Abstract
The accuracy of the y vectors estimated by near infrared spectroscopic models depends on the quality of the reference method. In this paper the reference data values are augmented with noise. The level of noise added ranged from 0 to 20% of the variability for the mean y values. Partial least squares models are then calculated for each addition of noise resulting in many regression coefficient vectors that are used to produce the final calibration model. This final model is selected as the one with the highest R2 with the median of all these coefficient vectors. This model, applied on unchanged independent test sets, produced a root mean square error of prediction that leads to a clear improvement over the original value without noise. The ability of this technique is investigated on a simulated and on an industrial data set.