Global Prediction of Soil Saturated Hydraulic Conductivity Using Random Forest in a Covariate‐Based GeoTransfer Function (CoGTF) Framework

Saturated hydraulic conductivity (Ksat) is a key soil hydraulic parameter for representing infiltration and drainage in land surface models. For large scale applications, Ksat is often estimated from pedotransfer functions (PTFs) based on easy-to-measure soil properties like soil texture and bulk density. The reliance of PTFs on data from uniform arable lands and the omission of soil structure limits the applicability of texture-based predictions of Ksat in vegetated lands. To include effects of terrain, climate, and vegetation in the derivation of a new global Ksat map at 1 km resolution, we harness technological advances in machine learning and availability of remotely sensed surrogate information. For model training and testing, a global compilation of 6,814 geo-referenced Ksat measurements from the literature was used. The accuracy assessment based on spatial cross-validation shows a concordance correlation coefficient (CCC) of 0.16 and a root mean square error (RMSE) of 1.18 for log10 Ksat values in cm/day (CCC = 0.79 and RMSE = 0.72 for non-spatial cross-validation). The generated maps of Ksat represent spatial patterns of soil formation processes more distinctly than previous global maps of Ksat based on easy-to-measure soil properties. The validation of the model indicates that Ksat could be modeled without bias using Covariate-based GeoTransfer Functions (CoGTFs) that harness spatially distributed surface and climate attributes, compared to soil information based PTFs. The relatively poor performance of all models in the validation (low CCC and high RMSE) highlights the need for the collection of additional Ksat values to train the model for regions with sparse data.