NIR PLSR results obtained by calibration with noisy, low-precision reference values: Are the results acceptable?

Abstract

Abstract Both spectral noise and reference method noise affect the accuracy and the precision NIR predicted values. The reference noise is often neglected, and the few reports dealing with it only consider random noise artificially added to the original sound reference data. A calibration for lignin content of maritime pine (Pinus pinaster Ait.) wood meal was developed, but due to low precision and accuracy in the reference data set, NIR partial least-squares regression (PLSR) yielded a slope of 0.51 and an intercept at 14% Klason lignin. We demonstrate with an independent data set for external validation, obtained with higher precision and accuracy, that the NIR PLSR model based on the noisy reference data led to better results. The slope of the correlation between predicted and reference values was 0.89 and the intercept was 3.9. Thus, the model performed much better than expected from the cross-validation results. The predictability can be explained by the facts that the loadings of the first principal component (PC) of the calibration and test samples are very similar and dominated by lignin-related bands, and that most of the variation in the test set can be explained by the first PC. This only explains why the Klason lignin content could be predicted with the model without giving many spectral outliers, but not the good result of the external validation. We show that the latter can be explained by the inverse calibration used for PLSR and that predicted values can be more accurate and precise than the reference values used for calibration.