Global mapping of the soil water characteristics curve using machine-learning, a comprehensive dataset and spatial covariates

Abstract

<p>The representation of land surface processes in hydrologic and climatic models is critically dependent on the soil water characteristics curve (SWCC) that defines the hydrologic behavior of unsaturated soil. The SWCC depends not only on soil texture, but it is also shaped by biopores, soil structure, and clay type. To capture climate, vegetation and other soil formation processes on SWCC in spatial context, we predict how SWCC parameter values vary with local environmental covariates using a machine learning approach. The model was trained using (i) a novel and comprehensive compilation of global dataset of soil water retention measurements collected from the literature (approximately 13,000 pairs of water content and matric potential data) and (ii) global maps of environmental covariates and soil texture developed at 250 m resolution. Because in many cases only few measurements per sample are available to fit the SWCC, the estimated parameters are often highly uncertain and could yield unrealistic predictions of related physical quantities. To address these limitations, we added constraints to the values of residual and saturated water content based on clay content and mineralogy and ensured that the shape parameters related to air-entrance and pore size distribution honor other physical constraints, such as the characteristic length of evaporation and the ponding time. The resulting global maps of SWCC parameters are compared with predictions using pedotransfer functions (PTFs) based on soil information alone that were trained on data mainly collected for samples from arable land in temperate regions. We anticipate that our model including environmental covariates and geospatial data (covariate-based geotransfer functions CoGTFs) would enable us to provide more reliable predictions (compared to traditional PTFs) of SWCC that can be implemented in Earth system models.</p>