Freshness Grading Prediction of Berries Based on Volatile Organic Compounds and Machine Learning: A Comparative Study of Kiwifruit and Grapes

To enable early, non-destructive freshness grading of berries under cold-chain conditions, this study investigated Kyoho grapes (non-climacteric) and Xuxiang kiwifruit (climacteric) stored at 5°C for 15 days. Firmness, weight loss, respiration rate, and color parameters were measured on days 0, 3, 6, 9, 12, and 15. In parallel, eight volatile organic compounds (VOCs) were quantified by GC–MS using an internal standard method for quality evolution characterization and model development. Based on the integrated changes in multiple quality indicators, a five-level freshness grading scheme was established, and four machine-learning classifiers were developed using the concentrations of the eight VOCs. The results showed that pronounced changes in characteristic VOCs occurred approximately 3–6 days earlier than the observable declines in firmness and darkening in color. On the independent test set, SVM achieved the best performance for grapes (accuracy: 93.00%), whereas RF performed best for kiwifruit (accuracy: 85.36%). This study proposes a grading strategy of “key VOC fingerprints + an algorithm tailored to the fruit type,” demonstrating its potential for non-destructive, early freshness grading and cold-chain quality early warning for postharvest berries.