Since June 1997, CMC and MRB have replaced their previous statistical interpolation (SI) scheme with a 3D variational data assimilation (3D-var) for its global analysis (Gauthier et al., 1999). The regional analysis is also done operationally with the 3D-var since September 1997 (Laroche et al., 1999). In its first implementation, the 3D-var was configured to be as close as possible to the previous analysis based on optimal interpolation. In this formulation, the analysis increments for winds, geopotential and dew-point depression are produced on a prescribed set of pressure levels. This leads to many problems that can be avoided by doing the analysis directly on the model's own vertical levels in terms of temperature and specific humidity. This will be the formulation used in the next version of the 3D-var of CMC that will also include a new covariance model for the background error (Gauthier et al., 1998a). Work has been going on over the last few years on the assimilation of satellite data like cloud-cleared TIROS-N Operational Vertical Sounding (TOVS) radiances (Chouinard and Hallé, 1997) and measurements of total precipitable water from a Special Sensor Microwave Imager (SSM/I) (Deblonde, 1999). It is planned to include those in the operational analysis in 2000.

We have also implemented in GEM a hybrid coordinate which redresses more rapidly towards a pressure coordinate with height in view of performing 3D-var assimilation in the stratosphere with the new TOVS assimilation system. These stratospheric efforts are the first step toward the development of an expertise for making design choices in development of new space-borne instruments. Preliminary observing system simulation experiments (OSSE) with ozone profiles for the ORACLE (NASA,CSA) and ODIN (CSA, Sweden) projects have been done in passive tracer mode (Gauthier et al.,1998b; Kaminski et al, 1998; Brunet et al., 1999).

Several projects have been initiated, beginning in 1993, to develop the key elements leading to a 4D data assimilation system. The development of variational assimilation went along two main lines. The first one was the development of 3D-var that focused on making it possible to handle as many types of data as possible with a special emphasis on satellite data. Because the background error covariances play a key role in the analysis, it is also designed to handle more complex representations of the background error covariances such as those coming out of an ensemble forecasting system, a simplified Kalman filter or other means (Houtekamer and Mitchell, 1998). In parallel, the tangent linear and adjoint models of GEM model were developed with a view of eventually including the observation operators of the 3D-var to obtain a full 4D-var assimilation system (Tanguay et al., 1999). The background term used in 4D-var could use error covariances obtained from an ensemble Kalman filter as described above.