Successful seasonal prediction of summer precipitation over land at mid-latitudes may depend on our ability to specify accurate soil moisture initial conditions. Such initial conditions can be derived by assimilating satellite observations of soil moisture into a land model up to the start time of the seasonal forecast using the Ensemble Kalman filter (EnKF). The land models used for soil moisture assimilation are typically one-dimensional models that treat each catchment (or alternatively, each grid box) independently of the other catchments. This does not mean, however, that errors in the forcing data for neighboring catchments are necessarily independent. Atmospheric forcing data in particular are likely to contain errors that are correlated over scales larger than the catchment scale of the land model. The potential value of such horizontal error correlations lies in the ability to spread information from observed to unobserved catchments.

In the Kalman filtering framework, the model and forcing errors are characterized by their mean value and covariance, including length scales for horizontal error correlations. One-dimensional approaches to the filtering problem assimilate observations independently into each catchment, effectively setting all horizontal error correlation scales to zero. Such one-dimensional filtering is computationally attractive but neglects potentially useful information. By contrast, three-dimensional filtering approaches take horizontal error correlations into account but are more computationally demanding.

Arguably the most important forcing input for soil moisture estimation is precipitation. We have computed error statistics from a comparison of two precipitation data sets for the contiguous United States. The data sets are (1) a reanalysis product from the European Centre for Medium-range Weather Forecasting (ECMWF) and (2) a daily gridded product based on gauge measurements from the Climate Prediction Center of the National Weather Service (CPC-NWS). The ECMWF reanalysis product represents typical precipitation data used for soil moisture assimilation, while the CPC-NWS gauge-based data is thought to be more accurate and better represents actual precipitation. The difference between gauge-based and reanalysis data is taken to be representative of errors in typical precipitation inputs used for soil moisture assimilation.

Our analysis reveals length scales for horizontal error correlations ranging from around 2 degrees during summer to 3 degrees during winter, which is larger than the average 0.5 degree scale of the catchments used in our model. In twin experiments with synthetic soil moisture observations we analyze the impact of such horizontal error correlations on estimation accuracy by comparing soil moisture output from one- and three-dimensional versions of the EnKF. Preliminary results indicate that the three-dimensional EnKF produces more accurate soil moisture estimates. For a given application, the gain in accuracy must be traded off against the additional computational burden.