Explainable machine learning-based prediction of shelf-life and anthocyanin accumulation in table grapes

Area of Science:

• Agricultural Science
• Data Science
• Food Science

Background:

• Table grapes (Vitis vinifera L.) are a globally popular fruit but are highly perishable.
• Accurate prediction of shelf-life and nutritional content is vital for the fresh produce supply chain.

Purpose of the Study:

• To develop intelligent machine learning models for predicting table grape shelf-life and total anthocyanin content (TAC).
• To enhance the interpretability of these models using SHapley Additive exPlanations (SHAP).

Main Methods:

• Utilized artificial neural networks (ANN) for shelf-life prediction.
• Employed extreme gradient boosting (XGBoost) for TAC prediction.
• Applied SHAP-based sensitivity analysis to identify critical quality-influencing factors.

Main Results:

• ANN achieved high accuracy in shelf-life estimation (MAE 1.624 days, R² 0.9873).
• XGBoost effectively predicted TAC (MAE 0.0113 mg C3G g⁻¹, R² 0.9355).
• Berry abscission rate, weight loss rate, and postharvest treatments were identified as key factors.

Conclusions:

• Developed a reliable machine learning approach for monitoring table grape quality.
• Provided insights into reducing postharvest losses in fresh agricultural produce.
• Demonstrated the utility of SHAP for improving model interpretability in agricultural applications.