Estimating textural fractions of the USDA using those of the International System: A quantile approach

Abstract

In soil science, the two most frequently used classification systems for the soil particle size distribution are the schemes by the United States Department of Agriculture (USDA) and the so-called International System (IS), whose difference is the upper particle size limit of the silt fraction, namely, 0.02 mm for the IS and 0.05 mm for the USDA system. The existence of these and other systems creates a disparity that hinders and prevents the use and exchange of soil information worldwide. To solve this problem, it is necessary to devise methodologies for the conversion of textural fractions between the different classification systems. This work focuses on the estimation of the USDA silt fraction from the basic textural fractions (sand, silt and clay) in the IS. Five models are currently available for that purpose: the log-linear interpolation method, the Minasny-McBratney-Bristow regression formula, the Shirazi-Boersma-Johnson interpolation method, the Minasny-McBratney regression formula, and the Padarian-Minasny-McBratney regression formula. The accuracy of some of these methods has already been assessed, but in this work we develop a new methodology, based on a local quantile regression, which improves and enriches this evaluation, providing both the regions of the textural triangle where the predictions of the models are acceptable, and the regions where each model is most appropriate. The data used were taken from the publicly available National Cooperative Soil Survey Soil Characterization Database, from which more than 270,000 soil horizon samples were selected for having valid texture data. The analysis carried out concludes that the Padarian-Minasny-McBratney regression formula is the best model of those evaluated. In addition, the tool developed for the evaluation of the models becomes a new model that provides point estimates of the USDA silt fraction from the basic textural fractions in the IS, with further improvement, compared to the 5 models evaluated, as it also provides a prediction interval for those estimates.