Bayesian inference of root architectural model parameters from synthetic field data

Abstract

Background and aims: Characterizing root system architectures of field-grown crops is challenging as root systems are hidden in the soil. We investigate the possibility of estimating root architecture model parameters from soil core data in a Bayesian framework. Methods: In a synthetic experiment, we simulated wheat root systems in a virtual field plot with the stochastic CRootBox model. We virtually sampled soil cores from this plot to create synthetic measurement data. We used the Markov chain Monte Carlo (MCMC) DREAM(ZS) sampler to estimate the most sensitive root system architecture parameters. To deal with the CRootBox model stochasticity and limited computational resources, we essentially added a stochastic component to the likelihood function, thereby turning the MCMC sampling into a form of approximate Bayesian computation (ABC). Results: A few zero-order root parameters: maximum length, elongation rate, insertion angles, and numbers of zero-order roots, with narrow posterior distributions centered around true parameter values were identifiable from soil core data. Yet other zero-order and higher-order root parameters were not identifiable showing a sizeable posterior uncertainty. Conclusions: Bayesian inference of root architecture parameters from root density profiles is an effective method to extract information about sensitive parameters hidden in these profiles. Equally important, this method also identifies which information about root architecture is lost when root architecture is aggregated in root density profiles.