Towards Optimizing Experiments for Maximum-confidence Model Selection between Different Soil-plant Models

Abstract

We present a method for experimental design, optimizing data acquisition for maximum confidence in the soil-plant model selection task. The method considers uncertainty in parameters, measurements and model structures. We combine Bayesian Model Averaging (BMA) with worth-of-data analysis and investigate how decisive the model weights are under different selections of data types. This allows assessing the power of different data types, data densities and data locations in identifying the best model structure from among a suite of plausible models. The models considered in this study are the crop models CERES, SUCROS, GECROS and SPASS, which are coupled to identical routines for simulating soil processes within the modeling framework Expert-N. The four models considerably differ in the degree of detail at which crop growth and root water uptake are represented. Monte-Carlo simulations were conducted for each of these models considering their uncertainty in soil hydraulic properties and selected crop model parameters. With Bayesian model updating by a Bootstrap filter, the models were then conditioned on field measurements of soil moisture, leaf area index (LAI), and evapotranspiration rates during a vegetation period of winter wheat in Nellingen, Southwestern Germany. Following our approach, we derived the BMA model weights when using all data or different subsets thereof. We discuss to which degree the posterior BMA mean outperformed the prior BMA mean and all individual posterior models, how informative the data types were for reducing prediction uncertainty of actual evapotranspiration, and how well the model structure can be identified based on the different data types and subsets.