Portable Vis/NIR Spectroscopy Combined with Optimized SVM for Grade Classification of Apple Watercore

Abstract

To achieve rapid and non-destructive detection of watercore in apples, the apple Vis–NIR spectral data were used as the research object in this study. A total of 256 × 2 wavelength-based spectral features were extracted, followed by dimensionality reduction via PCA. Subsequently, four machine learning classification models—PLS-DA, SVM, Logistic Regression, and Random Forest—were constructed to identify apples with varying severities of watercore. The results showed that the SVM model achieved the optimal comprehensive performance, with an accuracy of 85.30% and a precision of 91.67% on 34 independent test samples. The five-fold cross-validation accuracy of SVM reached 87.20%, indicating the best generalization stability among the four models. PCA analysis showd that after dimensionality reduction of the core wavelength bands obtained via feature screening, eight principal components can cumulatively explain 95.1% of the total spectral variance; normal apples and watercore-affected apples exhibited significant distribution differences in the principal component space. This study adopted a preprocessing scheme that combined feature selection with SMOTE sample balancing, which effectively alleviated the impact of class imbalance on model performance. It also indicated that insufficient fine-grained classification accuracy and the low recall rate of watercore samples was the key directions for subsequent optimization.