Abstract
INTRODUCTION: Accurately determining the moisture content of cigar leaves during the air-curing process is crucial for quality preservation. Traditional measurement techniques are often subjective and destructive, limiting their practical application. METHODS: In this study, we propose a stacking ensemble learning model for non-destructive moisture prediction, leveraging image-based analysis of naturally suspended cigar leaves. In this study, front and rear surface images of cigar leaves were collected throughout the air-curing process. Color and texture features were extracted from these images, and a filtering method was applied to remove redundant variables. To ensure optimal model selection, the entropy weight method was employed to comprehensively evaluate candidate machine learning models, leading to the construction of a stacking ensemble model. Furthermore, we applied the SHAP method to quantify the contribution of each input feature to the prediction results. RESULTS: The stacking ensemble model, comprising MLP, RF, and GBDT as base learners and LR as the meta-learner, achieved superior prediction accuracy (R (2) (test) =0.989) and outperforms than traditional machine learning models (R (2) (test) ranged from 0.961 to 0.982). SHAP analysis revealed that front surface features (45.5%) and leaf features (38.5%) were the most influential predictors, with airing period (AP), a (f) (*), G (f), and ASM (f) identified as key predictors. CONCLUSION: This study provides a feasible and scalable solution for real-time and non-destructive monitoring of cigar leaf moisture content, offering effective technical support for similar agricultural and food drying applications.