An Improved Ensemble Learning Method for Protein Content Analysis of Corn with Small Sample by Near-Infrared Spectroscopy

Abstract

Near-infrared spectroscopy has become an important methodology for rapid and non-destructive detection in food and agricultural fields. However, the accuracy of quantitative analysis was seriously restricted by the severe overlap of spectra and the high cost of standard samples. In order to reduce the impact of these problems especially that of small sample size problem, a novel method named weighted clustering ensemble partial least squares (WCE-PLS) is proposed for the protein content analysis of corn. Firstly, the clustering and sampling strategy is introduced in the calibration sets of corn to create different subsets for generating sub-models. Then, root mean square errors of cross-validation (RMSECV) in those sub-models as the crucial criterion are computed for model optimization. Finally, in integrating step, two Gaussian weighted functions used to determine the weights of sub-models are defined. The validation performance of the proposed method is tested with the near infrared spectral data sets of corn and compared with single PLS, bagging PLS, boosting PLS, and data augmentation (DA) PLS. To further demonstrate the effectiveness of the method, another data set of soil was used for supplementary verification. Results of the prediction sets indicated that the RMSEP values of the WCE-PLS are obviously smaller than that of boosting PLS. And the RMSEP of WCE-PLS and bagging PLS is relatively small in most cases. Furthermore, the correlation coefficients between predicted value and chemical value are respectively 0.96587 and 0.90849 for two data sets, which computed by the WCE-PLS is obviously higher than that computed by the other four methods. And the t test also showed the WCE-PLS has smaller t values and larger p values.