Abstract: Near-infrared (NIR) spectroscopy was applied to rapidly and non-destructively distinguish between bruised and intact apples and identify different apple varieties. Besides, the effect of different discrimination methods on the distinguishing and identification models obtained was investigated. The results indicated that discrimination models were developed based on the first five principal components in the range from 12000 cm-1 to 4000 cm-1 from the NIR spectra of bruised and intact apples subjected to wavelet pretreatment using three different discrimination methods, multilayer perceptron (MLP) neural network, radial basis function (RBF) neural network and Fisher line discriminant analysis (Fisher-DA) with discrimination accuracy rates of 97.8%, 87.2% and 84.8%, respectively for unknown samples. The Fisher linear discriminant analysis model established based on multiple linear regression and the loading weights showed a discrimination accuracy rate of 89.1% compared with 100% for the model established using partial least squares discriminant analysis (PLS-DA). The discrimination accuracy rates of the PLS-DA model for the training and validation sets were both 100%, and therefore the PLS-DA model was superior to others. A better discrimination model was established based on the first six principal components of the NIR spectra of different apple varieties subjected to smoothing pretreatment in the full wavelength range of 12000－4000 cm-1 than in the empirical range of 8000－4500 cm-1 with discrimination accuracy rates of 90.9% and 92.1% for the predication and validation sets, respectively. In conclusion, the combination of NIR and chemometrics can provide a rapid and non-destructive approach to discriminate whether apples are bruised and identify different apple varieties.

Key words: apple, NIR spectroscopy, artificial neural network, PLS-DA, Fisher-DA