• Standalone handheld acoustic device enables real-time in-field durian grading. • On-device acquisition and ML inference deliver <500 ms end-to-end latency. • Mixed-domain DT achieves 97.77% accuracy and 96.90% precision across years. • Robust acoustic features resist tapping force variation and orchard noise. Reliable in-field assessment of durian maturity remains challenging due to the subjectivity and limited practicality of conventional methods. This study presents a low-cost, handheld acoustic sensing device that performs real-time, on-device inference for on-tree durian maturity classification under orchard conditions. The system integrates acoustic acquisition, feature extraction, and machine learning inference entirely on-device, enabling immediate decision feedback without external computing. Field experiments were conducted on 142 Ri 6 durian fruits collected from two orchards over two harvest years, yielding 1,026 valid in-field acoustic recordings obtained via tapping-based excitation. Twenty-eight acoustic features, including peak frequency, spectral centroid, and Mel-frequency cepstral coefficients, were used as inputs to multiple machine learning models for classification. The proposed device achieved accurate acoustic recording and demonstrated strong robustness to environmental noise through a dedicated sound insulation design. Mixed-domain training improved robustness to inter-annual variability, with the decision tree model achieving the best performance (97.77% accuracy and 96.90% precision). Inference time remained below 500 ms for all models. These results demonstrate that the proposed device offers a practical, resource-efficient solution for real-time durian maturity assessment, with a low hardware cost (approximately USD 55) and strong potential for large-scale field deployment and extension to other fruits with similar structural characteristics, such as jackfruit and other acoustic-based sensing applications.
Nguyen et al. (Wed,) studied this question.