Statistical Characterization of the Root System Architecture Model CRootBox

Abstract

Core Ideas The stochasticity of 3D root architecture models needs to be recognized by statistical analysis. Probability density functions, regression, and correlation analyses reveal the impact of model input parameters on different CRootBox measures. Distributions of ratios of root system measures are highly asymmetric. Multivariate approaches (copulas) are envisioned for future root architecture model analysis. The connection between the parametrization of three‐dimensional (3D) root architecture models and characteristic measures of the simulated root systems is often not obvious. We used statistical methods to analyze the simulation outcome of the root architecture model CRootBox and built meta‐models that determine the dependency of root system measures on model input parameters. Starting with a reference parameter set, we varied selected input parameters one at a time and used CRootBox to compute 1000 root system realizations as well as their root system measures. The obtained data sets were then statistically analyzed with regard to dependencies between input parameters, as well as distributions and correlations between different root system measures. While absolute root system measures (e.g., total root length) were approximately normally distributed, distributions of ratios of root system measures (e.g., root tip density) were highly asymmetric and could be approximated with inverse gamma distributions. We derived regression models (meta‐models) that link significant model parameters to 18 widely used root system measures and determined correlations between different root system measures. Statistical analysis of 3D root architecture models helps to understand the impact of input parametrization on specific root architectural measures. Our developed meta‐models can be used to determine the effect of parameter variations on the distribution of root system measures without running a full simulation. Model intercomparison and benchmarking of root architecture models is still missing. Our approach provides a means to compare different models with each other and with experimental data.