Can spectral analyses improve measurement of key soil fertility parameters with X-ray fluorescence spectrometry?

While laboratory methods of elemental analysis of soil nutrients are used frequently to support soil studies, the implementation of more portable and cost-efficient methods lingers behind. The portable (handheld) X-ray fluorescence spectrometer (XRF) is one such tool enabling onsite elemental analysis in a straightforward manner. However, in soil studies the use of XRF often remains cumbersome, following the poor performance of the method for low-Z elemental analysis and the complex nature of the soil matrix, introducing background noise. Here, we therefore evaluate how the potential use of a portable XRF for predicting potassium (K), phosphorus (P), magnesium (Mg) and calcium (Ca) can be improved through the analysis of XRF spectral data with the Random Forest (RF) machine learning method. A total of 105 soil samples from a wide range of soils collected from 10 different countries (D.R. Congo, Belgium, Ivory Coast, Italy, The Netherlands, Saudi-Arabia, South Africa, Spain, Switzerland, and Zimbabwe) were scanned using an Oxford XMET8000 XRF spectrometer (Oxford Instruments, UK). Spectral data of the calibration set (n = 74) were pulled in one matrix alongside measured elemental concentrations and subjected to RF analysis. Resulting models were validated using an independent validation set (n = 31). The best RF prediction result was obtained for K followed successively by Ca, Mg and P with coefficient of determination (R-2) values of 0.83, 0.76, 0.69, and 0.47, and root mean squared error of prediction (RMSEP) of 2283.8, 6818.7, 1511.8, and 538.08 mg kg(-1), respectively. The RF modelling procedure provided improved prediction performance compared to the calibration models provided by the manufacturer (R-2 = 0.65, 0.75, 0.65, and 0.22, for K, Ca, Mg and P, respectively). Our results suggest that portable XRF instruments coupled with spectral data analysis by RF allows for rapid and low-cost analysis of soil K and Ca with remarkable accuracy. Still, the lower measurement accuracy for P and Mg suggests further work is needed to test whether the prediction can be improved by better calibration models, and how such approaches can help overcoming instrumental limitations.