Abstract:
Accurate characterization of land surface variables is critical for weather and climate prediction, hydrological forecasting, land-atmosphere process understanding, and model improvement. Land data assimilation is the attempt to merge information from land observations and models in an optimal manner, but the field is still very much in its infancy, primarily due to a lack of quality large-scale observations.

To date land data assimilation has largely been conducted with land surface models that are not coupled to the atmosphere. Such uncoupled (or off-line) land assimilation relies heavily on two kinds of measurements: (1) meteorological data such as precipitation and air temperature that are used to force the land model and (2) observations of land surface states such as soil moisture or snow that are assimilated into the model.

The Global Land Data Assimilation System (GLDAS) headquartered at NASA and its sister projects make use of various new satellite- and ground-based observation systems to drive several land surface models in near-real time. Its suite of forcing variables and the on-going production of output from three global land models since 2001 make GLDAS a unique archive of land data. While the emphasis at GLDAS has so far been on forcing data, incorporation of advanced methods for the assimilation of soil moisture, soil temperature and snow is now under way.

Due to the nonlinear nature of land processes and the computational complexity of large-scale systems it is not trivial to choose the most appropriate method for assimilating observations of state variables. In the context of the NASA Seasonal-to-Interannual Prediction Project (NSIPP) we have used twin experiments to compare techniques for soil moisture estimation, including variational methods, the Ensemble Kalman filter (EnKF), and the Extended Kalman filter (EKF). While all methods produce satisfactory estimates, the variational method is more accurate than the EnKF which in turn is more accurate than the EKF. Nevertheless, the EnKF is probably the most suitable approach for land assimilation because it does not require an adjoint (which is rarely available for land models) and because it can account for the widest range of model errors. We will discuss results from NSIPP and GLDAS experiments including applications of land assimilation for the initialization of weather and climate forecasts.