Abstract

Portable X-ray fluorescence (PXRF) measurements on 1520 soil samples were used to create national prediction models for copper (Cu), zinc (Zn), and cadmium (Cd) concentrations in agricultural soil. The models were validated at both national and farm scales. Multiple linear regression (MLR), random forest (RF), and multivariate adaptive regression spline (MARS) models were created and compared. National scale cross-validation of the models gave the following R-2 values for predictions of Cu (R-2 = 0.63), Zn (R-2 = 0.92), and Cd (R-2 = 0.70) concentrations. Independent validation at the farm scale revealed that Zn predictions were relatively successful regardless of the model used (R-2 > 0.90), showing that a simple MLR model can be sufficient for certain predictions. However, predictions at the farm scale revealed that the non-linear models, especially MARS, were more accurate than MLR for Cu (R-2 = 0.94) and Cd (R-2 = 0.80). These results show that multivariate modelling can compensate for some of the shortcomings of the PXRF device (e.g., high limits of detection for certain elements and some elements not being directly measurable), making PXRF sensors capable of predicting elemental concentrations in soil at comparable levels of accuracy to conventional laboratory analyses.

Keywords

PXRF; soil; copper; zinc; cadmium; machine learning; precision agriculture

Published in

Sensors (Basel, Switzerland)
2020, volume: 20, number: 2, article number: 474
Publisher: MDPI

Authors' information