Rationale

ELDAS is using a common infrastructure implemented at three participating institutes: ECMWF, DWD and CNRM/Meteo France. The procedure is followed to be able to generate soil moisture fields for a suite of land surface schemes (LSM's), without the need of implementing a complex coupler of these LSM's to a common atmospheric host model. In addition, after ELDAS has finished, the operational implementation of the assimilation infrastructure should be as smooth as possible. Simultaneously, differences between soil moisture fields generated by different LSM's caused by differences in the forcings or data assimilation procedure should be kept to a minimum, to be able to trace back departures from observations as much as possible to properties of the land surface models.

The system layout

The data assimilation system can be thought of as a group of modules between which data are exchanged (see figure). The analysis method is already available at CNRM and operational at DWD, and will be made so at ECMWF. They all follow on the work by Rhodin et al (1999) and Hess et al (2001). The first guess and perturbed simulations will be used to generate the cost function gradient and the Jacobian matrix of the surface scheme using a finite difference approach. The analysis method will provide an optimal soil moisture analysis with its associated error. The Jacobian of the surface scheme will allow the propagation in time of the covariance matrix of the analysis error as given by the Kalman Filter theory. This propagated matrix will be used as an updated background error matrix. These simulations will be carried out with the full atmospheric model, but with model precip and radiation replaced by the datasets provided. The data assimilation will take place once per day, generating a soil moisture field at 0:00 UTC for Europe, using all observations of the day starting at this analysis time.

The whole process will take place in a sequential cycled way, updating the atmospheric initial fields using analyses, and propagating the soil fields as first guess. This is probably most easily achieved by using an SMS-like controlling system, but another scripting structure may also be applicable.

Grid

Generation of soil moisture increments will be carried out in model space, that is, by generating observation fields that are compatible with the resolution of the model for which the soil moisture fields are generated. The model grid will be different for different case- or validation studies. All observation fields that will enter the data assimilation (precipitation, radiation, screen level quantities, satellite heating rates, microwave data) should be on a predefined lat-lon grid, and should be accompagnied by a common set of interpolation/aggregation routines (or, more generally, up/down scaling procedures) to reproject these lat-lon grids to the model grids. The interpolation should be carried out according to the needs of the target model grid, preferrably without using any background information, as this information is supposed to be already in the observation fields generated. The interpolation software should be common to all partners and kept in the ELDAS web site.

DWD and CNRM will use limited area versions of their global forecast models to generate the regional soil moisture fields. ECMWF does not have such limited area version available, and will use either the global model at moderate resolution, or an ensemble of single column simulations to account for high resolution in small areas.

Observations

Missing observations, in particular the precipitation and radiation forcing, will be obtained from the extensive archive of ECMWF (re)analyses. For all observations (in particular for precipitation), it is important that the processors also provide the associated covariance errors that are needed by the the data assimilation.

A common GRIB format will be used in all datasets above (eg, forcing and assimilation output).

Single column experimentation

ECMWF will start the ELDAS work with a single column version of the above scheme, focusing at a data set for which the necessary forcings (including microwave data and satellite heating rates) are available. CNRM has offered the MUREX data which they used to design their soil assimilation and contains good soil moisture ground truth. Microwave data should be generated synthetically. Alternatively, we could make use of the SGP data for which ESTAR microwave data are available.

This single column set-up allows testing of the system in an easy and efficient manner. Furthermore, it is compatible with the deliverables of the microwave workpackage, in which we will explore the impacts of future SMOS data on the soil moisture product, in first instance by assessing the necessary signal/noise ratio that is necessary for SMOS to be of any significance, and to gain experience with the benefits of long-term data assimilation to transport the information on surface moisture down into the soil column. Gaining experience with the error statistics is also beneficial for the heating rates. This single column model (with the TESSEL surface scheme) should be made available to the groups that wish to experiment with it (in particular those involved in the SMOS and heating rate workpackages).