Near-Infrared spectroscopy combined with machine learning methods for distinguishment of the storage years of rice

• Published in: Infrared physics & technology Vol. 133; p. 104835
• Main Authors: Huang, Fuping, Peng, Yimei, Li, Linghui, Ye, Shitong, Hong, Shaoyong
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-09-2023
• Subjects: Least square support vector machine; Machine learning; Near-infrared spectroscopy; Principal component neural network; Random forest; Rice; Storage time; Least square support vector machine; Storage time; Machine learning; Random forest; Principal component neural network; Near-infrared spectroscopy; Rice
• ISSN: 1350-4495; 1879-0275
• DOI: 10.1016/j.infrared.2023.104835

[Display omitted]
•Near-infrared model is established for distinguishment of the storage year of rice.•LSSVM, RF and PC-NN methods are studied for adaptive training of the NIR models.•Grid search technique is designed on the regularization parameter for LSSVM model.•Tree estimators and their split depth are tuned for the random forest model.•Principal components are extracted to refine the input of the neural network model.
Rice is one of the most important food crops that provide essential nutrients, micronutrients and daily energy for humans. The freshness of rice determines the quality and nutrition supply property, but the freshness decreases along with the storage time. A simple, nondestructive and rapid detection technology is needed to estimate the time of storage rice as for a fast evaluation of the rice quality. To accomplish this objective, near-infrared spectroscopy (NIRS) is employed in combination with three machine learning methods, including least square support vector machine (LSSVM), random forest (RF) and principal component-neural network (PC-NN). With specific design on grid search of the relevant parameters, the LSSVM model optimally performed classification with the highest accuracy of 95.7% in the distinguishment of three labeled storage years, the RF model and PC-NN models have close accuracies in model training and optimization processes. In comparison to the PLS method, which is the typical chemometric method in NIRS data analysis, the three presented machine learning methods all perform excellent over the PLS model for model training and for model testing. Especially the RF and PC-NN model were optimized by hyperparameter training, to obtain 90% of testing accuracy and reduced the error differences to ∼5.0% between model training and testing. This study indicated the potential of NIRS in combination with machine learning methods as practical chemometric tools for discrimination of the rice storage freshness by distinguishing their storage years. The design of adaptive tuning on hyperparameters provide a valuable approach to improve the model prediction abilities.