Optimization criteria in sample selection step of local regression for quantitative analysis of large soil NIRS database

Abstract

Large soil spectral libraries compiling thousands of NIR (Near Infrared) reflectance spectra have been created encompassing a wide diversity and heterogeneity of spectra. Among the many chemometric approaches to the calibration of chemical and physical properties from these large libraries, local calibrations have the advantage of being able to select the most similar spectra to the spectrum of a target sample. This is particularly relevant when dealing with highly heterogeneous media such as soils, where the mineral matrix has a strong influence on spectral features. A crucial step in the implementation of local calibration procedures is the construction of local neighbourhoods. In this study, we investigate the influence of index computation and neighbour selection on calibration results using local PLSR models on a large soil spectral database. Our indices combine two spectral compression methods (Principal Component Analysis or Fast Fourier Transform) with two distinct distance metrics (Mahalanobis distance or correlation coefficient). Based on a large collection of soil samples provided by the French National Soil Quality Monitoring programme, we constructed calibration models to estimate two chemical (organic carbon and cationic exchange capacity) and two physical (clay and sand content) factors. After neighbour selection, local Partial Least Squares regressions were applied to the selected spectra. Our results highlight the utility of the Fourier transformation of the spectra compared to the classical PCA compression method in achieving a more appropriate neighbourhood selection. We propose an index based on the coefficient correlation with FFT compression that led to a neighbourhood selection giving the best prediction results for the four considered soil constituents.