Nondestructive Maturity Measurement of Durian Using Acoustic Signal

This study introduces a non-destructive method for assessing the ripeness of the Vietnamese “Dona” durian using acoustic knocking sounds and machine learning. We collected audio data by recording knocking sounds from durians, followed by segmenting the raw audio into specific duration clips to isolate relevant acoustic features. The Short-Time Fourier Transform (STFT) was then applied for feature extraction and by utilizing a MobileNetV2 model pre-trained on ImageNet, we achieved a validation accuracy of $\mathbf{9 1. 7 2 \%}$. The model demonstrates effective classification capabilities, though some misclassifications of ripe durians occurred. These inaccuracies stem from the limited size of the dataset, which may not fully capture the variations in acoustic characteristics of different types. In order to facilitate practical use, we implemented the model in an application that allows users to record or upload knocking sounds for real-time ripeness prediction. This study paves the way for broader adoption of machine learning in agriculture, as future work could focus on refining data collection processes, expanding the dataset, and enhancing the model to detect diseased durians by analyzing additional indicators.