Inverse modelling of in situ soil water dynamics: investigating the effect of different prior distributions of the soil hydraulic parameters

In situ observations of soil water state variables under natural boundary conditions are often used to estimate the soil hydraulic properties. However, many contributions to the soil hydrological literature have demonstrated that the information content of such data is insufficient to accurately and precisely estimate all the soil hydraulic parameters. In this case study, we explored to which degree prior information about the soil hydraulic parameters can help improve parameter identifiability in inverse modelling of in situ soil water dynamics under natural boundary conditions. We used percentages of sand, silt, and clay as input variables to the ROSETTA pedotransfer function that predicts the parameters in the van Genuchten-Mualem (VGM) model of the soil hydraulic functions. To derive additional information about the correlation structure of the predicted parameters, which is not readily provided by ROSETTA, we employed a Monte Carlo approach. We formulated three prior distributions that incorporate to different extents the prior information about the VGM parameters derived with ROSETTA. The inverse problem was posed in a formal Bayesian framework and solved using Markov chain Monte Carlo (MCMC) simulation with the DiffeRential Evolution Adaptive Metropolis (DREAM) algorithm. Synthetic and real-world soil water content data were used to illustrate the approach. The results of this study demonstrated that prior information about the soil hydraulic parameters significantly improved parameter identifiability and that this approach was effective and robust, even in case of biased prior information. To be effective and robust, however, it was essential to use a prior distribution that incorporates information about parameter correlation