LAUSR

The hyperspectral imaging methods used previously for analyzing food quality and safety focused on spectral data analysis to elucidate the spectral characteristics relevant to the quality and safety of food and agricultural commodities. However, the use of spatial information, including physical size, geometric characteristics, orientation, shape, color, and texture, in hyperspectral imaging analysis of food safety and quality has been limited. In this study, image processing techniques were employed for extracting information related to the morphological features of fifteen different foodborne bacterial species and serotypes, including eight Gram-negatives and seven Gram-positives, for classification. The values of nine morphological features (maximum axial length, minimum axial length, orientation, equivalent diameter, solidity, extent, perimeter, eccentricity, and equivalent circular diameter) of bacterial cells were calculated from their spectral images at 570 nm, which were selected from hyperspectral images at 89 wavelengths based on peak scattering intensity. First, two classes (Gram-negative and Gram-positive) were classified using a support vector machine (SVM) algorithm, resulted in a classification accuracy of 82.9% and kappa coefficient (kc) of 0.65. Thereafter, a classification model was developed with two features (cell orientation and perimeter) selected by principal component analysis. In addition, a decision tree (DT) algorithm was used for classification with all nine morphological features. With respect to differentiation into two classes (Gram-positive and Gram-negative), the classification accuracy for five selected bacteria species (, , Typhimurium, , and ) decreased to 80.0% (0.74 of kc) with the DT algorithm and to only 72.5% (0.64 of kc) with the SVM algorithm. Thus, the hyperspectral microscopy image analysis with morphological features is limited for classifying foodborne pathogens, so additional spectral features would be helpful for classification of foodborne bacteria.