Using the Frequency Components of near Infrared Spectra: Optimising Calibration and Standardisation Processes

Abstract

The modification of frequency components (Fourier coefficients and wavelet detail component) of near infrared spectra for the optimisation of calibration and standardisation processes was investigated. High-frequency components were smoothed and approximated to remove components most likely to represent noise and background information. Savitzky-Golay smoothing and signal correction were used for that purpose. Frequency modification methods were used in addition to wavelength domain processing techniques. Whole soybean protein and oil calibrations were developed on four instruments with their own calibration set (two Foss Infratecs and two Bruins OmegAnalyzerGs). A validation strategy with two sample sets of known and of new variability was implemented. Frequency modification methods showed improvements of the prediction precision in calibration (relative predictive determinant ( RPD) increased from 8.57 to 9.25 for protein and from 7.01 to 7.28 for oil with Fourier coefficients-based smoothing for Infratec 124.10350). Frequency based pre-processing methods were also successful when transferring prediction models in intra and inter-brand situations ( RPD of the secondary unit of 9.21 compared to original RPD of 8.45 in intra-brand for Foss network for protein; RPD of secondary unit of 9.33 compared to original RPD of 8.74 for inter-brand scenario with Foss Infratec 1241 master of Bruins units for protein). The smoothing of Fourier coefficients showed the best results. Prediction accuracies were not modified by the frequency-based modifications, except in the inter-brand scenario. An appropriate pre-processing limited the need for other standardisation methods except in inter-brand situations where a bias correction should be implemented. Frequency-based pre-treatment methods tend to specialise the calibration set to optimise predictions. This may not be suitable when the variability of the future samples is not included in the calibration set (i.e. yearly variability of agricultural products).