Many spatially distributed environmental models have been developed for small spatial units (e.g. individual plots or fields). Their application at a regional scale (e.g. large drainage basins) was hitherto not very successful since most models are 'scale-specific'. In most cases the lack of input data is a limiting factor for an appropriate model application at a regional scale. The use of simplified models on the other hand, can only be successful if they have the appropriate degree of complexity. This leads to the following paradox: on the one hand a model should be based on process knowledge so that it will react correctly to changes imposed by the modelling system. Including all available process knowledge, however, will lead to an overparametrisation. The error propagation involved in the poor data quality can then deteriorate dramatically the accuracy of the output results. For optimal model predictions at a regional scale the model complexity has to be in balance with the quality of the available input data. The error involved in a model application can be split in two parts: on the one hand the intrinsic model error because of an incomplete description of the processes and on the other hand the input error because of the use of low quality data. The total error is the sum of both parts. A simplification of the modelling structure will lead to an increase of the intrinsic model error and a decrease of the input error. If observed field values are available they can be used to determine which model structure gives the best results, given the available data. In many cases such validation data are not available. However, also in these cases, a model user should be able to select an optimal model structure. In this paper a technique is presented that allows the determination of the optimal degree of model complexity for an application at a regional scale if no observed field data are available. If the uncertainty of a model parameter exceeds a certain threshold, the model structure must be further simplified by replacement of this parameter by an aggregated value or a constant. The use of the technique is illustrated via the application of a soil erosion model.
