Improved prediction of protein content in wheat kernels with a fusion of scatter correction methods in NIR data modelling

Abstract

The study aims to test the hypothesis that modelling of near-infrared (NIR) spectroscopic data based on a single scatter correction technique is sub-optimal. Better predictive performance of the multivariate analysis method can be obtained when the information from differently scatter corrected data is jointly used. To demonstrate it, an open-source NIR spectroscopy data set related to protein prediction in wheat kernels was used. Two different pre-processing fusion approaches i.e., sequential and parallel fusion, were used for fusing the complementary information from four different scatter correction techniques, namely standard normal variate (SNV), variable sorting for normalisation (VSN), 2nd derivative, and multiplicative scatter correction (MSC). As a comparison, partial least-squares regression (PLSR) was performed on the SNV pre-processed data. The results showed that fusion of scatter correction can improve the predictive performance of NIR spectroscopic models. The results revealed that both sequential and parallel fusion approaches improved the predictive performance compared to the PLSR performed using a single scatter correction technique. The R2p was improved by up to 3% and the RMSEP was reduced by up to 13% compared to the results obtained with conventional PLSR model developed with a single scatter correction technique.