Rapid elemental prediction of heterogeneous tropical soils from pXRF data: a comparison of models via linear regressions and machine learning algorithms

Abstract

Context: USEPA 3051a is a standard analytical methodology for the extraction of inorganic substances in soils. However, these analyses are expensive, time-consuming and produce chemical residues. Conversely, proximal sensors such as portable X-ray fluorescence (pXRF) spectrometry reduce analysis time, costs and consequently offer a valuable alternative to laboratory analyses.

Aim: We aimed to investigate the feasibility to predict the results of the USEPA 3051a method for 28 chemical elements from pXRF data.

Methods: Samples (n = 179) representing a large area from Brazil were analysed for elemental composition using the USEPA 3051a method and pXRF scanning (Al, Ca, Cr, Cu, Fe, K, Mn, Ni, P, Pb, Sr, Ti, V, Zn and Zr). Linear regressions (simple linear regression – SLR and stepwise multiple linear regressions – SMLR) and machine learning algorithms (support vector machine – SVM and random forest – RF) were tested and compared. Modelling was developed with 70% of the data, while the remaining 30% were used for validation.

Key results: Results demonstrated that SVM and RF performed better than SLR and SMLR for the prediction of Al, Ba, Bi, Ca, Cd, Ce, Co, Cr, Cu, Fe, Mg, Mn, Mo, P, Pb, Sn, Sr, Ti, Tl, V, Zn and Zr; R2 and RPD values ranged from 0.52 to 0.94 and 1.43 to 3.62, respectively, as well as the lowest values of RMSE and NRMSE values (0.28 to 0.70 mg kg−1).

Conclusions and implications: Most USEPA 3051a results can be accurately predicted from pXRF data saving cost, time, and ensuring large-scale routine geochemical characterisation of tropical soils in an environmentally friendly way.