A Non-Destructive Acoustic Tapping System with Deep Learning for Pineapple Juiciness Classification in Postharvest Quality Assessment

Non-destructive fruit quality assessment is essential for improving consistency, scalability, and objectivity of postharvest decision-making. This study presents a deep learning-assisted acoustic tapping system for classifying pineapple juiciness using recorded tapping signals. A total of 300 Pattavia pineapples were evaluated, producing 3600 audio recordings and 6840 curated tapping-sound samples evenly distributed across three juiciness classes. The proposed workflow integrates mobile-phone-based sound acquisition, tap-event segmentation, acoustic feature extraction, and supervised classification. Three audio representations Mel-Frequency Cepstral Coefficients (MFCCs), YAMNet embeddings, and VGGish embeddings were evaluated with deep learning, conventional machine learning, and ensemble models, including CNN, LSTM with attention, GRU with attention, hybrid CNN–LSTM–attention, random forest, logistic regression, gradient boosting, multilayer perceptron, voting, and stacking classifiers. The results showed that MFCC-based deep learning models provided the most reliable classification performance. The best-performing CNN achieved an accuracy of 0.9415, F1-score of 0.9411, Cohen’s kappa of 0.9123, and AUC of 0.9923. Additional analyses using confusion matrices, ROC curves, McNemar’s tests, bootstrap confidence intervals, and t-SNE visualization confirmed the robustness and discriminative capability of the proposed approach. These findings demonstrate that acoustic tapping combined with deep learning offers a practical, low-cost, and non-destructive method for pineapple juiciness classification, with potential application in postharvest sorting, quality control, and precision agriculture.