Potential of MIR, TXRF and XRD as Complementary Techniques for Assessment of Soil Properties

Abstract

Objectives: To evaluated the potential of combinations of MIR, TXRF and XRD techniques for prediction of soil properties. Methods: Using Random Forests (RF) algorithm, MIR prediction of soil properties for a set of geo-referenced soil samples associated with the Africa Soil Information Service (AfSIS) project, from 44 sentinel sites, randomized across sub-Sahara Africa, were done . RF regression models were used to model the residuals of the predictions against TXRF spectra combined with MIR first derivative spectra and also against dominant mineralogy groupings measured using XRD. Results:  MIR resulted in good prediction models (R2>0.80) using RF out-of-bag validation for organic and total C and N, extractable Ca and Al, and pH. Acceptable models (R2>0.60) were obtained for exchangeable bases, but models were poor (R2<0.60) for exchangeable Na, and Mehlich-3 extractable P, K, S, Na, Cu, and Zn. The prediction performance of combined MIR and TXRF spectra was better than MIR alone for several soil properties. Relative improvements in prediction accuracy over MIR alone were: exchangeable Na (828%), Mehlich-3 S 350%; Mehlich-3 Zn; 150%; and Mehlich-3 Cu, 10%. However, TXRF data was generally limited in explaining the residuals of the MIR predictions for extractable P and K. Conclusions: Element concentrations measured using TXRF could be predicted relatively well from XRD-measured mineralogy composition with prediction r-square of 0.71 for raw mineralogy traces and 0.81 for dominant mineralogy grouping. The results indicate that TXRF and/or XRD fingerprinting reflects mineralogical composition and could supplement MIR prediction of soil properties and help stabilize calibration models across soil types with widely different mineralogy.