Nondestructive freshness evaluation of Prunus armeniaca L via multi-sensor fusion and explainable machine learning

The rapid postharvest quality decline of ‘Xiaobai’ apricots warrants intelligent and non-destructive freshness evaluation. Herein, we developed a real-time multi-sensor platform based on the Modbus RTU protocol to continuously monitor key environmental parameters at −1°C, 5°C and 25°C. Physiological indicators, including firmness, soluble solid content (SSC), weight loss and colour, were systematically measured. Five machine learning models, namely, the back propagation neural network (BPNN), support vector machine (SVM), random forest (RF), radial basis function neural (RBF) network and extreme learning machine (ELM), were compared with the traditional Arrhenius kinetic model. Low-temperature storage effectively delayed quality loss: firmness declined by only 27.46% after 7 days of storage at -1°C but by 58.98% at 25°C. The machine learning models substantially outperformed the Arrhenius model, with the RF algorithm showing the best performance ( R 2 = 0.9609, RMSE = 0.1331). SHAP interpretability analysis identified O 2 and CO 2 dynamics as the most critical features driving shelf-life predictions, offering novel mechanistic insights into postharvest gas–metabolism interplay. T he proposed integration of multi-sensor data and explainable machine learning provides a robust, non-destructive framework for real-time freshness monitoring and a scalable tool for intelligent quality control in fruit supply chains.