Checks and Mass Balances for In Situ Quantification of Mineral Composition using Proximal Soil Sensors

Abstract

Soil mineral composition affects soil behavior, but field estimation of mineral composition has proved difficult. To investigate the potential of predicting soil mineral composition in situ, 15 soils representing diverse mineral composition from New South Wales, Australia, were scanned with visible near‐infrared (VisNIR) and portable X‐ray fluorescence (pXRF) spectrometers to a depth of 1 m at 2.5‐cm scanning increments. The presence of phyllosilicate and Fe oxides was assessed using a pattern‐matching algorithm with VisNIR spectra from mineral end‐member libraries. Rule‐based iterative partitioning was then applied on pXRF elemental compositions based on mineral stoichiometry to determine the abundance of kaolinite, smectite, illite, hematite, goethite, calcite, gypsum, and quartz. This fusion model correctly identified the most abundant mineral in a sample with 72% accuracy. Mineral predictions were stable under variable moisture and surface conditions, as experienced when scanning samples in situ and under air‐dry and ground conditions in the laboratory. Relative changes in mineral composition within a profile and across horizon boundaries were accurately expressed. The fusion model accurately quantified the abundance of quartz (Cohen's kappa coefficient, 0.67), and CaCO3 (Cohen's kappa coefficient, 0.76; Lin's concordance correlation coefficient, 0.96; RMSE, 20.9 g kg−1). The dominant phyllosilicate mineral was identified correctly with 86% accuracy, although accurate quantification of phyllosilicates and Fe oxides was not achieved. Conjoint use of VisNIR and pXRF spectroscopy as part of a fusion model approach showed great potential to provide comprehensive estimation of soil mineral composition in situ.Core Ideas VisNIR and pXRF were used to quantify soil mineral composition in situ. The fusion model approach leveraged the complementary strengths of VisNIR and pXRF. The most abundant mineral of a sample was predicted with 78% accuracy. Quantitative predictions were achieved for quartz, CaCO3, and gypsum. This is the most comprehensive investigation of its kind to date.