Keywords

Gaussian Process, APSIM, Sugarcane, BACCO

Location

Session C2: Accounting for Uncertainty in Decision Support by Treating Model Assumptions as Scenarios

Start Date

19-6-2014 10:40 AM

End Date

19-6-2014 12:20 PM

Abstract

While several statistical methods are available to analyse model sensitivity, their application to complex process-based models is often impractical due to the large number of simulation runs required. A Bayesian approach to global sensitivity analysis can greatly reduce the number of simulation runs required by building an emulator of the model which is less computationally demanding. A Gaussian Emulation Machine (GEM) was used to efficiently assess the sensitivity of key agronomic outputs from the APSIM-Sugar crop model to influential input parameters. The sensitivity of simulated biomass and sucrose at harvest was assessed on 14 parameters representing varietal differences and growth response to water stress. Analysis was performed under irrigated and water stressed conditions. Simulated biomass and sucrose were found to be insensitive to 4 of the parameters tested under both irrigated and stressed conditions. Both outputs were most sensitive to radiation use efficiency under irrigated conditions and transpiration efficiency under stressed conditions. Output sensitivity was often non-linear and for a given parameter, could vary between well irrigated and water stressed conditions. Understanding how these parameters affect simulation outputs and which parameters are most influential can help improve simulations of interactions between sugarcane varieties and growing environments. This in turn can help better guide management decisions in the future. The Bayesian approach to sensitivity analysis proved an efficient alternative requiring far fewer model simulations than other approaches to sensitivity analysis and effectively provided insight into influential and negligible model parameters.

COinS
 
Jun 19th, 10:40 AM Jun 19th, 12:20 PM

Global sensitivity analysis of key parameters in a process-based sugarcane growth model - A Bayesian approach

Session C2: Accounting for Uncertainty in Decision Support by Treating Model Assumptions as Scenarios

While several statistical methods are available to analyse model sensitivity, their application to complex process-based models is often impractical due to the large number of simulation runs required. A Bayesian approach to global sensitivity analysis can greatly reduce the number of simulation runs required by building an emulator of the model which is less computationally demanding. A Gaussian Emulation Machine (GEM) was used to efficiently assess the sensitivity of key agronomic outputs from the APSIM-Sugar crop model to influential input parameters. The sensitivity of simulated biomass and sucrose at harvest was assessed on 14 parameters representing varietal differences and growth response to water stress. Analysis was performed under irrigated and water stressed conditions. Simulated biomass and sucrose were found to be insensitive to 4 of the parameters tested under both irrigated and stressed conditions. Both outputs were most sensitive to radiation use efficiency under irrigated conditions and transpiration efficiency under stressed conditions. Output sensitivity was often non-linear and for a given parameter, could vary between well irrigated and water stressed conditions. Understanding how these parameters affect simulation outputs and which parameters are most influential can help improve simulations of interactions between sugarcane varieties and growing environments. This in turn can help better guide management decisions in the future. The Bayesian approach to sensitivity analysis proved an efficient alternative requiring far fewer model simulations than other approaches to sensitivity analysis and effectively provided insight into influential and negligible model parameters.