Crop model parameter estimation and sensitivity analysis for large scale data using supercomputers

Abstract

Global crop production must be doubled by 2050 to feed 9 billion people. Novel crop improvement methods and management strategies are the sine qua non for achieving this goal. This requires reliable quantitative methods for predicting the behavior of crop cultivars in novel, time-varying environments. In the last century, two different mathematical prediction approaches emerged (1) quantitative genetics (QG) and (2) ecophysiological crop modeling (ECM). These methods are completely disjoint in terms of both their mathematics and their strengths and weaknesses. However, in the period from 1996 to 2006 a method for melding them emerged to support breeding programs. The method involves two steps: (1) exploiting ECM’s to describe the intricate, dynamic and environmentally responsive biological mechanisms determining crop growth and development on daily/hourly time scales; (2) using QG to link genetic markers to the values of ECM constants (called genotype-specific parameters, GSP’s) that encode the responses of different varieties to the environment. This can require huge amounts of computation because ECM’s have many GSP’s as well as site-specific properties (SSP’s, e.g. soil water holding capacity). Moreover, one cannot employ QG methods, unless the GSP’s from hundreds to thousands of lines are known. Thus, the overall objective of this study is to identify better ways to reduce the computational burden without minimizing ECM predictability. The study has three parts: (1) using the extended Fourier Amplitude Sensitivity Test (eFAST) to globally identify parameters of the CERES-Sorghum model that require accurate estimation under wet and dry environments; (2) developing a novel estimation method (Holographic Genetic Algorithm, HGA) applicable to both GSP and SSP estimation and testing it with the CROPGRO-Soybean model using 182 soybean lines planted in 352 site-years (7,426 yield observations); and (3) examining the behavior under estimation of the anthesis data prediction component of the CERES-Maize model. The latter study used 5,266 maize Nested Associated Mapping lines and a total 49,491 anthesis date observations from 11 plantings. Three major problems were discovered that challenge the ability to link QG and ECM’s: 1) model expressibility, 2) parameter equifinality, and 3) parameter instability. Poor expressibility is the structural inability of a model to accurately predict an observation. It can only be solved by model changes. Parameter equifinality occurs when multiple parameter values produce equivalent model predictions. This can be solved by using eFAST as a guide to reduce the numbers of interacting parameters and by collecting additional data types. When parameters are unstable, it is impossible to know what values to use in environments other than those used in calibration. All of the methods that will have to be applied to solve these problems will expand the amount of data used with ECM’s. This will require better optimization methods to estimate model parameters efficiently. The HGA developed in this study will be a good foundation to build on. Thus, future research should be directed towards solving these issues to enable ECM’s to be used as tools to support breeders, farmers, and researchers addressing global food security issues.