Determination of Total Soil Organic C and Hot Water-Extractable C from VIS-NIR Soil Reflectance with Partial Least Squares Regression and Spectral Feature Selection Techniques

Abstract

The calibration of soil organic C (SOC) and hot water-extractable C (HWE-C) from visible and near-infrared soil reflectance spectra is hindered by the complex spectral interaction of soil chromophores that usually varies from one soil or soil type to another. The exploitation of spectral variables from spectroradiometer data is further affected by multicollinearity and noise. In this study, a set of soil samples (Fluvisols, Podzols, Cambisols and Chernozems; n = 48) representing a wide range of properties was analysed. Spectral readings with a fibre-optics visible to near-infrared instrument were used to estimate SOC and HWE-C contents by partial least squares regression (PLS). In addition to full-spectrum PLS, spectral feature selection techniques were applied with PLS (uninformative variable elimination, UVE-PLS, and a genetic algorithm, GA-PLS). On the basis of normalized spectra (mean centring + vector normalization), the order of prediction accuracy was GA-PLS ¿¿ UVE-PLS ¿ PLS for SOC; for HWE-C, it was GA-PLS ¿ UVE-PLS, PLS. With GA-PLS, acceptable cross-validated (cv) prediction accuracies were obtained for the complete dataset (SOC, R-cv(2) = 0.83, RPDcv = 2.42; HWE-C-cv, R-cv(2) = 0.78, RPDcv = 2.13). Splitting the soil data into two groups with different basic properties (Podzols compared with Fluvisols/Cambisols; n = 21 and n = 23, respectively) improved SOC predictions with GA-PLS distinctly (Podzols, R-cv(2) = 0.89, RPDcv = 3.14; Fluvisols/Cambisols, R-cv(2) = 0.92, RPDcv = 3.64). This demonstrates the importance of using stratified models for successful quantitative approaches after an initial rough screening. GA selection frequencies suggest that the spectral region over 1900 nm, and in particular the hydroxyl band at 2200 nm are of great importance for the spectral prediction of both SOC and HWE-C.