The feasibility of Vis/NIR spectroscopy for detection of flaws in hazelnut kernels (Corylus avellana L. cv. Tonda Gentile Romana) is demonstrated. Feature datasets comprising raw absorbance values, raw absorbance ratios (Abs[lambda(1)]:Abs[lambda(2)]) and differences (Abs[lambda(1)] - Abs[lambda(2)] for all possible pairs of wavelengths from 306.5 nm to 1710.9 nm were extracted from the spectra for use in an iterative LDA routine. For each dataset, several spectral pretreatments were tested. Each group of features selected was subjected to Partial Least Squares Discriminant Analysis (PLS-DA), Receiver Operating Characteristics (ROCs) analysis, and evaluation of performance through the Area Under ROC Curve. The best result (5.4% false negative, 5.0% false positive, 5.2% total error) was obtained using a Savitzky-Golay second derivative on the dataset of raw absorbance differences. The optimal features were Abs[564 nm]-Abs[600 nm], Abs[1223 nm] -Abs[1338] and Abs[1283 nm]-Abs[1338 nm]. The results indicate the feasibility of a rapid, online detection system. 2013 Elsevier Ltd. All rights reserved.