Non-destructive quality classification of rice taste properties based on near-infrared spectroscopy and machine learning algorithms

Abstract

The taste quality of rice is determined by protein and amylose percentages, with low levels indicating high-quality taste in Japan. However, accurate non-destructive screening remains a challenge for the industry. We explored the use of machine learning models and near-infrared spectra to classify rice taste quality. Three models were optimized using 796 brown rice samples from Hokkaido, Japan, produced between 2008 and 2016, and tested on 278 distinct samples from the same region produced between 2017 and 2019. Logistic regression and support vector machine models outperformed the partial least-squares discriminant analysis model, achieving high accuracy (94%), f1-score (90%), average precision (0.94), and low classification error (4%) and allowing accurate non-destructive classification of rice quality. These results not only improve rice quality, post-harvest technology, and producer output in Japan but also could enhance quality control processes and foster the production of high-quality products for other agricultural goods and food commodities worldwide.

Keywords: Logistic regression; Near-infrared spectroscopy; Oryza sativa subsp. japonica; Partial least squares discriminant analysis; Rice taste; Support vector machine.