Mapping soil hydraulic properties using random-forest-based pedotransfer functions and geostatistics

Abstract

Abstract. Spatial 3-D information on soil hydraulic properties for areas larger than plot scale is usually derived using indirect methods such as pedotransfer functions (PTFs) due to the lack of measured information on them. PTFs describe the relationship between the desired soil hydraulic parameter and easily available soil properties based on a soil hydraulic reference dataset. Soil hydraulic properties of a catchment or region can be calculated by applying PTFs on available soil maps. Our aim was to analyse the performance of (i) indirect (using PTFs) and (ii) direct (geostatistical) mapping methods to derive 3-D soil hydraulic properties. The study was performed on the Balaton catchment area in Hungary, where density of measured soil hydraulic data fulfils the requirements of geostatistical methods. Maps of saturated water content (0 cm matric potential), field capacity (−330 cm matric potential) and wilting point (−15 000 cm matric potential) for 0–30, 30–60 and 60–90 cm soil depth were prepared. PTFs were derived using the random forest method on the whole Hungarian soil hydraulic dataset, which includes soil chemical, physical, taxonomical and hydraulic properties of some 12 000 samples complemented with information on topography, climate, parent material, vegetation and land use. As a direct and thus geostatistical method, random forest combined with kriging (RFK) was applied to 359 soil profiles located in the Balaton catchment area. There were no significant differences between the direct and indirect methods in six out of nine maps having root-mean-square-error values between 0.052 and 0.074 cm3 cm−3, which is in accordance with the internationally accepted performance of hydraulic PTFs. The PTF-based mapping method performed significantly better than the RFK for the saturated water content at 30–60 and 60–90 cm soil depth; in the case of wilting point the RFK outperformed the PTFs at 60–90 cm depth. Differences between the PTF-based and RFK mapped values are less than 0.025 cm3 cm−3 for 65 %–86 % of the catchment. In RFK, the uncertainty of input environmental covariate layers is less influential on the mapped values, which is preferable. In the PTF-based method the uncertainty of mapping soil hydraulic properties is less computationally intensive. Detailed comparisons of maps derived from the PTF-based method and the RFK are presented in this paper.

Using Machine Learning for Prediction of Saturated Hydraulic Conductivity and Its Sensitivity to Soil Structural Perturbations

“…Chaney et al (2019) similarly employed RF to build a map of predicted soil properties over the United States. Recently Szabó et al (2019) have developed PTFs based on RF and BRT to map soil hydraulic properties across a watershed. Koestel and Jorda (2014) showed that the RF algorithm can be used to accurately model soil preferential solute transport.…”

Pedotransfer Function for the Brunswick Soil Hydraulic Property Model and Comparison to the van Genuchten‐Mualem Model

“…Another advantage is that there are different VGM PTFs available (van Looy et al, 2017), with different requirements in input, so that, depending on the case specific information available, the case-specific double prediction step can be made. However, it is possible that this might not replace a newly constructed hyPTF based on machine learning tools, deemed superior for these kind of predictions (Szabó et al, 2019). A difficulty for that is that existing large data sets such as the EU-HYDI (Weynants et al, 2013) and HYPRESS (Wösten et al, 1999) data sets are not publically available, which in itself hampers model improvement.…”

Application of Hybrid Prediction Methods in Spatial Assessment of Inland Excess Water Hazard

“…Random Forest combined with Ordinary Kriging (RFK) is a relatively new hybrid method used in digital environmental mapping [53,68,69], which combines predictive model of Random Forest with kriging of the Random Forest residuals [70]. In RFK, the deterministic component is described by Random Forest (RF) as opposed to RK, where the deterministic component was described by MLR.…”

“…For example, Koch et al [51,52] used random forest regression kriging and random forest combined with residual Gaussian simulation for modelling the shallow groundwater and the depth of redox interface, respectively. Szabó et al [53] compared the performance of random-forest-based pedotransfer functions and random forest combined with kriging in deriving 3D soil hydraulic properties. Pásztor et al [54] mapped risk of IEW of a Hungarian county, Bozán et al [55] mapped relative frequency of IEW inundation on the Great Hungarian Plain.…”