CR 55:253-265 (2013)  -  DOI: https://doi.org/10.3354/cr01138

ABSTRACT: A new methodology was developed to create statistical emulators from an existing complex plant growth simulation model, in order to provide simple models for predicting perennial pasture production based on meteorological data. The goal was to find emulators that could fit simulation data and provide reliable predictions of new simulation output. Four types of statistical emulators were compared—linear models, non-linear models, linear splines and cubic splines—which varied in complexity due to the use of different functional forms to fit main effects and different methods to determine model parameters. Two methods of model validation were used to ensure that the emulator selected provided good fits to the initial simulation output and gave adequate predictions of future simulation output. We illustrated the various techniques of model development and validation using the Agricultural Production Systems sIMulator (APSIM) plant growth model. The effects of rainfall amount, rainfall frequency, temperature and radiation on APSIM-generated biomass production of the perennial pasture species lucerne Medicago sativa were used as a case study. We conclude that, in this case, the best statistical emulator is the cubic spline. The novel approach illustrated here provides a rigorous, flexible and powerful means of developing and testing emulators for any complex simulation model. 

KEY WORDS: Statistical emulator · Linear mixed effects model · Model validation · Cubic splines · APSIM · Medicago sativa · Biomass production