modeling method, called supervised principal component regression, with which the model of the near-infrared (NIR) spectroscopy quantitative analysis was established. Usually, there are many difficulties such as collinearity when establishing the quantitative analysis model for the high dimension of the spectral data. Using this new method, firstly according to some criterion, the wavelength information is selected in order to reduce the dimension of spectral data. Then the selected lower dimensional spectral data set is used to establish the principal component regression model. Taking sixty-six wheat samples as experiment materials, forty samples were chosen randomly to establish the predicting model, while the remaining twenty-sixth wheat samples were viewed as prediction set. In this example, 4 wavelengths, 4 632, 4 636, 5 994 and 5 997 cm(-1), were selected at first according to the coefficients between the response variable and each spectral data. Then two principal components of the spectral data at those four wavelengths were extracted to establish the principal component regression model. The model was used to the prediction set. The coefficient was 0.991 and the average relative error was 1.5% between the model predication results and Kjeldahl's value for the protein content. It is very important to select the most significant part of wavelength information from a large number of spectral data, not only because this procedure can alleviate the influence of collinearity in modeling, but also because it can be used to guide the design of special NIR analysis instrument for analyzing specific component in some samples.