A procedure to reduce parameter uncertainty for complex models by comparison with real system output illustrated on a potato growth model

Abstract

The parameter estimates in simulation models of (agro) ecosystems have a limited accuracy, which causes uncertainty in model output. Running the model for all biologically plausible values of the parameters gives rise toa model output range. In this paper a method to reduce model output uncertainty by comparing the model output range to real system output is presented. A first step is the definition of an acceptable model output: only part of the model output range will be acceptably close to the real system outputs. It is argued that this is not only a statistical question, but also depends on model purpose. The acceptable part of the model output range corresponds to a part of the original parameter range. The paper presents an algorithm to find these acceptable parameters ranges by an adaptive random search technique, which is relatively efficient and does not require any assumptions on model behavior in response to changes in parameter values. The issues in the paper are illustrated on a potato growth model and a series of observations on the growth of the Russet Burbank cultivar obtained from Wisconsin. The use of an acceptable range of parameter values as an outcome of the calibration procedure is illustrated by calculating the model response for two sets of input data. It appears that the range in model response for the same range of parameter values depends on the input data. For a simulation with increased irradiance, the model predicted similar yields with similar uncertainty ranges as for the calibration year; for a simulation with increased temperatures, the predicted yield was much lower and the relative uncertainty larger.