An unbiased method capable of classifying hazelnut (Corylus avellana L.) in compliance with the statements set out by the Commission Regulation (EC) No. 1284/2002 is economically important to the fresh and processed industries. Thus, in this study, the feasibility of High Dynamic Range (HDR) hyperspectral imaging for hazelnut kernel sorting (cv. Tonda Gentile Romana) of four quality classes (‘Class Extra’, ‘Class I’, ‘Class II’ and ‘Waste’) has been investigated. Two different exposure times (5 and 8 ms) were selected for experiments, and the respective spectra were combined to obtain a high dynamic range over the full spectral range. The illumination setup was also optimized to improve the intensity and uniformity of the light along the field of view of the camera. PLS-DA was used to classify the pixels based on their spectra and the spectral pre-treatment was optimized through an iterative routine. The performance of each PLS-DA model was defined based on its sensitivity, selectivity, α-error, β-error and accuracy rates. All of the selected models provided a very-good (> 90%) or good (>80%) sensitivity and selectivity for the predefined classes. Misclassified kernels were primarily assigned to the low-quality classes (i.e. ‘Class II’ and Waste). Moreover, the spatial domain was used to evaluate the feasibility of distinguishing hazelnut classes on the basis of their size and shape. It was found that hazelnut dimensions can be used to improve the accuracy of the classification of the kernels. Thanks to this combination of both spectral and spatial information spectral imaging could be used for quality sorting of hazelnuts.