Evaluating the impact of NIR pre-processing methods via multiblock partial least-squares

Abstract Near-infrared (NIR) spectral data are used in many applications to predict physical and chemical properties. However, it can result in poor predictive models when untreated spectra are directly used to estimate these properties. Many pre-preprocessing techniques are available to reduce noise and variance unrelated to the studied property but choosing which one to apply can be tricky. Existing methods to select a pre-processing are time-consuming or do not allow for a meaningful comparison of the different techniques. Even though new methods focus on extracting complementary information from each pre-processing, an optimal combination is still required to obtain efficient predictive models and avoid extensive computational costs. Here, we propose an approach using multiblock partial least squares (MBPLS) to simultaneously compare the impact of the pre-processing techniques on spectral data and as a result on the regression models. Superloadings provide qualitative and quantitative information on pre-processed data. This tool helps compare and determine which pre-processing technique, or combinations thereof, that may be appropriate for a dataset, not just a single “best” one. Using this, the analyst is then better equipped to make a final choice when selecting which ones to include. This method is tested on artificial signals and NIR spectra from corn samples.