摘要：Core Ideas A robust PTF was developed to predict water contents at −1, −10, and −158 m tension. Drip infiltrometer experiments were inversely modeled to predict soil hydraulic properties. Both θ ( h ) and K ( h ) can be accurately estimated from experimental data together with PTFs. A transient flow experiment using automated drip infiltrometers (ADIs) was performed on soil columns (about 6 dm 3 ) large enough to incorporate macropore flow effects. We investigated to what extent the estimated soil hydraulic parameters obtained from inverse modeling of these experiments are reliable. A machine learning based pedotransfer function (PTF) for prediction of water content at −1, −10, and −158 m pressure head was developed. Sensitivity analysis of the van Genuchten parameters (residual and saturated water content θ r and θ s , fitting parameters α , n , and λ, and saturated hydraulic conductivity K s ) in soils of sandy, silty, and clayey textures showed that the temporal variation of pressure heads in ADI scenarios was not sensitive to θ r and θ s . The other parameters were accurately estimated from numerically synthesized data. The uniqueness of the estimated parameters did not change when a bias, representing experimental error, was added to the data set. In actual columns, using the temporal and spatial pressure head data from the ADIs and the water contents in the drier range predicted by the developed PTF resulted in a precise estimation of the van Genuchten parameters. Not including the PTF water contents resulted in non‐uniquely estimated van Genuchten parameters.
关键词：ADI; automated drip infiltrometer; GPR; Gaussian process regression; MSPE; mean square percentage error; NSE; Nash–Sutcliffe efficiency; PTF; pedotransfer function; RMSE; root mean square error; SHP; soil hydraulic property; VG; van Genuchten.
