Crispness evaluation of Korla pear based on explainable machine learning: Instrumental mechanical-acoustic synchronously at high sampling rate.

Crispness is a critical indicator for fruit texture evaluation, which is directly associated with fruit freshness and consumer preference. In this study, a new instrumental method was applied to synchronously collect mechanical-acoustic signals at high sampling rate of 51,200 Hz for "noisy" Korla pear in puncture. The mechanical-acoustic jagged analytic spectral were fused at the data-level to imitate human perception behavior. Three types of features totaling 26 were extracted from the fusion signals to establish four machine learning models. Among four models, the extreme gradient boosting (XGBoost) model exhibited the superior performance in crispness prediction. Furthermore, the Shapley additive explanations (SHAP) analysis was performed to interpret the XGBoost model. The 14 features with the positive impact outcome were then selected to improve the model. The explainable model achieved an value of 0.92, an of 0.32, and an of 3.66 with a higher accuracy, stability and reliability. Hence, the use of synchronous acquisition at high sampling rate, data-level fusion strategy and positive features selection can significantly enhance the crispness prediction performance of pear. Our proposed method can be applicable to other fruit and vegetables for instrumental crispness measurement.