Detection of Early Damage in Kiwifruit Based on Near-Infrared Technology

Abstract

The internal quality of kiwifruit directly affects its taste. During harvesting or transportation, kiwifruit sustained surface invisible damage due to collisions or pressure. To conduct non-destructive detection of minor mechanical damage in kiwifruit, this study investigated two widely cultivated varieties in China. Near-infrared spectroscopy was employed to collect spectral data from both intact samples and early-damaged samples. These datasets were utilized to develop classification models aimed at assessing the extent of damage in kiwifruit. Initially, the first derivative method was applied as a spectral preprocessing technique. Three feature selection methods—Competitive Adaptive Reweighted Sampling (CARS), Genetic Algorithm (GA), and Bootstrap Soft Shrinkage (BOSS)—were implemented to extract characteristic wavelengths from the preprocessed spectra. Subsequently, classification models were constructed based on both the selected feature spectra and the original spectra. A novel Stacking ensemble model was developed using Support Vector Machine (SVM), Extreme Learning Machine (ELM), and Extreme Gradient Boosting (XGBoost) as first-level classifiers, with Logistic Regression serving as the second-level classifier. By establishing training and testing datasets while comparing performance metrics against those of individual first-level classifiers, the study evaluated the model's efficacy. The results indicated that the Stacking model consistently demonstrated high accuracy across all feature selection algorithms; notably, when combined with CARS feature selection, it achieved accuracy rates of 100% and 98.60% on training and testing sets, respectively, underscoring its superior performance. This suggested that integrating the Stacking model with CARS provided optimal predictive capabilities for this dataset. In conclusion, employing near-infrared spectroscopy for classifying varying degrees of damage in kiwifruit was not only feasible but also offered a robust reference point for evaluating market-related damage levels.