In view of the uncertainty of sugar heart occurrence in Zhaotong apples and the varying effects of different sugar heart degrees on fruit flavor and storage life, this study aimed to develop a rapid and non-destructive method for accurately sorting apples with or without different degrees of sugar heart, thereby providing technical support for enhancing the commercial value of fruits and optimizing their storage and grading. This study employed a laboratory-developed online fruit sorting system to acquire visible/near infrared (VIS/NIR) spectral signals of Zhaotong sugar-heart apples. Feature wavelengths identified by the variable importance in projection (VIP) method served as inputs for establishing a stacking ensemble model integrating random forest (RF), extreme gradient boosting (XGBoost), logistic regression (LR), and support vector machine (SVM) algorithms. The performance of the stacking model was compared against that of standalone models. The results showed that the stacking model integrating four distinct machine learning algorithms exhibited superior recognition performance with an accuracy of 95.47%, a true positive rate (TPR) of 94.82%, and a true negative rate (TNR) of 97.83% compared with standalone models. The proposed stacking ensemble approach significantly enhances predictive capability by combining the strengths of base models without substantially increasing the computational load, showing great potential for applications in online fruit sorting.

A Crown pear is one of the most primary fresh fruit varieties worldwide, due to the high juice content, sweetness, and rich nutritional value. Among them, the firmness can represent one of the most important quality indicators for pears. It can directly present how ripe the pear is and how well the pear can be stored after harvest. For these reasons, it is often required to accurately measure the firmness of the pear. The conventional measurement of fruit firmness can depend mainly on the destructive tests, including the Magness-Taylor puncture and compression. However, the large amount of food waste cannot fully meet the large-scale testing in recent years. Also, the current studies still share the two challenges, whether the regular prediction has low accuracy or the deep learning is difficult to interpret. This research aims to examine the connection between fruit firmness and vibration frequency in Crown pears. A recognition model was also established using a standard convolutional neural network (CNN). After that, an improved CNN was combined with Deep Shapley Additive Explanations. A series of experiments was carried out to verify the improved model. The 508 Crown pears were taken as the test samples. Five groups were divided for the samples, each of which was tested every thirteen days. Vibration data was gathered from the Crown pears using a Laser Doppler Vibrometer (LDV). An improved one-dimensional CNN was applied to construct a firmness prediction model, according to the characteristic frequencies. Deep Shapley Additive Explanations (Deep SHAP) structure was used to explain the function of the prediction model. Furthermore, a comparison was made of several standard prediction models, including the partial least squares regression, support vector regression, extreme gradient boosting, and adaptive gradient boosting. The results show that the improved one-dimensional CNN model was achieved in the high-precision prediction of the Crown pear hardness. Among them, the feature frequency of the R_P² was 0.945, the RMSE_P was 0.594 N/mm, and the RPD_P was 4.272. The performance also outperformed all conventional models over all metrics. The most outstanding performance was achieved, where the R² value was closest to the ideal value of 1, the RMSE was the lowest, the next-best was XGBoost model, and the RPD value was significantly higher than the rest of the models, indicating the superior generalization. The characteristic frequencies were utilized to predict the Crown pear firmness with high accuracy. Vibration frequency features between 300 and 700 Hz shared a strong relationship with the pear firmness. An interpretable deep learning model with the LDV vibration data was achieved in the high-precision prediction for interpretability and practicality. There was a strong correlation between characteristic frequency and firmness. The finding can provide an efficient and reliable non-destructive testing of the Crown pear firmness in the assessment of the fruit quality.