Main content area

Detection of blueberry internal bruising over time using NIR hyperspectral reflectance imaging with optimum wavelengths

Fan, Shuxiang, Li, Changying, Huang, Wenqian, Chen, Liping
Postharvest biology and technology 2017 v.134 pp. 55-66
blueberries, hyperspectral imagery, image analysis, least squares, models, reflectance, selection methods, storage quality, support vector machines, wavelengths
Early detection of internal bruising is one of the major challenges in blueberry postharvest quality sorting processes. The potential of using near infrared (NIR) hyperspectral reflectance imaging (950–1650nm) with reduced spectral features was investigated for blueberry internal bruising detection 30min to 12h after mechanical impact. A least squares support vector machine (LS-SVM) was used to develop classification models to compute the spatial distribution of bruising based on the spectra extracted from regions of interest (ROIs) at four measurement times (30min, 2h, 6h, and 12h after mechanical impact). Three feature selection methods were used to select optimum wavelengths or band ratio images for bruising detection. The classification model, developed using optimum wavelengths selected by competitive adaptive reweighted sampling (CARS) (CARS-LS-SVM model) and full spectra (full spectra-LS-SVM), had similar performance in the identification of bruised blueberries. Band ratio images (1235nm/1035nm) achieved a comparable accuracy with the CARS-LS-SVM model at 6h, and higher accuracy than CARS-LS-SVM and full spectra-LS-SVM models at 12h. The overall classification accuracies of 77.5%, 83.8%, 92.5%, and 95.0% were obtained by band ratio images for blueberries 30min, 2h, 6h, and 12h after impact, respectively. In order to evaluate the performance of the proposed methods, additional validation samples were processed by the detection algorithm. The overall discrimination accuracies for healthy and bruised blueberries in the validation set were 93.3% and 98.0%, respectively, for CARS-LS-SVM model, and 93.3% and 95.9%, respectively, for band ratio images. The overall results indicated that NIR reflectance imaging can detect blueberry internal bruising as early as 30min after mechanical impact, and band ratio images computed from two wavelengths showed great potential to detect blueberry internal bruising on the packing line.