Selecting parameters for Bayesian calibration of a process-based model: a methodology based on canonical correlation analysis

Bayesian statistics is becoming increasingly common in the environmental sciences because of developments in computers and sampling-based techniques for parameter estimation. However, the use of the Bayesian approach is still limited in forest research, especially for models with many parameters. Some studies have used parameter screening to make the calibration of a computationally expensive model possible. In this paper we introduce a new methodology for parameter screening, based on canonical correlation analysis. Furthermore we show how parameter screening impacts the performance of a process-based model. The methodology presented here can be generally applied and is particularly suitable for complex process-based models because it is not computationally demanding and is easy to implement. It provides an overall ranking in relation to all outputs of the model, as opposed to common sensitivity methods that analyze one model output variable at a time. We found that parameter screening can be used to reduce the computational load of Bayesian calibration, but only the least important parameters should be excluded from the calibration if we do not want to affect model performance. In this exercise, 25% of the parameters of a process-based forest model could be excluded from the calibration without affecting model performance. When calibration was limited to a more restricted number of parameters, model performance significantly deteriorated.