Abstract:
Successful soil moisture data assimilation requires information about the uncertainties in the land surface model and the observations. While the errors in the observations are relatively easy to quantify, it is very difficult to specify the errors in the model, including approximations in the dynamical equations, wrong model parameters, or inaccurate forcing inputs.

In synthetic experiments, it is straightforward to derive optimal model error covariances by minimizing the actual estimation errors (estimates minus synthetic ``true'' fields) as a function of the model error parameters. When satellite observations are used, however, we cannot easily compute the actual estimation errors. By contrast, the innovations (observations minus model forecast) are always readily available.

In a twin experiment we generate synthetic 'true' land surface fields and an open loop model trajectory using different model parameters and forcing fields. We also generate synthetic observations from the true fields which are then assimilated into the open loop model many times using different model error covariances. The resulting innovations sequences are analyzed and their statistics are contrasted to the actual estimation errors. We show that for our nonlinear problem the innovations sequence is a valuable tool for assessing the actual estimation errors and thereby for identifying appropriate model error covariances.