Detection of wheat saccharification power and protein content using stacked models integrated with hyperspectral imaging

• Published in: Journal of the science of food and agriculture 2024-05, Vol.104 (7), p.4145-4156
• Main Authors: Huang, Yuexiang, Tian, Jianping, Yang, Haili, Hu, Xinjun, Han, Lipeng, Fei, Xue, He, Kangling, Liang, Yan, Xie, Liangliang, Huang, Dan, Zhang, HengJing
• Format: Article
• Language: English
• Subjects: Accuracy; Adaptive sampling; Algorithms; Artificial neural networks; Back propagation networks; Ensemble learning; Feature extraction; Hyperspectral Imaging; Least-Squares Analysis; Liquor; Machine learning; Neural networks; Neural Networks, Computer; Nutrient content; Prediction models; Preprocessing; protein content; Proteins; Raw materials; Saccharification; saccharification power; stacked model; Support Vector Machine; Support vector machines; Test sets; Triticum; Wavelengths; Wheat
• ISSN: 0022-5142; 1097-0010
• DOI: 10.1002/jsfa.13296

BACKGROUND Wheat is one of the key ingredients used to make Chinese liquor, and its saccharification power and protein content directly affect the quality of the liquor. In pursuit of a non‐destructive assessment of wheat components and the optimization of raw material proportions in liquor, this study introduces a precise predictive model that integrates hyperspectral imaging (HSI) with stacked ensemble learning (SEL). RESULTS This study extracted hyperspectral information from 14 different varieties of wheat and employed various algorithms for preprocessing. It was observed that multiplicative scatter correction (MSC) emerged as the most effective spectral preprocessing method. The feature wavelengths were extracted from the preprocessed spectral data using three different feature extraction methods. Then, single models (support vector machine (SVM), backpropagation neural network (BPNN), random forest (RF), and gradient boosting tree (XGBoost)) and a SEL model were developed to compare the prediction accuracies of the SEL model and the single models based on the full‐band spectral data and the characteristic wavelengths. The findings indicate that the MSC–competitive adaptive reweighted sampling–SEL model demonstrated the highest prediction accuracy, with Rp2 (test set‐determined coefficient) values of 0.9308 and 0.9939 for predicting the saccharification power and protein content and root mean square error of the test set values of 0.0081 U and 0.0116 g kg−1, respectively. CONCLUSION The predictive model established in this study, integrating HSI and SEL models, accurately detected wheat saccharification power and protein content. This validation underscores the practical potential of the SEL model and holds significant importance for non‐destructive component analysis of raw materials used in liquor. © 2024 Society of Chemical Industry.