Reducing model complexity via output sensitivity

Ecosystem models help us understand the mechanisms that influence ecosystem health indicators. However, if they are too complex, these mechanisms can be difficult to identify. On the other hand, if they are too simple the mechanisms may be distorted or even absent. Determining an appropriate level of model complexity is therefore desirable. This paper introduces two model simplification methods that are based on the sensitivity of performance measures to model rates and components. The first method identifies rates that have little influence on the performance measures and subsequently eliminates them. The second identifies, for a given performance measure, state variables that can be made constant. The methods can be implemented automatically, so that familiarity with the model is not required a priori. Demonstrating with a biogeochemical model of Port Phillip Bay, Australia, we find that significant reduction in model complexity is possible, including reductions in model order. Also, the process of implementing the methods reveals insights into the system that were not obvious beforehand.