Non-Destructive Prediction of Maturity from the Sound of Hand Hitting a Watermelon Using Machine Learning

Traditional methods for assessing the quality, taste, and ripeness of fruits and vegetables without cutting rely on attributes such as color, shape, surface patterns, and acoustic responses. The ripeness levels were verified by cutting the watermelons, and the corresponding sound data were examined using spectrogram analysis, extracting 120 features from each sample. Various machine learning algorithms, including Support Vector Classifier (SVC), Decision Trees (DTC), Random Forest Classifier (RFC), Multi-Layer Perceptron (MLP), and K-Nearest Neighbors Classifier (KNC), were applied to identify the most effective predictive model. The results indicate that the KNC model achieved the highest accuracy at 96.04%, followed by the RFC model with an accuracy of 95.47%. The RFC model classified ripe watermelons with 98.2% accuracy, while the KNC model distinguished overripe and underripe watermelons with accuracies of 96.3% and 96.2%, respectively. Despite the presence of background noise in the naturally recorded dataset, the system demonstrated high performance across all categories. The findings were compared with studies on acoustic pattern recognition in animals, environmental acoustic analysis, and healthcare applications. This study highlights that machine learning-based models provide a non-invasive approach to determining watermelon taste and ripeness, offering a practical solution for applications in the agriculture and food industries.