Integration of vis-NIR and pXRF spectroscopy for rapid measurement of soil lead concentrations

Heavy metals accumulate in soil over time and, with changing land use, humans may be exposed to elevated contaminant concentrations. It is therefore important to delineate contaminated sites in the most efficient and accurate manner. Sensors, such as portable X-ray fluorescence (pXRF) and visible near-infrared (vis-NIR) spectroscopy predict metal concentrations more rapidly and in a less hazardous manner compared to traditional laboratory analytical methods. The current study explored the potential for integrating vis-NIR and pXRF outputs to improve lead predictions in fine- (<62.5 mu m) and whole-fraction (<2 mm) soil samples. A multi-stage approach was taken to compare different data treatments and combination methods for the prediction of whole-fraction lead content. Data treatment included principal component analysis, and combination methods included concatenation of pXRF and vis-NIR spectra before modelling, and Granger-Ramanathan model averaging of pXRF and vis-NIR model outputs. The most accurate predictions of whole-fraction lead were obtained by Granger-Ramanathan model averaging of vis-NIR Cubist predictions and Compton-normalised pXRF output: Lin's Concordance Correlation Coefficient (LCCC) = 0.95, root mean square error (RMSE) = 86.4 mg kg(-1), Bias < 0.001 mg kg(-1) and ratio of performance to inter-quartile range (RPIQ) = 0.37. The most suitable modelling method was then used to predict fine-fraction lead, which provided a similarly accurate model fit (LCCC = 0.94, RMSE = 84.2 mg kg(-1), Bias < 0.001 mg kg(-1) and RPIQ = 0.34), indicating the potential to reduce the number of samples required for fine-fraction processing. In addition, the quality of the prediction interval estimates was examined - an important aspect in modelling which is underutilised in current literature related to soil spectroscopy.