The off-line contact measuring configuration, described in the section Off-line contact and non-contact measurements (summary given in Table 2), should theoretically give the best prediction result, due to the elimination of all surface reflection and stray light of the measurements. The off-line non-contact and the online measurements do allow some surface reflection and stray light to be collected. The regression model shows, not surprisingly, a slightly, but not significantly lower prediction performance for the off-line non-contact compared to the off-line contact measurements. However, the on-line measurements provided a significant increase in prediction performance compared to the offline non-contact configuration, but the PLS model needed one additional PLS factor, as described in the section On-line measurements. Since the dry matter content varies along the length of the tuber, the on-line measurements might therefore give a better spectral representation of the dry matter content in the tuber and a correspondingly better prediction. Another possible explanation is that the recording and averaging of multiple spectra across a moving potato may eliminate most of the contribution from local skin defects, such as scabs and scurfs common on potatoes [27]. In this study three replicates of each potato tuber were averaged before the calibration model were obtained. However, based on a PCA analysis of the replicates (data not shown) and the information in Fig. 4, no large replicate variance is seen. Hence, one replicate may be enough if the instrument is mounted above a conveyor belt. It should also be noted that the validation approach used in the current study, i.e. leave-one-out cross validation, often is regarded as an optimistic approach. Thus, for industrial applications additional data, based on independent test sets, should be added. However, in the present feasibility study, the validation approaches could be regarded as sufficient.