Selection of informative wavelengths is a crucial step for online spectral imaging applications. In this paper, four cultivars of apples were utilized to select effective wavelengths for bruise detection within 380-1000 nm. Each of the wavelength variables was considered as an independent classifier for bruised/normal classification, and all classifiers were evaluated and compared by receiver operating characteristic (ROC) analysis. According to the area under the ROC curve (AUC), the reflectance difference between two wavelengths (R(lambda(1)) - R(lambda(2))), was determined as the best wavelength pair for the Fuji, Jon-agold, Orin and Sinano Gold cultivars as follows: R(808 nm) - R(760 nm), R(832 nm) - R(772 nm), R(834 nm) - R(762 nm) and R(788 nm) - R(742 nm), respectively. The performance of the wavelengths selected in this paper was measured by comparing the predicted sensitivity, predicted specificity and predicted classification accuracy with those of the model proposed by partial least squares discriminant analysis (PLSDA), which used all of the wavelength variables. The results showed that the predictive ability of both methods was generally on the same level. (C) 2011 Elsevier Ltd. All rights reserved.