Non-destructive Detection and Classification of Postharvest Latent Damage in Korla Fragrant Pears Based on Multi-frequency Electrical Properties

Latent damage in Korla fragrant pears critically affects fruit quality and postharvest management, making accurate severity grading essential. This study developed a nondestructive grading method by collecting multi-frequency electrical parameters from pears with different injury levels. StepReg, Lasso, and Elastic Net were used for feature selection and combined with SIMCA, SVM, RF, and XGBoost to build grading models. Storage experiments further examined the relationship between predicted injury grades and storability. Results showed that multi-frequency electrical parameters combined with machine learning could effectively grade latent damage. Among all models, Enet-SVM achieved the best performance, with macro-F1 and accuracy of 0.9824 and 0.9825, respectively. Model-predicted grades were consistent with subsequent storage behavior: median storage lives for mild, moderate, and severe damage were 73, 48, and 29 days, respectively. Higher injury grades showed faster decreases in survival probability and greater decay risk, supporting postharvest decisions such as short-term sale, early diversion, processing, or culling. SHAP analysis indicated that multi-frequency capacitance, phase angle, quality factor, and equivalent resistance were key contributors. This study offers an efficient, interpretable approach for nondestructive graded detection and precise postharvest sorting of Korla fragrant pears.