Multivariate chemometrics as a key tool for prediction of K and Fe in a diverse German agricultural soil-set using EDXRF

Abstract

Within the framework of precision agriculture, the determination of various soil properties is moving into focus, especially the demand for sensors suitable for in-situ measurements. Energy-dispersive X-ray fluorescence (EDXRF) can be a powerful tool for this purpose. In this study a huge diverse soil set (n = 598) from 12 different study sites in Germany was analysed with EDXRF. First, a principal component analysis (PCA) was performed to identify possible similarities among the sample set. Clustering was observed within the four texture classes clay, loam, silt and sand, as clay samples contain high and sandy soils low iron mass fractions. Furthermore, the potential of uni- and multivariate data evaluation with partial least squares regression (PLSR) was assessed for accurate determination of nutrients in German agricultural samples using two calibration sample sets. Potassium and iron were chosen for testing the performance of both models. Prediction of these nutrients in 598 German soil samples with EDXRF was more accurate using PLSR which is confirmed by a better overall averaged deviation and PLSR should therefore be preferred.

Figure 1

PCA score plot (a) of EDXRF soil data for the first two principal components PC-1 and PC-2 of 13 certified reference materials with a total variance of 99%. Corresponding loading plot (b) of the first three principal components for CRM. Loadings are plotted against the energy [keV]. PC-1 corresponds to iron, PC-2 to calcium and PC-3 to silicon.

Figure 2

PCA score plot, projection of the EDXRF German agricultural soil data for the first two principal components PC-1 and PC-2 of 13 certified reference materials.

Figure 3

Univariate calibration model for potassium (a,b) and iron (c,d) based on 15 CRM. The net peak area of the characteristic fluorescence peak [cps] was fitted against the mass fraction in the CRM [wt-%] (a,c). The error bars represent the standard deviation of five measurements. For comparison purpose of the univariate calibration model: predicted values [wt-%] were fitted against the reference values [wt-%] (b,d).

Figure 4

Predicted values [wt-%] of potassium (a) and iron (b) using univariate analysis with 41 German agricultural soils fitted against the reference values of WDXRF [wt-%].

Figure 5

Partial least squares regression of 15 CRM for K (a) and Fe (b) using Savitzky-Golay derivation, linear baseline correction as data pre-treatment, Kernel as algorithm and leave-one-out cross-validation.

Figure 6

Partial least squares regression with 41 German agricultural soils for K (a) and Fe (b) using Savitzky-Golay derivation, linear baseline correction as data pre-treatment, Kernel as algorithm and leave-one-out cross-validation.