Optimization of machine learning models for predicting glutinous rice quality stored under various conditions

Abstract

The preservation of freshly harvested glutinous rice (GR) is essential for maintaining its nutritional and economic value. This study examines the impact of storage temperature and duration on key quality attributes, including moisture content (MC), germination growth rate (GGR), water absorption capacity (WAC) and head rice yield (HRY). GR samples were dried at 60 °C and stored under freeze (−10 °C), cold (6 °C), and ambient (∼26 °C) conditions for six months, with biweekly data collection. Statistical analysis using ANOVA revealed that storage duration significantly affected MC, GGR and HRY, while storage temperature primarily influenced MC. The Random Forest (RF) machine learning model demonstrated high predictive performance (R² > 0.9) with low error values for predicting quality attributes. Hyperparameter tuning (HPT) through grid search optimization further improved the model's performance, as validated by parity plots showing strong alignment (regression slopes >0.8) between predicted and experimental results. SHapley Additive exPlanations (SHAP) and contour plots provided detailed insights into the influence of storage parameters on quality attributes. This comprehensive approach offers actionable guidance for optimizing GR storage conditions, contributing to food security, and supporting to the achievement of the United Nations Sustainable Development Goals (SDGs).