Non-destructive estimation of mandarin orange fruit quality during the ripening stage using machine-learning-based spectroscopic techniques

This study endeavors to employ non-destructive, machine learning-based techniques to predict and estimate the key quality parameters such as moisture content, total soluble solids, sugar content and pH throughout the ripening stage (unripe, ripen and over-ripen) of Mandarin orange fruit. Regression models compatible with Tiny-machine learning (TinyML) were used to track fruit development stages, crucially identifying the onset of spoilage from the unripe to ripe stages until the fruit becomes overripe. Additionally, Visible–Near–Infrared (VisNIR) spectral sensors were used to capture internal physicochemical attributes, facilitating precise predictions. These models, were trained on the Edge Impulse Platform and implemented on ESP8266 NodeMCU CP2102 Board microcontroller units. The optimal neural network architecture, comprising 18 input nodes representing spectral sensor data, two hidden layers with 20 and 10 nodes, and an output layer predicting ripening stage, achieves accuracy with R² values of 0.9912 for ripening stage, 0.8164 for pH, 0.9657 for total soluble solids (TSS), 0.9956 for sugar content (SC), and 0.9882 for moisture content (MC). Furthermore, by utilizing models for accurately predicting fruit quality parameters and estimating ripening stages, this approach not only aids in optimizing supply chain management by scheduling fruit consumption at the optimal time but also ensures consumers benefit from the nutritional advantages of mandarin oranges while minimizing economic losses due to spoilage.