Approaches to Describe and Quantify Uncertainty in Bio-physical Agricultural Computer Simulation Models

Abstract

The objective was to explore the definition, sources, quantification, and management of uncertainty in bio-physical agricultural deterministic computer simulation models (crop models). One aim was to provide recommendations from a formalised statistical viewpoint on management of uncertainty for a small pool of crop model researchers in New Zealand. These researchers face unique issues with models that describe temperate, island-based conditions. An equally important goal was to identify ways to provide predictions complete with uncertainty bounds beyond those offered by sensitivity analysis. Given these objectives, my focus was on a single model. Section I proposes an uncertainty evaluation (UE) framework to explore how the combined components of a crop model contribute to the overall output uncertainty. Tools to curate information, diagnose the most important sources of uncertainty, and identify UE objectives have been developed. The framework links qualitative and quantitative analysis through a review of techniques for generating and analysing data from such models. Although many elements considered appear in the literature, amalgamation into a united framework is an original contribution. In Section II a detailed description of a case-study model that simulates wheat development is provided. This model provides a concrete foundation by which to demonstrate the UE framework, and illustrates the nature of crop models as constructs upon which mechanistic understanding of real-world systems continues to develop. A theoretical addition to a recent model that combines physiological and genetic characteristics of wheat is proposed based on laboratory based experimental work, reducing structural uncertainty. Finally, Section III is centred on the analysis of simulated data. In particular, it addressed the desire to provide credible intervals for state-space model estimates. This was achieved through fitting a probabilistic Bayesian hierarchical model with MCMC, a general form of data assimilation that recursively updates state predictions based on available data. Credible bounds of both an observed and a latent state variable were estimated for the case-study model. Finally, I summarise how these three sections tie together to resolve research objectives. I discuss the benefits of this research, recommendations and limitations, and propose directions for future work.