Identification of early bruising degrees in blueberries using visible and near-infrared spectroscopy coupled with...

The early detections of internal bruises in blueberries caused by external impacts after harvest play a crucial role in enhancing their economics. The aim of this study was to identify early bruise blueberries at different impact energies based on Partial Least Squares Discriminant Analysis (PLSDA), Convolutional Neural Network (CNN) and Tabular Transformer (TabTransformer) models using a new developed spectroscopic system to collect blueberry reflectance spectra with high signal-to-noise ratio covering 400-1000 nm and 1350-2200 nm. The microscopic structures of the peel and pulp between healthy and bruised blueberries were observed to determine the relationship between bruise degrees and cell structures. The results showed two deep learning models exhibited better detection performance than PLSDA model with highest accuracies for discriminating bruise degrees of 96.9 % for CNN model and 91.8 % for TabTransformer model. The binary classification of each bruise degree demonstrated slightly higher results for CNN model with accuracies over 95 % for almost each bruise degree. The lower classification accuracy observed in the 1350-2200 nm was likely due to some cell disruption and release of free water, which influenced the detection of strong water absorption bands beyond 1400 nm. These findings presented the potential to recognize the bruise degrees using developed optical system coupled with deep learning models.