Much attention has been devoted to retrieving surface soil moisture from active and passive microwave satellite sensors. Such retrievals are of interest for a variety of reasons, including water cycle studies and the initialization of weather and climate forecasts. Soil moisture retrievals, however, suffer from gaps in spatial and temporal coverage and measure moisture only in the top few centimeters of the soil at best. Moreover, land model integrations driven with observed meteorological forcing data offer an independent source of information. To maximize their potential, the satellite retrievals may be merged with model estimates in a land data assimilation system.

Unfortunately, satellite retrievals of soil moisture are subject to large uncertainties because the physical processes that relate brightness temperature to soil moisture are difficult to parameterize, and because the necessary parameters are difficult to obtain on the global scale. An important question for the design of new satellite sensors is just how uncertain satellite retrievals can be and still add useful information to a land data assimilation system. In this paper, we address this question with a fraternal twin experiment that is based on high-resolution (1 km) "true" soil moisture fields and associated passive microwave brightness temperatures from a long-term integration of the TOPLATS land surface model over the Red-Arkansas river basin. From the true fields, we simulate many different retrieval data sets at a typical satellite footprint scale (36 km). The different retrieval data sets reflect various realistic sources of uncertainty with different error structure and magnitude.

The simulated retrieval data sets are then assimilated into the NASA Catchment land surface model with an Ensemble Kalman filter (EnKF), fitted with a simple and effective method of bias removal (cumulative distribution function matching). Finally, the quality of the assimilation estimates (with respect to the synthetic truth) is compared with that of a baseline integration of the Catchment model without assimilation. This procedure permits us to quantify the maximum level of uncertainty in the satellite retrievals for which information is still added in the assimilation, depending on the application. Performance measures include the traditional absolute (RMS) error, which is important for water cycle studies, and the time series correlation coefficient. The latter measures how well (scaled) anomalies are estimated, which contain the key information for forecast initialization.