Evaluation of pre-processing and variable selection on energy dispersive X-ray fluorescence spectral data with partial least square regression: A case of study for soil organic carbon prediction

Abstract

Most studies which have reported soil fertility attributes employing Energy Dispersive X-ray Fluorescence (EDXRF) combined with multivariate calibration make use of elemental concentration data. This combination may cause relevant information loss contained in EDXRF spectra. However, a well-established soil EDXRF spectra data treatment procedure for multivariate calibration is not currently available. The objective of this study was to evaluate the influence of different pre-processing and variable selection methods in partial least square regression models using EDXRF spectral data. Measurements were obtained under two experimental conditions (15 kV and 50 kV at tube) for soil organic carbon determination. Poisson scaling + mean center proved to be the most suitable pre-processing for this data set. The variable selection by successive projection algorithm for interval selection in partial least squares improved the performance of all tested pre-processing (or at least kept constant in terms of the errors). The 15 kV condition models with Pareto scaling and Poisson scaling + mean center were the most accurate and precise. The ratio to performance of deviation values for these models was of 2.2. The figures of merit demonstrated the soil organic carbon determination feasibility using EDXRF spectral data with these pre-processing since the accuracy, precision and limits of detection were consistent with previous reports. Thus, this study contributes toward the establishment of an approach for soil EDXRF spectral data treatment for multivariate calibration. It also contributes to a better EDXRF variables interpretation which impacts soil organic carbon modeling, demonstrating the proposed methodology potential.