This paper presents a methodology for the definition of an optimal set of sampling data for the calibration of a river water quality model. Starting with an extensive set of measurements, it is the aim to reduce those data to obtain just as much data as necessary for a calibration with an acceptable uncertainty in the parameters. The method requires a model for the river under examination and the availability of samples for a first calibration of the model. With the model, synthetic time series are generated, which can be used as virtual observations. In the next step, the method of D-optimal design is applied. The amount, frequency, period, place and kind of variables measured of the water samples that gives the most reliable estimates of the parameters of the model are considered to be the best observations that can be made for that river. Also, the percentage of improvement of the reliability can be defined, as a function of the observations. The method is applied to the river Dender.