Using rule-based regression models to predict and interpret soil properties from X-ray powder diffraction data

Abstract Data mining is often used to derive calibrations for soil property prediction from diffuse reflectance spectroscopy, facilitating inference of organic and mineral contributions to given properties. In contrast to spectroscopy, X-ray powder diffraction (XRPD) offers a more direct probe into the complexities of soil mineralogy. Here a national scale XRPD dataset of Scottish soils is used in combination with the rule-based regression algorithm ‘Cubist’ for prediction of eight soil properties (total carbon and nitrogen, cation exchange capacity, pH, aqua regia extractable potassium, and the sand, silt and clay size fractions), and interpretation of soil property–mineralogy relationships. Precision sample preparation methods prior to XRPD analysis eliminated effects of preferred orientation, creating reproducible data appropriate for data mining. For direct comparison, Cubist was also applied to an equivalent dataset of near infrared spectroscopy (NIRS) measurements. In terms of predictive performance, XRPD surpassed NIRS for prediction of six of the eight soil properties investigated. Notably, diffuse scattering from X-ray amorphous organic matter facilitated relatively accurate predictions of total carbon and nitrogen from XRPD. Aqua regia extractable potassium was predicted with substantial accuracy and confirmed to reflect the phyllosilicate potassium. The particle size fractions were predicted with moderate-substantial agreement using combinations of quartz, phyllosilicate and feldspar variables. This approach introduces the value of XRPD datasets in enhancing the understanding of soil mineralogy–property relationships whilst contributing to soil mineralogy's advance into the digital soil typing paradigm.