Identifying soil provenance based on portable X-ray fluorescence measurements using similarity and inverse-mapping approaches – A case in the Lower Hunter Valley, Australia

Abstract There is a growing interest in the use of soil composition as a form of evidence in food provenance, forensics, biosecurity, and archaeology. Given a soil sample of unknown origin, we should like to know the likely geographical source of that material. In this study, we investigated whether data provided from a rapid and non-destructive sensor can be used to identify the provenance of a soil sample. A portable X-ray fluorescence (pXRF) spectrometer was used to measure the elemental abundance of 0–10 cm soil samples from a part of the Lower Hunter Valley, NSW, Australia (an area of 328 km2). Three methods, namely, two similarity methods (points of similarity and regions of similarity) based on distances to the unlocated specimen in the principal component (PC) space of the geochemical data, and an artificial neural network (ANN) method, effectively an inverse digital soil mapping (DSM) approach, which predicts location from the set of geochemical variables, were tested to determine the provenance of soil samples. In the PC approach, digital soil maps of the PC scores of eight major elements and two elemental ratios were created. The locations predicted by the PC approach seemed to follow the pattern of topography. In the ANN approach, the geographical coordinates (Eastings and Northings) of a sample were predicted simultaneously using the elemental concentrations and ratios. Using maps of elemental concentration classes (regions of similarity based on PC) provided a mean RMSE of 8.6 km for the 147 validation samples. The different effects of identification of geographical locations were compared using a 95% spatial confidence interval of prediction on a validation dataset. The points of similarity based on PC approach showed that the predicted search areas can capture 59% of the true locations of the test data. Meanwhile, the ANN approach can capture 69% of the true locations of the data. The mean RMSE for ANN prediction (2.8 km) was smaller than that for points of similarity prediction (4.3 km). Both soil provenancing approaches are potentially useful in identifying geographical areas of origin or similarity.