In an ensemble Kalman filter, a Monte Carlo procedure is used to obtain ensemble statistics that are representative of the ensemble mean error at all stages of a data-assimilation cycle. By running the ensemble Kalman filter with different subsets of the observational network, one can try to estimate the relative value of the different components of the network.

Simulations have been performed with a dry global primitive-equation model on a 144 x 72 horizontal grid and 21 vertical levels. The simulated observing system has only radiosonde, SATEM and aircraft reports. In an initial experiment, it is assumed that the forecast model has no error. It is found that the predicted error levels for the guess fields are much lower than what is typically observed in data-assimilation procedures at operational centers.

It is investigated to what extent the above results are a consequence of the modest resolution of the forecast model and of the use of a simple model forcing. It is found that these model-error components are indeed responsible for the low simulated error levels.

By accounting for model error, error levels that are similar to those typically observed can be obtained. However, to arrive at reliable quantitative statements about the observational network, it will be necessary to first perform experiments with real data in order to calibrate the model-error component.