Optimized Multivariate Analysis for the Discrimination of Cucumber Green Mosaic Mottle Virus-Infected Watermelon Seeds Based on Spectral Imaging

Abstract

This study proposes a nondestructive sorting method based on the short-wave infrared hyperspectral imaging technique (SWIR-HIT) to detect and classify watermelon seeds infected with the cucumber green mosaic mottle virus (CGMMV). Virus-infected watermelon seeds were collected from virus-infected watermelon plants. Five plates each with 81 seeds were scanned. A total of 304 mean reflectance spectra were used to develop and evaluate virus-infected seed classification models with multivariate analysis methods such as partial least squares discriminant analysis (PLS-DA), support vector machine (SVM), and least squares support vector machine (LS-SVM). To determine the optimal preprocessing method, three preprocessing methods were employed: multivariate scatter correct (MSC) as well as first- and second-derivative preprocessing with the Savitzky–Golay algorithm. Among these methods, second-derivative preprocessing with the LS-SVM method showed an approximately 75% accuracy with a 0.57 kappa coefficient for all three classification classes (infected, infection suspected, and sound seeds). Binary classification between infected and sound seeds by LS-SVM with second-derivative preprocessing showed an approximately 92% accuracy with a 0.75 kappa coefficient. To improve the classification accuracy, the genetic algorithm was implemented, and 9 bands were selected. The selected wavelengths were applied to develop and compare classification models with full wavelengths. The three-class classification with the selected bands showed an approximately 80% accuracy, whereas binary classification in infected and sound seeds showed a more than 93% accuracy with a 0.78 kappa coefficient. These results indicate that SWIR-HIT is a valuable nondestructive tool for rapidly classifying CGMMV-infected watermelon seeds using LS-SVM with raw spectra.